At Flöka, we believe meaningful interventions start with rich data-sets that grow by asking intuitive questions

FLÖKA

flka

COPA v. Wright, Court Filing, retrieved on January 29, 2024, is part of HackerNoon’s Legal PDF Series. You can jump to any part of this filing here. This part is 39 of 42. 37. The Tulip Trust Deed (Particulars of Claim at [66A] {A/2/10}) ({ID_001186}) / {L8/20/1}) The Fifth Pleaded Example 728. In the Kleiman Litigation, Dr Wright proffered a Deed of Trust document as evidence of the existence of a trust called the Tulip Trust. Dr Wright has claimed that the Tulip Trust held Bitcoin and/or an encrypted file with keys to that Bitcoin. The Deed of Trust document that was adduced by Dr Wright was dated 23 October 2012. (a) COPA’s Reasons for Alleging Forgery 729. The date of 23 October 2012 was false. Computer forensic analysis of this document shows that it was backdated and that it was not created until at least 22 May 2015 (over two years after the death of David Kleiman on 26 April 2013). This document was therefore forged. As to the basis on which the Claimant alleges forgery, the Claimant relies on the above matters and the fact that this Deed of Trust document was found to be backdated in the Order of Justice Reinhart dated 27 August 2019. 730. A copy of this document is at C00000560 {L8/17/1}, which was included in the Extended Disclosure served on 7 March 2023. The Re-Re-Amended Defence was served on 20 April 2023. (b) Dr Wright’s Explanations and COPA’s Rebuttal 731. The document appears as a Deed of Trust between two Seychelles companies, Wright International Investments (“WIIL”) and Tulip Trading Ltd (“TTL”) with a date of 23 October 2012. 732. In Wright11 Appendix C, Dr Wright addressed this 23 October 2012 Deed of Trust document by reference to one particular version of it. There are a number of copies disclosed which are all identical in substance: 732.1. (C00000560) / {L8/17/1} – (includes the Kleiman case reference at the top). 732.2. (ID_001186) / {L8/20/1}; 732.3. (ID_003790) / {L8/28/1}; 732.4. (ID_003982) / {L8/14/1}; and 732.5. (ID_003329) / L15/86/1} – this is the version Dr Wright refers to in Wright11. 733. Mr Madden addresses the identical nature of the above documents in PM48, {H/304/8}. In PM48, Mr Madden notes that the documents analysed indicate that WIIL was purchased as an aged shelf company in 2014, not being active prior to that date: PM48 at [29] {H/304/9}. Mr Madden also notes the effect of similar analysis in respect of other similar documents relating to TTL (PM4 [119-127], {H/29/38}). 734. In his Defence (after various amendments), Dr Wright states that, pursuant to an order of the US Court in the Kleiman proceedings, he disclosed a copy of a Deed of Trust document between two Seychelles companies, WIIL and TTL, with a date of 23 October 2012 (Defence [85A(1)] {A/3/26}. Dr Wright claims that: “The date 23 October 2012 is not a false date as the Deed of Trust was executed on 23 October 2012” (Defence [85A(5)] {A/3/27}. 735. In Wright11, Dr Wright gives the following account. He says that he set up the trust structure with the intention that he did not have any rights to see documents until 2020 (Wright11 Appendix C, [6.6] {CSW/3/14}). Dr Wright says that Dianne Pinder of Lloyds solicitors in Brisbane Australia drafted a deed of trust to replace a former trust structure that had been set up in such a way that he did not have a right to see documents until 2020 (Wright11 Appendix C, [6.6] {CSW/3/14}). Dr Wright says that, as he was not a party to the Tulip Trust Deed, he can only comment on what appears on the face of the documents, and that the version at {ID_003329} is not the trust deed for the Tulip Trust (Wright11 Appendix C, [6.11] {CSW/3/15}). The position stated in Wright11 therefore appears to be that, whilst a trust was settled on 23 October 2012 which was called the Tulip Trust, the disclosed versions set out above do not represent the deed of that trust or record its terms. 736. Dr Wright further states in Wright11 that he thinks that the version at {ID_003329} has been doctored by someone, as some of the parties acting for Savannah Ltd as noted in the document held no authority with the company, which he did not find out until after 2022 (Wright11 Appendix C, [6.12] {CSW/3/15}). 737. In his oral evidence at trial, Dr Wright at first accepted that he had sworn in the Kleiman proceedings that the Tulip Trust Deed put to him in those proceedings (which was the version at {L8/17/1}) was an authentic document {Day4/107:4} to {Day4/109:5}. However, he then went on to say that it does not record the terms of the trust. He said that this particular Tulip Trust Deed is not a real document and that it is not part of the trust. Dr Wright claimed that he told the Judge in Kleiman that it was a real document but that he did not know any different to say otherwise at that time: {Day4/109:2} to {Day4/112:2}. He claimed that he had subsequently spoken to some of the individuals whose signatures appear on this Tulip Trust Deed and they told him that they did not sign it: {Day4/113:6}. 738. Dr Wright claimed that this particular version of the Tulip Trust Deed was formed by someone merging documents together: {Day4/114:5}. Dr Wright said that it was a fake document used by Ira Kleiman, and that there had never been any agreement made between Wright International and Tulip Trading: {Day4/115:10} to {Day4/115/20}. 739. COPA submitted that Dr Wright’s explanation should be rejected for the following reasons: 739.1. On Dr Wright’s own admission, the Tulip Trust Deed is not a genuine document on the basis that (on his account) (i) there never has been any agreement between WIIL and TTL at all; (ii) the signatures are said not to be genuine; and (iii) the terms do not reflect what he understood to be the terms of the Tulip Trust Deed actually executed on 23 October 2012. It is, however, remarkable that he failed to mention this in his Re-Amended Defence. 739.2. There is no evidence of any of other version of the Tulip Trust Deed dated to 23 October 2012 with different terms. Dr Wright has not brought any of those who are said to deny signing this deed to Court as witnesses, nor has he called anyone as a witness to say what the “real” terms of the Tulip Trust are. 739.3. WIIL is purportedly a party to the Tulip Trust Deed, but that company was only first purchased as an off the shelf company by Dr Wright in October 2014. Mr Madden's findings in Appendix PM48 to the effect that alteration of documents supports that proposition {H/304/9} were not challenged in cross-examination. 739.4. Paragraph 3 of this Tulip Trust Deed {L8/17/2} refers to bitcoin having been transferred into TTL on 10 June 2011, although that company likewise was not purchased until October 2014 (as explained in relation to {ID_001421} and {ID_001930} in COPA’s original Forgeries Schedule at {A/2/91} and {A/2/101}. 739.5. There is nothing in the terms of this trust (as recorded in this Deed) to say that trust documents cannot be shown to Dr Wright. He claimed that he did not know the terms of the document, but also that those terms precluded him from talking directly to any of the individuals involved. He then said that he had spoken to them and found out that the signatures on the document were not genuine. Dr Wright's evidence about this document, and the Tulip Trust more generally, has been confused and internally inconsistent. In particular, he has invented the excuse of having been precluded from seeing the document to explain away his having sworn to its authenticity in the Kleiman proceedings and made a sworn declaration based on its terms, but then having denied its authenticity in the Granath proceedings. See generally: {Day3/105:2} to {Day3/119:5}. (c) Conclusions 740. I found Dr Wright’s contorted explanations to be wholly unconvincing and false, including his final excuse that this document was ‘doctored by someone’. 741. There is strong evidence that (and I so find) both Tulip Trading Limited and Wright International Investments Ltd were acquired by Dr Wright in 2014 as aged shelf companies, and that the ‘Tulip Trust’ itself was a fabrication designed to shield Dr Wright’s assets against possible bankruptcy. Dr Wright’s explanations of the terms of the Trust were inconsistent and did not ring true. He has never explained the plain inconsistency between his assertion that the Trust put his assets (including his IP in various documents) beyond reach, yet his continued ability apparently to reference and use all such documents to yield the large number of ‘White Papers’ which are said to have given rise to all the claimed patent filings for nChain. Continue Reading Here. About HackerNoon Legal PDF Series: We bring you the most important technical and insightful public domain court case filings.

This court case retrieved on January 29, 2024, judiciary.uk is part of the public domain. The court-created documents are works of the federal government, and under copyright law, are automatically placed in the public domain and may be shared without legal restriction. COPA v. Wright, Court Filing, retrieved on January 29, 2024, is part of HackerNoon’s Legal PDF Series. You can jump to any part of this filing here. This part is 39 of 42. COPA v. Wright, Court Filing, retrieved on January 29, 2024, is part of HackerNoon’s Legal PDF Series . You can jump to any part of this filing here . This part is 39 of 42. HackerNoon’s Legal PDF Series HackerNoon’s Legal PDF Series here here 37. The Tulip Trust Deed (Particulars of Claim at [66A] {A/2/10}) ({ID_001186}) / {L8/20/1}) The Fifth Pleaded Example 728. In the Kleiman Litigation, Dr Wright proffered a Deed of Trust document as evidence of the existence of a trust called the Tulip Trust. Dr Wright has claimed that the Tulip Trust held Bitcoin and/or an encrypted file with keys to that Bitcoin. The Deed of Trust document that was adduced by Dr Wright was dated 23 October 2012. Kleiman (a) COPA’s Reasons for Alleging Forgery (a) COPA’s Reasons for Alleging Forgery 729. The date of 23 October 2012 was false. Computer forensic analysis of this document shows that it was backdated and that it was not created until at least 22 May 2015 (over two years after the death of David Kleiman on 26 April 2013). This document was therefore forged. As to the basis on which the Claimant alleges forgery, the Claimant relies on the above matters and the fact that this Deed of Trust document was found to be backdated in the Order of Justice Reinhart dated 27 August 2019. 730. A copy of this document is at C00000560 {L8/17/1}, which was included in the Extended Disclosure served on 7 March 2023. The Re-Re-Amended Defence was served on 20 April 2023. (b) Dr Wright’s Explanations and COPA’s Rebuttal (b) Dr Wright’s Explanations and COPA’s Rebuttal 731. The document appears as a Deed of Trust between two Seychelles companies, Wright International Investments (“ WIIL ”) and Tulip Trading Ltd (“ TTL ”) with a date of 23 October 2012. WIIL TTL 732. In Wright11 Appendix C, Dr Wright addressed this 23 October 2012 Deed of Trust document by reference to one particular version of it. There are a number of copies disclosed which are all identical in substance: 732.1. (C00000560) / {L8/17/1} – (includes the Kleiman case reference at the top). 732.2. (ID_001186) / {L8/20/1}; 732.3. (ID_003790) / {L8/28/1}; 732.4. (ID_003982) / {L8/14/1}; and 732.5. (ID_003329) / L15/86/1} – this is the version Dr Wright refers to in Wright11. 733. Mr Madden addresses the identical nature of the above documents in PM48, {H/304/8}. In PM48, Mr Madden notes that the documents analysed indicate that WIIL was purchased as an aged shelf company in 2014, not being active prior to that date: PM48 at [29] {H/304/9}. Mr Madden also notes the effect of similar analysis in respect of other similar documents relating to TTL (PM4 [119-127], {H/29/38}). 734. In his Defence (after various amendments), Dr Wright states that, pursuant to an order of the US Court in the Kleiman proceedings, he disclosed a copy of a Deed of Trust document between two Seychelles companies, WIIL and TTL, with a date of 23 October 2012 (Defence [85A(1)] {A/3/26}. Dr Wright claims that: “The date 23 October 2012 is not a false date as the Deed of Trust was executed on 23 October 2012” (Defence [85A(5)] {A/3/27}. “The date 23 October 2012 is not a false date as the Deed of Trust was executed on 23 October 2012” 735. In Wright11, Dr Wright gives the following account. He says that he set up the trust structure with the intention that he did not have any rights to see documents until 2020 (Wright11 Appendix C, [6.6] {CSW/3/14}). Dr Wright says that Dianne Pinder of Lloyds solicitors in Brisbane Australia drafted a deed of trust to replace a former trust structure that had been set up in such a way that he did not have a right to see documents until 2020 (Wright11 Appendix C, [6.6] {CSW/3/14}). Dr Wright says that, as he was not a party to the Tulip Trust Deed, he can only comment on what appears on the face of the documents, and that the version at {ID_003329} is not the trust deed for the Tulip Trust (Wright11 Appendix C, [6.11] {CSW/3/15}). The position stated in Wright11 therefore appears to be that, whilst a trust was settled on 23 October 2012 which was called the Tulip Trust, the disclosed versions set out above do not represent the deed of that trust or record its terms. 736. Dr Wright further states in Wright11 that he thinks that the version at {ID_003329} has been doctored by someone, as some of the parties acting for Savannah Ltd as noted in the document held no authority with the company, which he did not find out until after 2022 (Wright11 Appendix C, [6.12] {CSW/3/15}). 737. In his oral evidence at trial, Dr Wright at first accepted that he had sworn in the Kleiman proceedings that the Tulip Trust Deed put to him in those proceedings (which was the version at {L8/17/1}) was an authentic document {Day4/107:4} to {Day4/109:5}. However, he then went on to say that it does not record the terms of the trust. He said that this particular Tulip Trust Deed is not a real document and that it is not part of the trust. Dr Wright claimed that he told the Judge in Kleiman that it was a real document but that he did not know any different to say otherwise at that time: {Day4/109:2} to {Day4/112:2}. He claimed that he had subsequently spoken to some of the individuals whose signatures appear on this Tulip Trust Deed and they told him that they did not sign it: {Day4/113:6}. 738. Dr Wright claimed that this particular version of the Tulip Trust Deed was formed by someone merging documents together: {Day4/114:5}. Dr Wright said that it was a fake document used by Ira Kleiman, and that there had never been any agreement made between Wright International and Tulip Trading: {Day4/115:10} to {Day4/115/20}. 739. COPA submitted that Dr Wright’s explanation should be rejected for the following reasons: 739.1. On Dr Wright’s own admission, the Tulip Trust Deed is not a genuine document on the basis that (on his account) (i) there never has been any agreement between WIIL and TTL at all; (ii) the signatures are said not to be genuine; and (iii) the terms do not reflect what he understood to be the terms of the Tulip Trust Deed actually executed on 23 October 2012. It is, however, remarkable that he failed to mention this in his Re-Amended Defence. 739.2. There is no evidence of any of other version of the Tulip Trust Deed dated to 23 October 2012 with different terms. Dr Wright has not brought any of those who are said to deny signing this deed to Court as witnesses, nor has he called anyone as a witness to say what the “real” terms of the Tulip Trust are. 739.3. WIIL is purportedly a party to the Tulip Trust Deed, but that company was only first purchased as an off the shelf company by Dr Wright in October 2014. Mr Madden's findings in Appendix PM48 to the effect that alteration of documents supports that proposition {H/304/9} were not challenged in cross-examination. 739.4. Paragraph 3 of this Tulip Trust Deed {L8/17/2} refers to bitcoin having been transferred into TTL on 10 June 2011, although that company likewise was not purchased until October 2014 (as explained in relation to {ID_001421} and {ID_001930} in COPA’s original Forgeries Schedule at {A/2/91} and {A/2/101}. 739.5. There is nothing in the terms of this trust (as recorded in this Deed) to say that trust documents cannot be shown to Dr Wright. He claimed that he did not know the terms of the document, but also that those terms precluded him from talking directly to any of the individuals involved. He then said that he had spoken to them and found out that the signatures on the document were not genuine. Dr Wright's evidence about this document, and the Tulip Trust more generally, has been confused and internally inconsistent. In particular, he has invented the excuse of having been precluded from seeing the document to explain away his having sworn to its authenticity in the Kleiman proceedings and made a sworn declaration based on its terms, but then having denied its authenticity in the Granath proceedings. See generally: {Day3/105:2} to {Day3/119:5}. Kleiman Granath (c) Conclusions (c) Conclusions 740. I found Dr Wright’s contorted explanations to be wholly unconvincing and false, including his final excuse that this document was ‘doctored by someone’. 741. There is strong evidence that (and I so find) both Tulip Trading Limited and Wright International Investments Ltd were acquired by Dr Wright in 2014 as aged shelf companies, and that the ‘Tulip Trust’ itself was a fabrication designed to shield Dr Wright’s assets against possible bankruptcy. Dr Wright’s explanations of the terms of the Trust were inconsistent and did not ring true. He has never explained the plain inconsistency between his assertion that the Trust put his assets (including his IP in various documents) beyond reach, yet his continued ability apparently to reference and use all such documents to yield the large number of ‘White Papers’ which are said to have given rise to all the claimed patent filings for nChain. Continue Reading Here. Continue Reading Here . Here About HackerNoon Legal PDF Series: We bring you the most important technical and insightful public domain court case filings. This court case retrieved on January 29, 2024, judiciary.uk is part of the public domain. The court-created documents are works of the federal government, and under copyright law, are automatically placed in the public domain and may be shared without legal restriction. About HackerNoon Legal PDF Series: We bring you the most important technical and insightful public domain court case filings. This court case retrieved on January 29, 2024, judiciary.uk is part of the public domain. The court-created documents are works of the federal government, and under copyright law, are automatically placed in the public domain and may be shared without legal restriction. judiciary.uk judiciary.uk

Forensic Analysis Exposes Flaws in Craig Wright's Tulip Trust Deed

Flare Network has announced its integration with Stargate V2, a leading omnichain liquidity transport protocol. The collaboration, revealed on July 18, 2024, promises to connect Flare to over 25 blockchain networks, including Ethereum, Arbitrum, Optimism, and Base. This strategic partnership comes at a pivotal time for both organizations. Flare, known for its focus on data-intensive blockchain applications, aims to deepen its liquidity and expand its Total Value Locked (TVL). Stargate, on the other hand, has been dominating the web3 interoperability space, handling up to 50% of daily cross-chain transaction volume during peak periods. Hugo Philion, Flare's Co-Founder and CEO, expressed enthusiasm about the integration. "FLRfiSummer is accelerating," Philion stated, referring to Flare's ongoing initiatives. He highlighted that the partnership would provide holders of USDC and USDT across multiple networks with "seamless access to industry-leading yields" through Flare's new DeFi emissions program. The integration leverages Stargate V2's Hydra feature, allowing for the smooth transfer of USDC, USDT, and ETH to Flare. This mechanism locks the original assets in native pools on source chains while minting equivalent Hydra assets on Flare, ensuring unified liquidity across all connected chains. Angus Lamps, Foundation Lead of Stargate, emphasized the potential impact of this collaboration. "As one of the most reliable and popular bridge protocols, Stargate V2 can help to onboard a significant user base of advanced DeFi users to Flare through unified liquidity," Lamps explained. The timing of this integration is particularly noteworthy, coinciding with the launch of a major DeFi incentives program on Flare. This program aims to boost stablecoin liquidity and lay the groundwork for Flare's upcoming FAssets system, which promises to bring DeFi capabilities to non-smart contract chains. For developers and users alike, this integration opens up new possibilities. It not only expands access to cross-chain liquidity but also provides a robust platform for building omnichain applications. Flare's unique selling point - its enshrined data protocols - combined with Stargate's liquidity solutions, creates a powerful toolkit for blockchain innovators. However, the road ahead is not without challenges. The blockchain space is notoriously volatile, and the success of this integration will depend on user adoption and market conditions. Additionally, regulatory scrutiny of cross-chain protocols has been increasing, which could potentially impact future operations. As the crypto industry continues to evolve, partnerships like this between Flare and Stargate highlight the ongoing trend towards greater interoperability and liquidity in the blockchain ecosystem. Whether this collaboration will lead to the "dramatic deepening of stablecoin liquidity" that Philion envisions remains to be seen, but it certainly marks a significant step towards a more interconnected blockchain future. Don’t forget to like and share the story! Vested Interest Disclosure: This author is an independent contributor publishing via our business blogging program. HackerNoon has reviewed the report for quality, but the claims herein belong to the author. #DYOR. Flare Network has announced its integration with Stargate V2, a leading omnichain liquidity transport protocol. The collaboration, revealed on July 18, 2024, promises to connect Flare to over 25 blockchain networks, including Ethereum, Arbitrum, Optimism, and Base. This strategic partnership comes at a pivotal time for both organizations. Flare, known for its focus on data-intensive blockchain applications, aims to deepen its liquidity and expand its Total Value Locked (TVL). Stargate, on the other hand, has been dominating the web3 interoperability space, handling up to 50% of daily cross-chain transaction volume during peak periods. Flare Network Hugo Philion, Flare's Co-Founder and CEO, expressed enthusiasm about the integration. "FLRfiSummer is accelerating," Philion stated, referring to Flare's ongoing initiatives. He highlighted that the partnership would provide holders of USDC and USDT across multiple networks with "seamless access to industry-leading yields" through Flare's new DeFi emissions program. The integration leverages Stargate V2's Hydra feature, allowing for the smooth transfer of USDC, USDT, and ETH to Flare. This mechanism locks the original assets in native pools on source chains while minting equivalent Hydra assets on Flare, ensuring unified liquidity across all connected chains. Angus Lamps, Foundation Lead of Stargate , emphasized the potential impact of this collaboration. "As one of the most reliable and popular bridge protocols, Stargate V2 can help to onboard a significant user base of advanced DeFi users to Flare through unified liquidity," Lamps explained. Stargate The timing of this integration is particularly noteworthy, coinciding with the launch of a major DeFi incentives program on Flare. This program aims to boost stablecoin liquidity and lay the groundwork for Flare's upcoming FAssets system, which promises to bring DeFi capabilities to non-smart contract chains. For developers and users alike, this integration opens up new possibilities. It not only expands access to cross-chain liquidity but also provides a robust platform for building omnichain applications. Flare's unique selling point - its enshrined data protocols - combined with Stargate's liquidity solutions, creates a powerful toolkit for blockchain innovators. However, the road ahead is not without challenges. The blockchain space is notoriously volatile, and the success of this integration will depend on user adoption and market conditions. Additionally, regulatory scrutiny of cross-chain protocols has been increasing, which could potentially impact future operations. As the crypto industry continues to evolve, partnerships like this between Flare and Stargate highlight the ongoing trend towards greater interoperability and liquidity in the blockchain ecosystem. Whether this collaboration will lead to the "dramatic deepening of stablecoin liquidity" that Philion envisions remains to be seen, but it certainly marks a significant step towards a more interconnected blockchain future. Don’t forget to like and share the story! Vested Interest Disclosure: This author is an independent contributor publishing via our business blogging program. HackerNoon has reviewed the report for quality, but the claims herein belong to the author. #DYOR. Vested Interest Disclosure: This author is an independent contributor publishing via our business blogging program . HackerNoon has reviewed the report for quality, but the claims herein belong to the author. #DYOR. business blogging program business blogging program

Flare Network Integrates with Stargate V2 to Enhance Cross-Chain Liquidity

ফ্লেয়ার নেটওয়ার্ক ক্রস-চেইন লিকুইডিটি বাড়ানোর জন্য স্টারগেট V2 এর সাথে একীভূত করে

Flare Network integriert Stargate V2, um die kettenübergreifende Liquidität zu verbessern

Flare Network se integra con Stargate V2 para mejorar la liquidez entre cadenas

Flare Network s'intègre à Stargate V2 pour améliorer la liquidité inter-chaînes

फ़्लेयर नेटवर्क ने क्रॉस-चेन लिक्विडिटी बढ़ाने के लिए स्टारगेट V2 के साथ एकीकरण किया

Flare Network が Stargate V2 と統合し、クロスチェーンの流動性を強化

Flare Network는 Stargate V2와 통합되어 체인 간 유동성을 향상시킵니다.

Flare Network se integra ao Stargate V2 para aumentar a liquidez entre cadeias

Flare Network интегрируется со Stargate V2 для повышения межцепочной ликвидности

Flare Network, Zincirler Arası Likiditeyi Artırmak İçin Stargate V2 ile Bütünleşiyor

Mạng Flare tích hợp với Stargate V2 để tăng cường thanh khoản chuỗi chéo

Flare Network 与 Stargate V2 集成以增强跨链流动性

Introduction The world of browser games has seen a dramatic evolution, transforming from simple Flash-based creations to complex, multi-faceted experiences using modern web technologies. This journey reflects broader changes in web development, user expectations, and technological advancements. What Are Browser Games? Browser games are video games played directly in a web browser, without needing additional software installation. Early browser games were simple, often text-based or featuring rudimentary graphics. They provided an easy entry point for developers and players alike, allowing for quick access and play. The Flash Era: A Golden Age The late 1990s and early 2000s marked the heyday of Flash games. Adobe Flash, initially developed by Macromedia, became the cornerstone of browser gaming due to its ability to handle animations and interactive content efficiently. Platforms like Newgrounds and Kongregate emerged as central hubs for Flash games, fostering a vibrant community of indie developers and players. Flash’s appeal lay in its accessibility and ease of use. Aspiring game developers could create visually impressive games without needing extensive programming knowledge. This democratization of game development led to a boom in creative and experimental titles. Notable successes from this era include "Alien Hominid" and "Super Meat Boy," which started as Flash games before transitioning to larger platforms​. Transition to Modern Technologies The decline of Flash began in the late 2000s, accelerated by security concerns, performance issues, and the rise of mobile gaming. In 2017, Adobe announced the end of Flash support, prompting developers to transition to newer technologies like HTML5, JavaScript, and WebGL. HTML5, in particular, revolutionized browser gaming by providing a standardized way to create multimedia-rich applications without relying on plugins. Combined with JavaScript and WebGL, developers could now build sophisticated, high-performance games that worked seamlessly across devices and browsers. These technologies offer several APIs, such as the Canvas API for 2D graphics, WebGL for 3D graphics, and Web Audio API for advanced audio processing, enabling richer gaming experiences​. Benefits and Challenges of Developing Browser Games Developing browser games offers several benefits: Cross-Platform Compatibility: Modern web technologies ensure that games run on any device with a browser, from PCs to smartphones.
Ease of Access: Players can access games instantly without downloads, promoting wider reach and engagement.
Frequent Updates: Developers can push updates directly to the browser, ensuring all players have the latest version without needing manual downloads. However, there are also challenges: Performance Limitations: Browser games must run efficiently within the constraints of a web browser, which can limit complex graphics or large-scale simulations.
Security Concerns: Ensuring the security of in-browser code and protecting against exploits remains a critical challenge.
Monetization: While ads and in-game purchases are common, finding sustainable monetization strategies can be difficult, especially compared to traditional gaming platforms​​. Impact of the Transition from Flash The shift from Flash to HTML5 and other modern technologies significantly impacted the browser gaming industry. While the transition required developers to learn new tools and frameworks, it also opened opportunities for creating more advanced and secure games. The end of Flash marked the close of an era but also paved the way for more robust and versatile web-based games​​. Notable Modern Browser Games Several modern browser games showcase the capabilities of new technologies. Games like "Drift Boss," "1v1.lol," and "Smash Karts" have gained massive popularity due to their simple yet addictive gameplay, cross-platform compatibility, and real-time multiplayer features. These games highlight how modern web technologies can deliver engaging and scalable gaming experiences directly in the browser. Browser games used to be low quality and simplistic in gameplay and design, but this is rapidly changing. Nowadays, they can feature graphics and sound design comparable to those produced by big studios. For example, playing the aforementioned 1v1.lol directly in the browser offers an amazing arena shooter experience without the hassle of downloading and installing anything extra! Conclusion The evolution of browser games from the Flash era to modern times highlights the dynamic nature of web development and gaming. As technologies advance, browser games are poised to become even more immersive and accessible, potentially surpassing traditional gaming platforms. This shift presents exciting possibilities and new frontiers for developers and players alike. If asked whether the future of gaming will be fully browser-based and livestreamed directly to our devices, I would confidently say yes. What do you think? Introduction Introduction The world of browser games has seen a dramatic evolution, transforming from simple Flash-based creations to complex, multi-faceted experiences using modern web technologies. This journey reflects broader changes in web development, user expectations, and technological advancements. What Are Browser Games? What Are Browser Games? Browser games are video games played directly in a web browser, without needing additional software installation. Early browser games were simple, often text-based or featuring rudimentary graphics. They provided an easy entry point for developers and players alike, allowing for quick access and play. The Flash Era: A Golden Age The Flash Era: A Golden Age The late 1990s and early 2000s marked the heyday of Flash games. Adobe Flash, initially developed by Macromedia, became the cornerstone of browser gaming due to its ability to handle animations and interactive content efficiently. Platforms like Newgrounds and Kongregate emerged as central hubs for Flash games , fostering a vibrant community of indie developers and players. Newgrounds Kongregate Flash games Flash’s appeal lay in its accessibility and ease of use. Aspiring game developers could create visually impressive games without needing extensive programming knowledge. This democratization of game development led to a boom in creative and experimental titles. Notable successes from this era include " Alien Hominid " and " Super Meat Boy ," which started as Flash games before transitioning to larger platforms​. Alien Hominid Super Meat Boy Transition to Modern Technologies Transition to Modern Technologies The decline of Flash began in the late 2000s, accelerated by security concerns, performance issues, and the rise of mobile gaming. In 2017, Adobe announced the end of Flash support, prompting developers to transition to newer technologies like HTML5, JavaScript, and WebGL. HTML5, in particular, revolutionized browser gaming by providing a standardized way to create multimedia-rich applications without relying on plugins. Combined with JavaScript and WebGL, developers could now build sophisticated, high-performance games that worked seamlessly across devices and browsers. These technologies offer several APIs, such as the Canvas API for 2D graphics, WebGL for 3D graphics, and Web Audio API for advanced audio processing, enabling richer gaming experiences​. Benefits and Challenges of Developing Browser Games Benefits and Challenges of Developing Browser Games Developing browser games offers several benefits: Cross-Platform Compatibility: Modern web technologies ensure that games run on any device with a browser, from PCs to smartphones. Ease of Access: Players can access games instantly without downloads, promoting wider reach and engagement. Frequent Updates: Developers can push updates directly to the browser, ensuring all players have the latest version without needing manual downloads. Cross-Platform Compatibility: Modern web technologies ensure that games run on any device with a browser, from PCs to smartphones. Cross-Platform Compatibility: Ease of Access: Players can access games instantly without downloads, promoting wider reach and engagement. Ease of Access: Frequent Updates: Developers can push updates directly to the browser, ensuring all players have the latest version without needing manual downloads. Frequent Updates: However, there are also challenges: Performance Limitations: Browser games must run efficiently within the constraints of a web browser, which can limit complex graphics or large-scale simulations. Security Concerns: Ensuring the security of in-browser code and protecting against exploits remains a critical challenge. Monetization: While ads and in-game purchases are common, finding sustainable monetization strategies can be difficult, especially compared to traditional gaming platforms​​. Performance Limitations: Browser games must run efficiently within the constraints of a web browser, which can limit complex graphics or large-scale simulations. Performance Limitations: Security Concerns: Ensuring the security of in-browser code and protecting against exploits remains a critical challenge. Security Concerns: Monetization: While ads and in-game purchases are common, finding sustainable monetization strategies can be difficult, especially compared to traditional gaming platforms​​. Monetization: Impact of the Transition from Flash Impact of the Transition from Flash The shift from Flash to HTML5 and other modern technologies significantly impacted the browser gaming industry. While the transition required developers to learn new tools and frameworks, it also opened opportunities for creating more advanced and secure games. The end of Flash marked the close of an era but also paved the way for more robust and versatile web-based games​​. Notable Modern Browser Games Notable Modern Browser Games Several modern browser games showcase the capabilities of new technologies. Games like " Drift Boss ," " 1v1.lol ," and " Smash Karts " have gained massive popularity due to their simple yet addictive gameplay, cross-platform compatibility, and real-time multiplayer features. These games highlight how modern web technologies can deliver engaging and scalable gaming experiences directly in the browser. Drift Boss 1v1.lol Smash Karts Browser games used to be low quality and simplistic in gameplay and design, but this is rapidly changing. Nowadays, they can feature graphics and sound design comparable to those produced by big studios. For example, playing the aforementioned 1v1.lol directly in the browser offers an amazing arena shooter experience without the hassle of downloading and installing anything extra! Conclusion Conclusion The evolution of browser games from the Flash era to modern times highlights the dynamic nature of web development and gaming. As technologies advance, browser games are poised to become even more immersive and accessible, potentially surpassing traditional gaming platforms. This shift presents exciting possibilities and new frontiers for developers and players alike. If asked whether the future of gaming will be fully browser-based and livestreamed directly to our devices, I would confidently say yes. What do you think?

The Evolution of Browser Games: From Flash to Modern Times

Authors:
(1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA;
(4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA;
(5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA;
(6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA;
(7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil;
(8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA;
(9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction 2 Related Works 3 Med-Flamingo 4 Evaluation 5 Results 6 Discussion, Acknowledgments, and References A Appendix A APPENDIX A.1 ADDITIONAL DETAILS FOR MTB DATASET Clustering the images In a post-hoc analysis, we clustered the image embeddings of the MTB dataset into a large number of clusters (100) and manually reviewed examples of each cluster to assign an annotation. We discard noisy or unclear clusters and display the remaining clusters and their frequency in Figure 7. Classification of book titles Here, we provide further details about the creation of Figure 3. Table 4 lists the categories used to prompt the Claude-1 model to classify each book title. We initially prompted with 3 more very rare categories (Geriatrics, Occupational medicine, Space medicine), but merge them into the ”Other” group for visualization purposes. A.2 ADDITIONAL DETAILS FOR VISUAL USMLE DATASET This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Authors: Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Works 2 Related Works 3 Med-Flamingo 3 Med-Flamingo 4 Evaluation 4 Evaluation 5 Results 5 Results 6 Discussion, Acknowledgments, and References 6 Discussion, Acknowledgments, and References A Appendix A Appendix A APPENDIX A.1 ADDITIONAL DETAILS FOR MTB DATASET Clustering the images In a post-hoc analysis, we clustered the image embeddings of the MTB dataset into a large number of clusters (100) and manually reviewed examples of each cluster to assign an annotation. We discard noisy or unclear clusters and display the remaining clusters and their frequency in Figure 7. Clustering the images Classification of book titles Here, we provide further details about the creation of Figure 3. Table 4 lists the categories used to prompt the Claude-1 model to classify each book title. We initially prompted with 3 more very rare categories (Geriatrics, Occupational medicine, Space medicine), but merge them into the ”Other” group for visualization purposes. Classification of book titles A.2 ADDITIONAL DETAILS FOR VISUAL USMLE DATASET This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. available on arxiv

Med-Flamingo: a Multimodal Medical Few-shot Learner - Appendix

Med-Flamingo: a Multimodal Medical Few-shot Learner - Discussion, Acknowledgments, and References

Authors:
(1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA;
(4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA;
(5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA;
(6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA;
(7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil;
(8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA;
(9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction 2 Related Works 3 Med-Flamingo 4 Evaluation 5 Results 6 Discussion, Acknowledgments, and References A Appendix 5 RESULTS In our experiments, we focus on generative medical visual question answering (VQA). While recent medical VLMs predominantly performed VQA in a non-generative but rather discriminative manner (i.e., by scoring different answer choices), we believe that this ex-post classification to carry less clinical usefulness, than directly generating responses. On the other hand, generative VQA is more challenging to evaluate, as automated metrics suffer from significant limitations as they do not fully capture the domain-specific context. Thus, we perform a human evaluation study where clinical experts review model generations (blinded) and score them (between 0 and 10) in terms of clinical usefulness. Conventional VQA datasets Table 1 shows the results for VQA-RAD, the radiological VQA dataset for which we created custom splits to address leakage (see Section4). Med-Flamingo few-shot shows strong results, improving the clinical eval score by ∼ 20% over the best baseline. In this dataset, the auxiliary metrics are rather aligned with clinical preference. Finetuning the MedVINT baseline did not lead to improved performance on this dataset which may be due to its small size. MedVINT zero-shot outperforms the other zero-shot ablations which may be partially attributed to its instruction tuning step on PMC-VQA. Table 2 shows for the results for Path-VQA, the pathology VQA dataset. Compared to the other datasets, all models overall perform poorer on the Path-VQA dataset in terms of clinical evaluation score. We hypothesize that this has to do with the fact the models are not pre-trained on actual large-scale and fine-grained pathology image datasets, but only on a rather small amount of pathology literature (which may not be enough to achieve strong performance). For instance, Figure 3 shows that only a small fraction of our training data covers pathology. In the automated metrics (BERT-sim and exact-match), Med-Flamingo improves upon the OpenFlamingo baseline, however the overall quality does not improve (as seen in the clinical evaluation score). MedVINT was fine-tuned on a sizeable training split which results in strong automated metrics, but did not result in a clinical evaluation score that matches any Flamingo variant. Visual USMLE Table 3 shows the results for the Visual USMLE dataset. Med-Flamingo (fewshot) results in the clinically most preferrable generations, whereas OpenFlamingo (zero-shot) is a close runner-up. As the ground truth answers were rather lengthy paragraphs, exact match was not an informative metric (constant 0 for all methods). The few-shot prompted models lead to lower automated scores than their zero-shot counterparts, which we hypothesize has to do with the fact that the USMLE problems are long (long vignettes as well as long answers) which forced us to summarize the questions and answers when designing few-shot prompts (for which we used GPT-4). Hence, it’s possible that those prompts lead to short answers that in terms of BERT-sim score may differ more from the correct answer than a more wordy zero-shot generation. Across datasets Overall, we find that Med-Flamingo’s multimodal in-domain few-shot learning abilities lead to favorable generative VQA performance, leading to the lowest average rank of 1.67 in terms of clinical evaluation score as averaged across all evaluation datasets. As runner-up, OpenFlamingo zero-shot achieves a rank of 2.33. Qualitative analysis Finally, we showcase few examples of Med-Flamingo generations in more detail in Figures 1,5, and 6. Figure 5 exemplifies that a medical few-shot learner like Med-Flamingo can be prompted to generate rationale for its VQA answer. The shown example is impressive in that the rationale is visually guiding the reader towards the object of interest (calcification of the aortic wall). We note, however, that at this stage, few-shot multimodal prompted rationales may not be robust, especially when a model arrives at a wrong answer. Figures 1 and 6 showcase two example problems from the Visual USMLE dataset. The problem descriptions were slightly rephrased and summarized using GPT-4 for display. In Figure 6, MedFlamingo generates the correct answer while not mentioning the underlying diagnosis (urothelial cancer) as it was not asked for. By contrast, we observed baselines to directly diagnose the patient (instead of answering the actual question in a targeted way). The problem in Figure 1 illustrates that Med-Flamingo has the ability to integrate complex medical history information together with visual information to synthesize a comprehensive diagnosis that draws from the information of both modalities. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Authors: Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Works 2 Related Works 3 Med-Flamingo 3 Med-Flamingo 4 Evaluation 4 Evaluation 5 Results 5 Results 6 Discussion, Acknowledgments, and References 6 Discussion, Acknowledgments, and References A Appendix A Appendix 5 RESULTS In our experiments, we focus on generative medical visual question answering (VQA). While recent medical VLMs predominantly performed VQA in a non-generative but rather discriminative manner (i.e., by scoring different answer choices), we believe that this ex-post classification to carry less clinical usefulness, than directly generating responses. On the other hand, generative VQA is more challenging to evaluate, as automated metrics suffer from significant limitations as they do not fully capture the domain-specific context. Thus, we perform a human evaluation study where clinical experts review model generations (blinded) and score them (between 0 and 10) in terms of clinical usefulness. Conventional VQA datasets Table 1 shows the results for VQA-RAD, the radiological VQA dataset for which we created custom splits to address leakage (see Section4). Med-Flamingo few-shot shows strong results, improving the clinical eval score by ∼ 20% over the best baseline. In this dataset, the auxiliary metrics are rather aligned with clinical preference. Finetuning the MedVINT baseline did not lead to improved performance on this dataset which may be due to its small size. MedVINT zero-shot outperforms the other zero-shot ablations which may be partially attributed to its instruction tuning step on PMC-VQA. Conventional VQA datasets Table 2 shows for the results for Path-VQA, the pathology VQA dataset. Compared to the other datasets, all models overall perform poorer on the Path-VQA dataset in terms of clinical evaluation score. We hypothesize that this has to do with the fact the models are not pre-trained on actual large-scale and fine-grained pathology image datasets, but only on a rather small amount of pathology literature (which may not be enough to achieve strong performance). For instance, Figure 3 shows that only a small fraction of our training data covers pathology. In the automated metrics (BERT-sim and exact-match), Med-Flamingo improves upon the OpenFlamingo baseline, however the overall quality does not improve (as seen in the clinical evaluation score). MedVINT was fine-tuned on a sizeable training split which results in strong automated metrics, but did not result in a clinical evaluation score that matches any Flamingo variant. Visual USMLE Table 3 shows the results for the Visual USMLE dataset. Med-Flamingo (fewshot) results in the clinically most preferrable generations, whereas OpenFlamingo (zero-shot) is a close runner-up. As the ground truth answers were rather lengthy paragraphs, exact match was not an informative metric (constant 0 for all methods). The few-shot prompted models lead to lower automated scores than their zero-shot counterparts, which we hypothesize has to do with the fact that the USMLE problems are long (long vignettes as well as long answers) which forced us to summarize the questions and answers when designing few-shot prompts (for which we used GPT-4). Hence, it’s possible that those prompts lead to short answers that in terms of BERT-sim score may differ more from the correct answer than a more wordy zero-shot generation. Visual USMLE Across datasets Overall, we find that Med-Flamingo’s multimodal in-domain few-shot learning abilities lead to favorable generative VQA performance, leading to the lowest average rank of 1.67 in terms of clinical evaluation score as averaged across all evaluation datasets. As runner-up, OpenFlamingo zero-shot achieves a rank of 2.33. Across datasets Qualitative analysis Finally, we showcase few examples of Med-Flamingo generations in more detail in Figures 1,5, and 6. Figure 5 exemplifies that a medical few-shot learner like Med-Flamingo can be prompted to generate rationale for its VQA answer. The shown example is impressive in that the rationale is visually guiding the reader towards the object of interest (calcification of the aortic wall). We note, however, that at this stage, few-shot multimodal prompted rationales may not be robust, especially when a model arrives at a wrong answer. Qualitative analysis Figures 1 and 6 showcase two example problems from the Visual USMLE dataset. The problem descriptions were slightly rephrased and summarized using GPT-4 for display. In Figure 6, MedFlamingo generates the correct answer while not mentioning the underlying diagnosis (urothelial cancer) as it was not asked for. By contrast, we observed baselines to directly diagnose the patient (instead of answering the actual question in a targeted way). The problem in Figure 1 illustrates that Med-Flamingo has the ability to integrate complex medical history information together with visual information to synthesize a comprehensive diagnosis that draws from the information of both modalities. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. available on arxiv

Med-Flamingo: a Multimodal Medical Few-shot Learner - Results

Authors:
(1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA;
(4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA;
(5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA;
(6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA;
(7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil;
(8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA;
(9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction 2 Related Works 3 Med-Flamingo 4 Evaluation 5 Results 6 Discussion, Acknowledgments, and References A Appendix 4 EVALUATION 4.1 AUTOMATIC EVALUATION Baselines To compare generative VQA abilities against the literature, we consider different variants of the following baselines: MedVINT Zhang et al. (2023b), a visual instruction-tuned VLM based on Llama. As this model was not designed to do few-shot learning (e.g. the image information is prepended to the overall input), we report two modes for MedVINT: zero-shot and fine-tuned, where the model was fine-tuned on the training split of the VQA dataset. Since the rather small Visual-USMLE dataset has no separate training split, we ommit the fine-tuned baseline for that dataset. We used the MedVInT-TD model with PMC-LLaMA and PMC-CLIP backbones.


OpenFlamingo Awadalla et al. (2023), a powerful VLM which was trained on generaldomain data, and which served as the base model to train Med-Flamingo. We report both zero-shot and few-shot performance. We expect Flamingo-type models to shine in the fewshot setting which they are designed for (as already the pre-training task includes multiple interleaved image-text examples). Evaluation datasets To evaluate our model and compare it against the baselines, we leverage two existing VQA datasets from the medical domain (VQA-RAD and PathVQA). Upon closer inspection of the VQA-RAD dataset, we identified severe data leakage in the official train / test splits, which is problematic given that many recent VLMs fine-tune on the train split. To address this, we created a custom train / test split by seperately splitting images and questions (each 90% / 10%) to ensure that no image or question of the train split leaks into the test split. On these datasets, 6 shots were used for few-shot. Furthermore, we create Visual USMLE, a challenging multimodal problem set of 618 USMLE-style questions which are not only augmented with images but also with a case vignette and potentially tables of laboratory measurements. The Visual USMLE dataset was created by adapting problems from the Amboss platform (using licenced user access). To make the Visual USMLE problems more actionable and useful, we rephrased the problems to be open-ended instead of multiple-choice. This makes the benchmark harder and more realistic, as the models have to come up with differential diagnoses and potential procedures completely on their own—as opposed to selecting the most reasonable answer choice from few choices. Figure 8 gives an overview of the broad range of specialties that are covered in the dataset, greatly extending existing medical VQA datasets which are narrowly focused on radiology and pathology. For this comparatively small dataset, instead of creating a training split for finetuning, we created a small train split of 10 problems which can be used for few-shot prompting. For this dataset (with considerably longer problems and answers), we used only 4 shots to fit in the context window. Evaluation metrics Previous works in medical vision-language modelling typically focused scoring all available answers of a VQA dataset to arrive at a classification accuracy. However, since we are interested in generative VQA (as opposed to post-hoc scoring different potential answers), for sake of clinical utility, we employ the following evaluation metrics that directly assess the quality of the generated answer: Clinical evaluation score, as rated by three medical doctors (including one board-certified radiologist) using a human evaluation app that we developed for this study. More details are provided in Section 4.2.


BERT similarity score (BERT-sim), the F1 BERT score between the generated answer and the correct answer Zhang et al. (2020).


Exact-match, the fraction of generated answers that exactly match (modulo punctuation) the correct answer. This metric is rather noisy and conservative as useful answers may not lexically match the correct answer. 4.2 HUMAN EVALUATION We implemented a human evaluation app using Streamlit to visually display the generative VQA problems for clinical experts to rate the quality of the generated answers with scores from 0 to 10. Figure 4 shows an examplary view of the app. For each VQA problem, the raters are provided with the image, the question, the correct answer, and a set of blinded generations (e.g., appearing as ”prediction 1” in Figure 4), that appear in randomized order. 4.3 DEDUPLICATION AND LEAKAGE During the evaluation of the Med-Flamingo model, we were concerned that there may be leakage between the pre-training datasets (PMC-OA and MTB) and the down-stream VQA datasets used for evaluation; this could inflate judgements of model quality, as the model could memorize image-question-answer triples. To alleviate this concern, we performed data deduplication based upon pairwise similarity between images from our pre-training datasets and the images from our evaluation benchmarks. To detect similar images, in spite of perturbations due to cropping, color shifts, size, etc, we embedded the images using Google’s Vision Transformer, preserving the last hidden state as the resultant embedding Dosovitskiy et al. (2021). We then found the k-nearest neighbors to each evaluation image from amongst the pre-training images (using the FAISS library) Johnson et al. (2019). We then sorted and visualized image-image pairs by least euclidean distance; we found that images might be duplicates until a pairwise distance of around 80; beyond this point, there were no duplicates. This process revealed that the pretraining datasets leaked into the PVQA evaluation benchmark. Out of 6700 total images in PVQA test set, we judged 194 to be highly similar to images in the pretraining datasets, and thus, we removed them from our down-stream evaluation. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Authors: Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Works 2 Related Works 3 Med-Flamingo 3 Med-Flamingo 4 Evaluation 4 Evaluation 5 Results 5 Results 6 Discussion, Acknowledgments, and References 6 Discussion, Acknowledgments, and References A Appendix A Appendix 4 EVALUATION 4.1 AUTOMATIC EVALUATION Baselines To compare generative VQA abilities against the literature, we consider different variants of the following baselines: Baselines MedVINT Zhang et al. (2023b), a visual instruction-tuned VLM based on Llama. As this model was not designed to do few-shot learning (e.g. the image information is prepended to the overall input), we report two modes for MedVINT: zero-shot and fine-tuned, where the model was fine-tuned on the training split of the VQA dataset. Since the rather small Visual-USMLE dataset has no separate training split, we ommit the fine-tuned baseline for that dataset. We used the MedVInT-TD model with PMC-LLaMA and PMC-CLIP backbones. OpenFlamingo Awadalla et al. (2023), a powerful VLM which was trained on generaldomain data, and which served as the base model to train Med-Flamingo. We report both zero-shot and few-shot performance. We expect Flamingo-type models to shine in the fewshot setting which they are designed for (as already the pre-training task includes multiple interleaved image-text examples). MedVINT Zhang et al. (2023b), a visual instruction-tuned VLM based on Llama. As this model was not designed to do few-shot learning (e.g. the image information is prepended to the overall input), we report two modes for MedVINT: zero-shot and fine-tuned, where the model was fine-tuned on the training split of the VQA dataset. Since the rather small Visual-USMLE dataset has no separate training split, we ommit the fine-tuned baseline for that dataset. We used the MedVInT-TD model with PMC-LLaMA and PMC-CLIP backbones. MedVINT Zhang et al. (2023b), a visual instruction-tuned VLM based on Llama. As this model was not designed to do few-shot learning (e.g. the image information is prepended to the overall input), we report two modes for MedVINT: zero-shot and fine-tuned, where the model was fine-tuned on the training split of the VQA dataset. Since the rather small Visual-USMLE dataset has no separate training split, we ommit the fine-tuned baseline for that dataset. We used the MedVInT-TD model with PMC-LLaMA and PMC-CLIP backbones. OpenFlamingo Awadalla et al. (2023), a powerful VLM which was trained on generaldomain data, and which served as the base model to train Med-Flamingo. We report both zero-shot and few-shot performance. We expect Flamingo-type models to shine in the fewshot setting which they are designed for (as already the pre-training task includes multiple interleaved image-text examples). OpenFlamingo Awadalla et al. (2023), a powerful VLM which was trained on generaldomain data, and which served as the base model to train Med-Flamingo. We report both zero-shot and few-shot performance. We expect Flamingo-type models to shine in the fewshot setting which they are designed for (as already the pre-training task includes multiple interleaved image-text examples). Evaluation datasets To evaluate our model and compare it against the baselines, we leverage two existing VQA datasets from the medical domain (VQA-RAD and PathVQA). Upon closer inspection of the VQA-RAD dataset, we identified severe data leakage in the official train / test splits, which is problematic given that many recent VLMs fine-tune on the train split. To address this, we created a custom train / test split by seperately splitting images and questions (each 90% / 10%) to ensure that no image or question of the train split leaks into the test split. On these datasets, 6 shots were used for few-shot. Evaluation datasets Furthermore, we create Visual USMLE, a challenging multimodal problem set of 618 USMLE-style questions which are not only augmented with images but also with a case vignette and potentially tables of laboratory measurements. The Visual USMLE dataset was created by adapting problems from the Amboss platform (using licenced user access). To make the Visual USMLE problems more actionable and useful, we rephrased the problems to be open-ended instead of multiple-choice. This makes the benchmark harder and more realistic, as the models have to come up with differential diagnoses and potential procedures completely on their own—as opposed to selecting the most reasonable answer choice from few choices. Figure 8 gives an overview of the broad range of specialties that are covered in the dataset, greatly extending existing medical VQA datasets which are narrowly focused on radiology and pathology. For this comparatively small dataset, instead of creating a training split for finetuning, we created a small train split of 10 problems which can be used for few-shot prompting. For this dataset (with considerably longer problems and answers), we used only 4 shots to fit in the context window. Evaluation metrics Previous works in medical vision-language modelling typically focused scoring all available answers of a VQA dataset to arrive at a classification accuracy. However, since we are interested in generative VQA (as opposed to post-hoc scoring different potential answers), for sake of clinical utility, we employ the following evaluation metrics that directly assess the quality of the generated answer: Evaluation metrics Clinical evaluation score, as rated by three medical doctors (including one board-certified radiologist) using a human evaluation app that we developed for this study. More details are provided in Section 4.2. BERT similarity score (BERT-sim), the F1 BERT score between the generated answer and the correct answer Zhang et al. (2020). Exact-match, the fraction of generated answers that exactly match (modulo punctuation) the correct answer. This metric is rather noisy and conservative as useful answers may not lexically match the correct answer. Clinical evaluation score, as rated by three medical doctors (including one board-certified radiologist) using a human evaluation app that we developed for this study. More details are provided in Section 4.2. Clinical evaluation score, as rated by three medical doctors (including one board-certified radiologist) using a human evaluation app that we developed for this study. More details are provided in Section 4.2. BERT similarity score (BERT-sim), the F1 BERT score between the generated answer and the correct answer Zhang et al. (2020). BERT similarity score (BERT-sim), the F1 BERT score between the generated answer and the correct answer Zhang et al. (2020). Exact-match, the fraction of generated answers that exactly match (modulo punctuation) the correct answer. This metric is rather noisy and conservative as useful answers may not lexically match the correct answer. Exact-match, the fraction of generated answers that exactly match (modulo punctuation) the correct answer. This metric is rather noisy and conservative as useful answers may not lexically match the correct answer. 4.2 HUMAN EVALUATION We implemented a human evaluation app using Streamlit to visually display the generative VQA problems for clinical experts to rate the quality of the generated answers with scores from 0 to 10. Figure 4 shows an examplary view of the app. For each VQA problem, the raters are provided with the image, the question, the correct answer, and a set of blinded generations (e.g., appearing as ”prediction 1” in Figure 4), that appear in randomized order. 4.3 DEDUPLICATION AND LEAKAGE During the evaluation of the Med-Flamingo model, we were concerned that there may be leakage between the pre-training datasets (PMC-OA and MTB) and the down-stream VQA datasets used for evaluation; this could inflate judgements of model quality, as the model could memorize image-question-answer triples. To alleviate this concern, we performed data deduplication based upon pairwise similarity between images from our pre-training datasets and the images from our evaluation benchmarks. To detect similar images, in spite of perturbations due to cropping, color shifts, size, etc, we embedded the images using Google’s Vision Transformer, preserving the last hidden state as the resultant embedding Dosovitskiy et al. (2021). We then found the k-nearest neighbors to each evaluation image from amongst the pre-training images (using the FAISS library) Johnson et al. (2019). We then sorted and visualized image-image pairs by least euclidean distance; we found that images might be duplicates until a pairwise distance of around 80; beyond this point, there were no duplicates. This process revealed that the pretraining datasets leaked into the PVQA evaluation benchmark. Out of 6700 total images in PVQA test set, we judged 194 to be highly similar to images in the pretraining datasets, and thus, we removed them from our down-stream evaluation. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. available on arxiv

Med-Flamingo: a Multimodal Medical Few-shot Learner - Evaluation

Authors:
(1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA;
(4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA;
(5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA;
(6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA;
(7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil;
(8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA;
(9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction 2 Related Works 3 Med-Flamingo 4 Evaluation 5 Results 6 Discussion, Acknowledgments, and References A Appendix 3 MED-FLAMINGO To train a Flamingo model adapted to the medical domain, we leverage the pre-trained OpenFlamingo9B model checkpoint Awadalla et al. (2023), which is a general-domain VLM that was built on top of the frozen language model LLaMA-7B Touvron et al. (2023) and frozen vision encoder CLIP ViT/L-14 Radford et al.. We perform continued pre-training in the medical domain which results in the model we refer to as Med-Flamingo. 3.1 DATA We pre-train Med-Flamingo by jointly training on interleaved image-text data and paired image-text data. As for the interleaved dataset, we created a interleaved dataset from a set of medical textbooks, which we subsequently refer to as MTB. As for the paired datasets, we used PMC-OA Lin et al. (2023). MTB We construct a new multimodal dataset from a set of 4 721 textbooks from different medical specialties (see Figure 3). During preprocessing, each book is first converted from PDF to HTML with all tags removed, except the image tags are converted to  tokens. We then carry out data cleaning via deduplication and content filtering. Finally, each book with cleaned text and images is then chopped into segments for pretraining so that each segment contains at least one image and up to 10 images and a maximum length. In total, MTB consists of approximately 0.8M images and 584M tokens. We use 95% of the data for training and 5% of the data for evaluation during the pre-training. We construct a new multimodal dataset from a set of 4 721 textbooks from different medica specialties (see Figure 3). During preprocessing, each book is first converted from PDF to HTML with all tags removed, except the image tags are converted to tokens. We then carry out data cleaning via deduplication and content filtering. Finally, each book with cleaned text and images is then chopped into segments for pretraining so that each segment contains at least one image and up to 10 images and a maximum length. In total, MTB consists of approximately 0.8M images and 584M tokens. We use 95% of the data for training and 5% of the data for evaluation during the pre-training. PMC-OA We adopt the PMC-OA dataset Lin et al. (2023) which is a biomedical dataset with 1.6M image-caption pairs collected from PubMedCentral’s OpenAccess subset. We use 1.3M image-caption pairs for training and 0.16M pairs for evaluation following the public split[2]. 3.2 OBJECTIVES We follow the original We follow the original Flamingo model approach Alayrac et al., which considers the following language modelling problem: 3.3 TRAINING We performed multi-gpu training on a single node with 8x 80GB NVIDIA A100 GPUs. We trained the model using DeepSpeed ZeRO Stage 2: Optimizer states and gradients are sharded across devices. To further reduce memory load, we employed the 8-bit AdamW optimizer as well as the memory-efficient attention implementation of PyTorch 2.0. Med-Flamingo was initialized at the checkpoint of the Open-Flamingo model and then pre-trained for 2700 steps (or 6.75 days in wall time, including the validation steps), using 50 gradient accumulation steps and a per-device batch size of 1, resulting in a total batch size of 400. The model has 1.3B trainable parameters (gated cross attention layers and perceiver layers) and roughly 7B frozen parameters (decoder layers and vision encoder), which results in a total of 8.3B parameters. Note that this is the same number parameters as in the OpenFlamingo-9B model (version 1). This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. [2] https://huggingface.co/datasets/axiong/pmc_oa_beta Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Authors: Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Works 2 Related Works 3 Med-Flamingo 3 Med-Flamingo 4 Evaluation 4 Evaluation 5 Results 5 Results 6 Discussion, Acknowledgments, and References 6 Discussion, Acknowledgments, and References A Appendix A Appendix 3 MED-FLAMINGO To train a Flamingo model adapted to the medical domain, we leverage the pre-trained OpenFlamingo9B model checkpoint Awadalla et al. (2023), which is a general-domain VLM that was built on top of the frozen language model LLaMA-7B Touvron et al. (2023) and frozen vision encoder CLIP ViT/L-14 Radford et al.. We perform continued pre-training in the medical domain which results in the model we refer to as Med-Flamingo. 3.1 DATA We pre-train Med-Flamingo by jointly training on interleaved image-text data and paired image-text data. As for the interleaved dataset, we created a interleaved dataset from a set of medical textbooks, which we subsequently refer to as MTB. As for the paired datasets, we used PMC-OA Lin et al. (2023). MTB We construct a new multimodal dataset from a set of 4 721 textbooks from different medical specialties (see Figure 3). During preprocessing, each book is first converted from PDF to HTML with all tags removed, except the image tags are converted to  tokens. We then carry out data cleaning via deduplication and content filtering. Finally, each book with cleaned text and images is then chopped into segments for pretraining so that each segment contains at least one image and up to 10 images and a maximum length. In total, MTB consists of approximately 0.8M images and 584M tokens. We use 95% of the data for training and 5% of the data for evaluation during the pre-training. We construct a new multimodal dataset from a set of 4 721 textbooks from different medica specialties (see Figure 3). During preprocessing, each book is first converted from PDF to HTML with all tags removed, except the image tags are converted to tokens. We then carry out data cleaning via deduplication and content filtering. Finally, each book with cleaned text and images is then chopped into segments for pretraining so that each segment contains at least one image and up to 10 images and a maximum length. In total, MTB consists of approximately 0.8M images and 584M tokens. We use 95% of the data for training and 5% of the data for evaluation during the pre-training. MTB PMC-OA We adopt the PMC-OA dataset Lin et al. (2023) which is a biomedical dataset with 1.6M image-caption pairs collected from PubMedCentral’s OpenAccess subset. We use 1.3M image-caption pairs for training and 0.16M pairs for evaluation following the public split[2]. PMC-OA 3.2 OBJECTIVES We follow the original We follow the original Flamingo model approach Alayrac et al., which considers the following language modelling problem: 3.3 TRAINING We performed multi-gpu training on a single node with 8x 80GB NVIDIA A100 GPUs. We trained the model using DeepSpeed ZeRO Stage 2: Optimizer states and gradients are sharded across devices. To further reduce memory load, we employed the 8-bit AdamW optimizer as well as the memory-efficient attention implementation of PyTorch 2.0. Med-Flamingo was initialized at the checkpoint of the Open-Flamingo model and then pre-trained for 2700 steps (or 6.75 days in wall time, including the validation steps), using 50 gradient accumulation steps and a per-device batch size of 1, resulting in a total batch size of 400. The model has 1.3B trainable parameters (gated cross attention layers and perceiver layers) and roughly 7B frozen parameters (decoder layers and vision encoder), which results in a total of 8.3B parameters. Note that this is the same number parameters as in the OpenFlamingo-9B model (version 1). This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. available on arxiv [2] https://huggingface.co/datasets/axiong/pmc_oa_beta

Med-Flamingo: a Multimodal Medical Few-shot Learner - Med-Flamingo

Authors:
(1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA;
(4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA;
(5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA;
(6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA;
(7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil;
(8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA;
(9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction 2 Related Works 3 Med-Flamingo 4 Evaluation 5 Results 6 Discussion, Acknowledgments, and References A Appendix 2 RELATED WORKS The success of large language models (LLMs) Brown et al.; Liang et al. (2022); Qin et al. (2023) has led to significant advancements in training specialized models for the medical domain. This has resulted in the emergence of various models, including BioBERT Lee et al. (2020), ClinicalBERT Huang et al. (2019), PubMedBERT Gu et al. (2021), BioLinkBERT Yasunaga et al. (b), DRAGON Yasunaga et al. (a), BioMedLM Bolton et al., BioGPT Luo et al. (2022), and Med-PaLM Singhal et al.. Although these medical language models are typically smaller than general-purpose LLMs like GPT-3 Brown et al., they can match or even surpass their performance on medical tasks, such as medical question answering. Recently, there has been a growing interest in extending language models to handle vision-language multimodal data and tasks Su et al. (2019); Ramesh et al.; Alayrac et al. (2022); Aghajanyan et al.; Yasunaga et al. (2023). Furthermore, many medical applications involve multimodal information, such as radiology tasks that require the analysis of both X-ray images and radiology reports Tiu et al. (2022). Motivated by these factors, we present a medical vision-language model (VLM). Existing medical VLMs include BiomedCLIP Zhang et al. (2023a), MedVINT Zhang et al. (2023b). While BiomedCLIP is an encoder-only model, our focus lies in developing a generative VLM, demonstrating superior performance compared to MedVINT. Finally, Llava-Med is another recent medical generative VLM Li et al. (2023), however the model was not yet available for benchmarking. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Authors: Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Works 2 Related Works 3 Med-Flamingo 3 Med-Flamingo 4 Evaluation 4 Evaluation 5 Results 5 Results 6 Discussion, Acknowledgments, and References 6 Discussion, Acknowledgments, and References A Appendix A Appendix 2 RELATED WORKS The success of large language models (LLMs) Brown et al.; Liang et al. (2022); Qin et al. (2023) has led to significant advancements in training specialized models for the medical domain. This has resulted in the emergence of various models, including BioBERT Lee et al. (2020), ClinicalBERT Huang et al. (2019), PubMedBERT Gu et al. (2021), BioLinkBERT Yasunaga et al. (b), DRAGON Yasunaga et al. (a), BioMedLM Bolton et al., BioGPT Luo et al. (2022), and Med-PaLM Singhal et al.. Although these medical language models are typically smaller than general-purpose LLMs like GPT-3 Brown et al., they can match or even surpass their performance on medical tasks, such as medical question answering. Recently, there has been a growing interest in extending language models to handle vision-language multimodal data and tasks Su et al. (2019); Ramesh et al.; Alayrac et al. (2022); Aghajanyan et al.; Yasunaga et al. (2023). Furthermore, many medical applications involve multimodal information, such as radiology tasks that require the analysis of both X-ray images and radiology reports Tiu et al. (2022). Motivated by these factors, we present a medical vision-language model (VLM). Existing medical VLMs include BiomedCLIP Zhang et al. (2023a), MedVINT Zhang et al. (2023b). While BiomedCLIP is an encoder-only model, our focus lies in developing a generative VLM, demonstrating superior performance compared to MedVINT. Finally, Llava-Med is another recent medical generative VLM Li et al. (2023), however the model was not yet available for benchmarking. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. available on arxiv

Med-Flamingo: a Multimodal Medical Few-shot Learner - Related Works

Authors:
(1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work;
(3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA;
(4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA;
(5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA;
(6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA;
(7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil;
(8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA;
(9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction 2 Related Works 3 Med-Flamingo 4 Evaluation 5 Results 6 Discussion, Acknowledgments, and References A Appendix ABSTRACT Medicine, by its nature, is a multifaceted domain that requires the synthesis of information across various modalities. Medical generative vision-language models (VLMs) make a first step in this direction and promise many exciting clinical applications. However, existing models typically have to be fine-tuned on sizeable down-stream datasets, which poses a significant limitation as in many medical applications data is scarce, necessitating models that are capable of learning from few examples in real-time. Here we propose Med-Flamingo, a multimodal few-shot learner adapted to the medical domain. Based on OpenFlamingo-9B, we continue pre-training on paired and interleaved medical image-text data from publications and textbooks. Med-Flamingo unlocks few-shot generative medical visual question answering (VQA) abilities, which we evaluate on several datasets including a novel challenging open-ended VQA dataset of visual USMLE-style problems. Furthermore, we conduct the first human evaluation for generative medical VQA where physicians review the problems and blinded generations in an interactive app. Med-Flamingo improves performance in generative medical VQA by up to 20% in clinician’s rating and firstly enables multimodal medical few-shot adaptations, such as rationale generation. We release our model, code, and evaluation app under https://github.com/snap-stanford/med-flamingo. 1 INTRODUCTION Large, pre-trained models (or foundation models) have demonstrated remarkable capabilities in solving an abundance of tasks by being provided only a few labeled examples as context Bommasani et al. (2021). This is known as in-context learning Brown et al. (2020), through which a model learns a task from a few provided examples specifically during prompting and without tuning the model parameters. In the medical domain, this bears great potential to vastly expand the capabilities of existing medical AI models Moor et al. (2023). Most notably, it will enable medical AI models to handle the various rare cases faced by clinicians every day in a unified way, to provide relevant rationales to justify their statements, and to easily customize model generations to specific use cases. Implementing the in-context learning capability in a medical setting is challenging due to the inherent complexity and multimodality of medical data and the diversity of tasks to be solved. Previous efforts to create multimodal medical foundation models, such as ChexZero Tiu et al. (2022) and BiomedCLIP Zhang et al. (2023a), have made significant strides in their respective domains. ChexZero specializes in chest X-ray interpretation, while BiomedCLIP has been trained on more diverse images paired with captions from the biomedical literature. Other models have also been developed for electronic health record (EHR) data Steinberg et al. (2021) and surgical videos Kiyasseh et al. (2023). However, none of these models have embraced in-context learning for the multimodal medical domain. Existing medical VLMs, such as MedVINT Zhang et al. (2023b), are typically trained on paired image-text data with a single image in the context, as opposed to more general streams of text that are interleaved with multiple images. Therefore, these models were not designed and tested to perform multimodal in-context learning with few-shot examples[1] Here, we propose Med-Flamingo, the first medical foundation model that can perform multimodal incontext learning specialized for the medical domain. Med-Flamingo is a vision-language model based on Flamingo (Alayrac et al., 2022) that can naturally ingest data with interleaved modalities (images and text), to generate text conditioned on this multimodal input. Building on the success of Flamingo, which was among the first vision-language models to exhibit in-context learning and few-shot learning abilities, Med-Flamingo extends these capabilities to the medical domain by pre-training on multimodal knowledge sources across medical disciplines. In preparation for the training of Med-Flamingo, our initial step involved constructing a unique, interleaved image-text dataset, which was derived from an extensive collection of over 4K medical textbooks (Section 3). Given the critical nature of accuracy and precision within the medical field, it is important to note that the quality, reliability, and source of the training data can considerably shape the results. Therefore, to ensure accuracy in medical facts, we meticulously curated our dataset from respected and authoritative sources of medical knowledge, as opposed to relying on potentially unreliable web-sourced data. In our experiments, we evaluate Med-Flamingo on generative medical visual question-answering (VQA) tasks by directly generating open-ended answers, as opposed to scoring artificial answer options ex post–as CLIP-based medical vision-language models do. We design a new realistic evaluation protocol to measure the model generations’ clinical usefulness. For this, we conduct an in-depth human evaluation study with clinical experts which results in a human evaluation score that serves as our main metric. In addition, due to existing medical VQA datasets being narrowly focused on image interpretation among the specialties of radiology and pathology, we create Visual USMLE, a challenging generative VQA dataset of complex USMLE-style problems across specialties, which are augmented with images, case vignettes, and potentially with lab results. Averaged across three generative medical VQA datasets, few-shot prompted Med-Flamingo achieves the best average rank in clinical evaluation score (rank of 1.67, best prior model has 2.33), indicating that the model generates answers that are most preferred by clinicians, with up to 20% improvement over prior models. Furthermore, Med-Flamingo is capable of performing medical reasoning, such as answering complex medical questions (such as visually grounded USMLE-style questions) and providing explanations (i.e., rationales), a capability not previously demonstrated by other multimodal medical foundation models. However, it is important to note that Med-Flamingo’s performance may be limited by the availability and diversity of training data, as well as the complexity of certain medical tasks. All investigated models and baselines would occasionally hallucinate or generate low-quality responses. Despite these limitations, our work represents a significant step forward in the development of multimodal medical foundation models and their ability to perform multimodal in-context learning in the medical domain. We release the Med-Flamingo-9B checkpoint for further research, and make our code available under https://github.com/snap-stanford/med-flamingo. In summary, our paper makes the following contributions: We present the first multimodal few-shot learner adapted to the medical domain, which promises novel clinical applications such as rationale generation and conditioning on retrieved multimodal context.


We create a novel dataset that enables the pre-training of a multimodal few-shot learner for the general medical domain.


We create a novel USMLE-style evaluation dataset that combines medical VQA with complex, across-specialty medical reasoning.


We highlight shortcomings of existing evaluation strategies, and conduct an in-depth clinical evaluation study of open-ended VQA generations with medical raters using a dedicated evaluation app. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. [1] For example, a challenge with multimodal in-context learning for existing medical vision language models is the potential for image information to leak across examples, potentially misleading the model. Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Authors: Authors: (1) Michael Moor, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (2) Qian Huang, Department of Computer Science, Stanford University, Stanford, USA and these authors contributed equally to this work; (3) Shirley Wu, Department of Computer Science, Stanford University, Stanford, USA; (4) Michihiro Yasunaga, Department of Computer Science, Stanford University, Stanford, USA; (5) Cyril Zakka, Department of Cardiothoracic Surgery, Stanford Medicine, Stanford, USA; (6) Yash Dalmia, Department of Computer Science, Stanford University, Stanford, USA; (7) Eduardo Pontes Reis, Hospital Israelita Albert Einstein, Sao Paulo, Brazil; (8) Pranav Rajpurkar, Department of Biomedical Informatics, Harvard Medical School, Boston, USA; (9) Jure Leskovec, Department of Computer Science, Stanford University, Stanford, USA. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Works 2 Related Works 3 Med-Flamingo 3 Med-Flamingo 4 Evaluation 4 Evaluation 5 Results 5 Results 6 Discussion, Acknowledgments, and References 6 Discussion, Acknowledgments, and References A Appendix A Appendix ABSTRACT Medicine, by its nature, is a multifaceted domain that requires the synthesis of information across various modalities. Medical generative vision-language models (VLMs) make a first step in this direction and promise many exciting clinical applications. However, existing models typically have to be fine-tuned on sizeable down-stream datasets, which poses a significant limitation as in many medical applications data is scarce, necessitating models that are capable of learning from few examples in real-time. Here we propose Med-Flamingo, a multimodal few-shot learner adapted to the medical domain. Based on OpenFlamingo-9B, we continue pre-training on paired and interleaved medical image-text data from publications and textbooks. Med-Flamingo unlocks few-shot generative medical visual question answering (VQA) abilities, which we evaluate on several datasets including a novel challenging open-ended VQA dataset of visual USMLE-style problems. Furthermore, we conduct the first human evaluation for generative medical VQA where physicians review the problems and blinded generations in an interactive app. Med-Flamingo improves performance in generative medical VQA by up to 20% in clinician’s rating and firstly enables multimodal medical few-shot adaptations, such as rationale generation. We release our model, code, and evaluation app under https://github.com/snap-stanford/med-flamingo. 1 INTRODUCTION Large, pre-trained models (or foundation models) have demonstrated remarkable capabilities in solving an abundance of tasks by being provided only a few labeled examples as context Bommasani et al. (2021). This is known as in-context learning Brown et al. (2020), through which a model learns a task from a few provided examples specifically during prompting and without tuning the model parameters. In the medical domain, this bears great potential to vastly expand the capabilities of existing medical AI models Moor et al. (2023). Most notably, it will enable medical AI models to handle the various rare cases faced by clinicians every day in a unified way, to provide relevant rationales to justify their statements, and to easily customize model generations to specific use cases. Implementing the in-context learning capability in a medical setting is challenging due to the inherent complexity and multimodality of medical data and the diversity of tasks to be solved. Previous efforts to create multimodal medical foundation models, such as ChexZero Tiu et al. (2022) and BiomedCLIP Zhang et al. (2023a), have made significant strides in their respective domains. ChexZero specializes in chest X-ray interpretation, while BiomedCLIP has been trained on more diverse images paired with captions from the biomedical literature. Other models have also been developed for electronic health record (EHR) data Steinberg et al. (2021) and surgical videos Kiyasseh et al. (2023). However, none of these models have embraced in-context learning for the multimodal medical domain. Existing medical VLMs, such as MedVINT Zhang et al. (2023b), are typically trained on paired image-text data with a single image in the context, as opposed to more general streams of text that are interleaved with multiple images. Therefore, these models were not designed and tested to perform multimodal in-context learning with few-shot examples[1] Here, we propose Med-Flamingo, the first medical foundation model that can perform multimodal incontext learning specialized for the medical domain. Med-Flamingo is a vision-language model based on Flamingo (Alayrac et al., 2022) that can naturally ingest data with interleaved modalities (images and text), to generate text conditioned on this multimodal input. Building on the success of Flamingo, which was among the first vision-language models to exhibit in-context learning and few-shot learning abilities, Med-Flamingo extends these capabilities to the medical domain by pre-training on multimodal knowledge sources across medical disciplines. In preparation for the training of Med-Flamingo, our initial step involved constructing a unique, interleaved image-text dataset, which was derived from an extensive collection of over 4K medical textbooks (Section 3). Given the critical nature of accuracy and precision within the medical field, it is important to note that the quality, reliability, and source of the training data can considerably shape the results. Therefore, to ensure accuracy in medical facts, we meticulously curated our dataset from respected and authoritative sources of medical knowledge, as opposed to relying on potentially unreliable web-sourced data. In our experiments, we evaluate Med-Flamingo on generative medical visual question-answering (VQA) tasks by directly generating open-ended answers, as opposed to scoring artificial answer options ex post–as CLIP-based medical vision-language models do. We design a new realistic evaluation protocol to measure the model generations’ clinical usefulness. For this, we conduct an in-depth human evaluation study with clinical experts which results in a human evaluation score that serves as our main metric. In addition, due to existing medical VQA datasets being narrowly focused on image interpretation among the specialties of radiology and pathology, we create Visual USMLE, a challenging generative VQA dataset of complex USMLE-style problems across specialties, which are augmented with images, case vignettes, and potentially with lab results. Averaged across three generative medical VQA datasets, few-shot prompted Med-Flamingo achieves the best average rank in clinical evaluation score (rank of 1.67, best prior model has 2.33), indicating that the model generates answers that are most preferred by clinicians, with up to 20% improvement over prior models. Furthermore, Med-Flamingo is capable of performing medical reasoning, such as answering complex medical questions (such as visually grounded USMLE-style questions) and providing explanations (i.e., rationales), a capability not previously demonstrated by other multimodal medical foundation models. However, it is important to note that Med-Flamingo’s performance may be limited by the availability and diversity of training data, as well as the complexity of certain medical tasks. All investigated models and baselines would occasionally hallucinate or generate low-quality responses. Despite these limitations, our work represents a significant step forward in the development of multimodal medical foundation models and their ability to perform multimodal in-context learning in the medical domain. We release the Med-Flamingo-9B checkpoint for further research, and make our code available under https://github.com/snap-stanford/med-flamingo. In summary, our paper makes the following contributions: We present the first multimodal few-shot learner adapted to the medical domain, which promises novel clinical applications such as rationale generation and conditioning on retrieved multimodal context. We create a novel dataset that enables the pre-training of a multimodal few-shot learner for the general medical domain. We create a novel USMLE-style evaluation dataset that combines medical VQA with complex, across-specialty medical reasoning. We highlight shortcomings of existing evaluation strategies, and conduct an in-depth clinical evaluation study of open-ended VQA generations with medical raters using a dedicated evaluation app. We present the first multimodal few-shot learner adapted to the medical domain, which promises novel clinical applications such as rationale generation and conditioning on retrieved multimodal context. We present the first multimodal few-shot learner adapted to the medical domain, which promises novel clinical applications such as rationale generation and conditioning on retrieved multimodal context. We create a novel dataset that enables the pre-training of a multimodal few-shot learner for the general medical domain. We create a novel dataset that enables the pre-training of a multimodal few-shot learner for the general medical domain. We create a novel USMLE-style evaluation dataset that combines medical VQA with complex, across-specialty medical reasoning. We create a novel USMLE-style evaluation dataset that combines medical VQA with complex, across-specialty medical reasoning. We highlight shortcomings of existing evaluation strategies, and conduct an in-depth clinical evaluation study of open-ended VQA generations with medical raters using a dedicated evaluation app. We highlight shortcomings of existing evaluation strategies, and conduct an in-depth clinical evaluation study of open-ended VQA generations with medical raters using a dedicated evaluation app. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. available on arxiv [1] For example, a challenge with multimodal in-context learning for existing medical vision language models is the potential for image information to leak across examples, potentially misleading the model.

Med-Flamingo: a Multimodal Medical Few-shot Learner - Abstract and Introduction 

I grew interested in how the 2FA usually works for quite some time, and of course, in a data-driven society, security is key to everything, right? So I thought of maybe developing one of my own and named it ProAuth. Well, it’s not as ideal as someone would hope it to be(I mean feature-like), but for security, it’s…. what could I say SUPERLATIVE! I’m sure you’ll love it. Here’s a recap of how it works: You log in with your Google account, and you’ll be prompted to a dynamic code generator page where unique codes will be generated every 60 seconds. Amazing! I know right? I’m afraid I didn’t have anyone to join hands with in the build-up process(I know it would’ve been great). Purely, I was targeting organizations that still don’t have their own internal 2FA verification system. If this is what your organization lacks then Viola! Welcome. You’d be surprised how easy it is to navigate through my app. First, the user is prompted to sign in with Google, and by the way, that’s just another secure user authentication. Once allowed in, the backend is just like I mentioned: time-sensitive code is generated(and an API for validating generated codes). The code is then stored in a global variable (memory), and thereafter, the API endpoint is used for code verification to ensure it matches. You must have noticed that I used Python (flask) for the entire backend process, with a bit of JS script spiced inside HTML and CSS for the frontend processing and code display. Well, I don’t have to mention the authentication again, right? (it’s Google). For serving this application, I chose WSGI HTTP, which we both know is well-suited and relied on by most, if not all, flask apps. Nginx was just for taking care of reverse proxy and handling incoming HTTP requests, serving static contents, and managing SSL/TLS. Oooh! I used Cerbot to obtain the SSL certificate so that all connections would appear secure and trustworthy on the internet. I encountered a significant technical challenge that revolved around deploying the app to the server and integrating Google OAuth for user authentication. The challenge became apparent when I repeatedly faced "Resource temporarily unavailable" errors while attempting to SSH into the remote server. Furthermore, attempts to set up OAuth resulted in Google's OAuth service blocking requests, citing security concerns. After many attempts and consultations with the mentors, it was discovered that the problem stemmed from a network misconfiguration at the server level. The server was incorrectly routed, causing intermittent connectivity issues. Once the server's network settings were corrected, SSH access was restored. From this project, what I learned as an engineer is the importance of resilience and a methodical approach to solving problems, especially those that involve network settings that can be daunting. It has informed my engineering path to adopt an iterative development approach with continuous testing in environments that closely mimic production, which will be crucial to identifying and resolving issues at an early stage. Generally about me, I am a dedicated engineer with a passion for creating secure, user-focused web applications.  My background focuses  on both front and backend development. I thrive in solving complex problems and continuously seek  to enhance my technical skills through hands-on projects and collaboration. Beyond coding, I enjoy learning about new technologies and sharing knowledge with my peers, and oh! I love music; perhaps someday I’ll develop an app for this. GitHub repo: https://github.com/Carey99/full_proauth
Deployed Project Page: https://proauth.awiki.org
LinkedIn Profile: https://www.linkedin.com/in/cery-akida-a619911a8/ Feel free to explore the project and connect with me for further discussions or collaborations! I grew interested in how the 2FA usually works for quite some time, and of course, in a data-driven society, security is key to everything, right? So I thought of maybe developing one of my own and named it ProAuth. Well, it’s not as ideal as someone would hope it to be(I mean feature-like), but for security, it’s…. what could I say SUPERLATIVE! I’m sure you’ll love it. Here’s a recap of how it works: You log in with your Google account, and you’ll be prompted to a dynamic code generator page where unique codes will be generated every 60 seconds. Amazing! I know right? I’m afraid I didn’t have anyone to join hands with in the build-up process(I know it would’ve been great). Purely, I was targeting organizations that still don’t have their own internal 2FA verification system. If this is what your organization lacks then Viola! Welcome. You’d be surprised how easy it is to navigate through my app. First, the user is prompted to sign in with Google, and by the way, that’s just another secure user authentication. Once allowed in, the backend is just like I mentioned: time-sensitive code is generated(and an API for validating generated codes). The code is then stored in a global variable (memory), and thereafter, the API endpoint is used for code verification to ensure it matches. You must have noticed that I used Python (flask) for the entire backend process, with a bit of JS script spiced inside HTML and CSS for the frontend processing and code display. Well, I don’t have to mention the authentication again, right? ( it’s Google ). For serving this application, I chose WSGI HTTP, which we both know is well-suited and relied on by most, if not all, flask apps. Nginx was just for taking care of reverse proxy and handling incoming HTTP requests, serving static contents, and managing SSL/TLS. Oooh! I used Cerbot to obtain the SSL certificate so that all connections would appear secure and trustworthy on the internet. it’s Google I encountered a significant technical challenge that revolved around deploying the app to the server and integrating Google OAuth for user authentication. The challenge became apparent when I repeatedly faced " Resource temporarily unavailable " errors while attempting to SSH into the remote server. Furthermore, attempts to set up OAuth resulted in Google's OAuth service blocking requests, citing security concerns. Resource temporarily unavailable After many attempts and consultations with the mentors, it was discovered that the problem stemmed from a network misconfiguration at the server level. The server was incorrectly routed, causing intermittent connectivity issues. Once the server's network settings were corrected, SSH access was restored. From this project, what I learned as an engineer is the importance of resilience and a methodical approach to solving problems, especially those that involve network settings that can be daunting. It has informed my engineering path to adopt an iterative development approach with continuous testing in environments that closely mimic production, which will be crucial to identifying and resolving issues at an early stage. Generally about me, I am a dedicated engineer with a passion for creating secure, user-focused web applications.  My background focuses  on both front and backend development. I thrive in solving complex problems and continuously seek  to enhance my technical skills through hands-on projects and collaboration. Beyond coding, I enjoy learning about new technologies and sharing knowledge with my peers, and oh! I love music; perhaps someday I’ll develop an app for this. GitHub repo: https://github.com/Carey99/full_proauth Deployed Project Page: https://proauth.awiki.org LinkedIn Profile: https://www.linkedin.com/in/cery-akida-a619911a8/ GitHub repo: https://github.com/Carey99/full_proauth https://github.com/Carey99/full_proauth Deployed Project Page: https://proauth.awiki.org https://proauth.awiki.org LinkedIn Profile: https://www.linkedin.com/in/cery-akida-a619911a8/ https://www.linkedin.com/in/cery-akida-a619911a8/ Feel free to explore the project and connect with me for further discussions or collaborations!

ProAuth: How I Used Flask and OAuth to Develop a Dynamic Code Authentication App

For startups and midsize organizations, flat structures allow you to be efficient and innovative Some people say flat structures cause chaos in organizations. And while that can certainly be true, my experience has been different. After years of working for horizontal and vertical hierarchies, I realized I prefer the former. At Nokia, for example, agreeing on the simplest matters involved several offices and many meetings. Since I’ve always wanted to found a tech company, when I did it back in 2005, I opted for horizontal management. Ten years later, I started another company with the same organizational style, and I never looked back. While flat organizational structures are not for everyone, they do work. In this article, I’ll explain what makes a successful flat organization structure, who thrives in it, and share examples of positive business outcomes. Why it works Flat structure organizations allow employees to have a say in how they do the work and provide more opportunities to showcase their talents. And that’s exactly what motivates people from within. According to self-determination theory, all of us have innate psychological needs for autonomy, competence, and relatedness. And we can fulfill those needs, we feel happier and can sustain our motivation for a long time. Looking back, I see how the principles of autonomy, competence, and relatedness are reflected in our corporate culture. These might be the very foundations of a successful flat-structure organization. Let’s dig deeper and see how each principle contributes. Autonomy Flat structure organizations give people much more autonomy in their jobs. Autonomy leads to increased accountability, making employees feel more important and involved in their company. While it’s not always possible to allow employees to choose tasks (after all, someone has to do the grunt work) or the people to work with, we can give them freedom over how and when they complete them. I’m not the leader who tells employees how to do their jobs: they’re the experts, so they should know better. As long as they can deliver results, I don’t care what their workflow looks like. I also believe employees’ schedules should be flexible. Let them work when they’re most productive or have an opportunity. They are humans. They may need that slow morning walk with their dog to improve their mental health. Competence The human urge to improve and learn is another intrinsic motivator. People are motivated by the sense of accomplishment and progress they experience as they develop their skills. Because a flat structure organization has few levels of middle management or none, it’s easier to recognize team members’ talents. As a result, managers can assign projects that speak to the team members’ strengths while encouraging growth. Employees can also easily champion their ideas. An engineer who has been with us for several years felt a little stagnated in her role. She proposed sharing her testing expertise on YouTube to HR. The marketing team jumped in on the same day. By the next week, we released our first video. This wouldn’t have been possible if our organization was an unwieldy bureaucratic machine where teammates were reluctant to promote ideas from the bottom up. Relatedness Fulfilling employees’ needs for relatedness  is about making them feel supported, appreciated, and involved. That’s when they feel like what they do is significant to the rest of the team and organization. Unlike hierarchical models, flat organizational structures eliminate rigid divisions, encouraging open communication and collaboration. Team members feel valued as contributors, not cogs in a machine. This fosters a sense of belonging and teamwork. How we do it It’s important to note that a flat structure doesn’t work for everyone. It does for us, but we do have to have the following in place: Crystal clear job titles and descriptions
There’s no confusion about whom to contact and who’s accountable for what. We’re not a holacracy where roles are flexible. While there are no department heads at Redwerk, we do have some employees in managerial positions. This helps keep things in order and reduces the number of people messaging me, the founder, for trivial matters.


Well-documented workflows
While they can change with new people bringing fresh ideas, it’s important to document these innovations and ensure there’s an up-to-date and easily accessible resource hub.


Being selective about who we hire
Hiring becomes more difficult because you need experts with particular soft skills, such as self-reliance, initiative, and superb organization. When we hire, we look for those who are comfortable managing their time and workload without a supervisor constantly checking in. Final Thoughts Flat organization structures are a great fit for startups and midsize businesses that want to be fast and innovative. They offer a natural breeding ground for the principles of autonomy, competence, and relatedness that are key for keeping your employees committed long-term. But they come with a catch, and it’s up to you to tackle or forgo it. Image source: Cine Insomnia/Pexels For startups and midsize organizations, flat structures allow you to be efficient and innovative Some people say flat structures cause chaos in organizations. And while that can certainly be true, my experience has been different. After years of working for horizontal and vertical hierarchies, I realized I prefer the former. At Nokia, for example, agreeing on the simplest matters involved several offices and many meetings. Since I’ve always wanted to found a tech company, when I did it back in 2005, I opted for horizontal management. Ten years later, I started another company with the same organizational style, and I never looked back. While flat organizational structures are not for everyone, they do work. In this article, I’ll explain what makes a successful flat organization structure, who thrives in it, and share examples of positive business outcomes. Why it works Flat structure organizations allow employees to have a say in how they do the work and provide more opportunities to showcase their talents. And that’s exactly what motivates people from within. According to self-determination theory, all of us have innate psychological needs for autonomy, competence, and relatedness. And we can fulfill those needs, we feel happier and can sustain our motivation for a long time. Looking back, I see how the principles of autonomy, competence, and relatedness are reflected in our corporate culture. These might be the very foundations of a successful flat-structure organization. Let’s dig deeper and see how each principle contributes. Autonomy Flat structure organizations give people much more autonomy in their jobs. Autonomy leads to increased accountability, making employees feel more important and involved in their company. While it’s not always possible to allow employees to choose tasks (after all, someone has to do the grunt work) or the people to work with, we can give them freedom over how and when they complete them. I’m not the leader who tells employees how to do their jobs: they’re the experts, so they should know better. As long as they can deliver results, I don’t care what their workflow looks like. I also believe employees’ schedules should be flexible. Let them work when they’re most productive or have an opportunity. They are humans. They may need that slow morning walk with their dog to improve their mental health. Competence The human urge to improve and learn is another intrinsic motivator. People are motivated by the sense of accomplishment and progress they experience as they develop their skills. Because a flat structure organization has few levels of middle management or none, it’s easier to recognize team members’ talents. As a result, managers can assign projects that speak to the team members’ strengths while encouraging growth. Employees can also easily champion their ideas. An engineer who has been with us for several years felt a little stagnated in her role. She proposed sharing her testing expertise on YouTube to HR. The marketing team jumped in on the same day. By the next week, we released our first video. This wouldn’t have been possible if our organization was an unwieldy bureaucratic machine where teammates were reluctant to promote ideas from the bottom up. Relatedness Fulfilling employees’ needs for relatedness  is about making them feel supported, appreciated, and involved. That’s when they feel like what they do is significant to the rest of the team and organization. Unlike hierarchical models, flat organizational structures eliminate rigid divisions, encouraging open communication and collaboration. Team members feel valued as contributors, not cogs in a machine. This fosters a sense of belonging and teamwork. How we do it It’s important to note that a flat structure doesn’t work for everyone. It does for us, but we do have to have the following in place: Crystal clear job titles and descriptions
There’s no confusion about whom to contact and who’s accountable for what. We’re not a holacracy where roles are flexible. While there are no department heads at Redwerk, we do have some employees in managerial positions. This helps keep things in order and reduces the number of people messaging me, the founder, for trivial matters. Well-documented workflows
While they can change with new people bringing fresh ideas, it’s important to document these innovations and ensure there’s an up-to-date and easily accessible resource hub. Being selective about who we hire
Hiring becomes more difficult because you need experts with particular soft skills, such as self-reliance, initiative, and superb organization. When we hire, we look for those who are comfortable managing their time and workload without a supervisor constantly checking in. Crystal clear job titles and descriptions
There’s no confusion about whom to contact and who’s accountable for what. We’re not a holacracy where roles are flexible. While there are no department heads at Redwerk, we do have some employees in managerial positions. This helps keep things in order and reduces the number of people messaging me, the founder, for trivial matters. Crystal clear job titles and descriptions There’s no confusion about whom to contact and who’s accountable for what. We’re not a holacracy where roles are flexible. While there are no department heads at Redwerk, we do have some employees in managerial positions. This helps keep things in order and reduces the number of people messaging me, the founder, for trivial matters. Crystal clear job titles and descriptions holacracy Well-documented workflows
While they can change with new people bringing fresh ideas, it’s important to document these innovations and ensure there’s an up-to-date and easily accessible resource hub. Well-documented workflows While they can change with new people bringing fresh ideas, it’s important to document these innovations and ensure there’s an up-to-date and easily accessible resource hub. Well-documented workflows Being selective about who we hire
Hiring becomes more difficult because you need experts with particular soft skills, such as self-reliance, initiative, and superb organization. When we hire, we look for those who are comfortable managing their time and workload without a supervisor constantly checking in. Being selective about who we hire Hiring becomes more difficult because you need experts with particular soft skills, such as self-reliance, initiative, and superb organization. When we hire, we look for those who are comfortable managing their time and workload without a supervisor constantly checking in. Being selective about who we hire Final Thoughts Flat organization structures are a great fit for startups and midsize businesses that want to be fast and innovative. They offer a natural breeding ground for the principles of autonomy, competence, and relatedness that are key for keeping your employees committed long-term. But they come with a catch, and it’s up to you to tackle or forgo it. Image source: Cine Insomnia/Pexels Image source: Cine Insomnia/Pexels

Why Flat Organizational Structures Drive Innovation in Startups

MVP Game Development With Flutter and Flame

ফ্লটার এবং ফ্লেম সহ MVP গেম ডেভেলপমেন্ট

MVP-Spieleentwicklung mit Flutter und Flame

Desarrollo de juegos MVP con Flutter y Flame

Développement de jeux MVP avec Flutter et Flame

फ़्लटर और फ्लेम के साथ एमवीपी गेम डेवलपमेंट

Flutter と Flame を使用した MVP ゲーム開発

Flutter 및 Flame을 사용한 MVP 게임 개발

Desenvolvimento de jogos MVP com Flutter and Flame

Разработка игр MVP с Flutter и Flame

Flutter ve Flame ile MVP Oyun Geliştirme

Phát triển trò chơi MVP với Flutter và Flame

使用 Flutter 和 Flame 进行 MVP 游戏开发

Flashift emerges as a premier platform, offering cutting-edge solutions to the blockchain community and digital currency enthusiasts. Through Flashift, users can effortlessly swap a wide range of cryptocurrencies across multiple blockchains. Leveraging strategic alliances with leading non-custodial exchanges that facilitate cryptocurrency swaps, Flashift identifies and delivers optimal market prices directly to its users. As a sophisticated aggregator, Flashift meticulously scours available offers, organizes them, and presents the best rates within a streamlined interface. One of the key features that distinguishes Flashift from other competitors is its use of artificial intelligence. Powered by its powerful AI, Flashift monitors all of its partner exchanges and identifies the best ones for you at the moment of transaction using special tags. 4 tags are assigned to exchanges: Recommended, Best Rate, Best in KYC, and No Hold to help you make the best choice based on your needs. Below is an introduction to these 4 tags: No Hold Tag In the realm of cryptocurrency exchanges, adhering to international Anti-Money Laundering (AML) regulations is paramount. These regulations aim to combat illicit financial activities and safeguard the integrity of the financial system. However, AML compliance can sometimes lead to complexities for exchange users, particularly in situations involving suspected fraudulent transactions. The Challenge of Asset Holds Many exchanges implement AML policies that empower them to restrict or freeze user assets in case of suspected fraudulent transactions. While this approach serves to deter illicit activities, it can also cause inconvenience for legitimate users, potentially hindering their access to their funds. Flashift's AI-Powered Solution Flashift, a leading cryptocurrency exchange aggregator, has addressed this challenge through its sophisticated AI-driven evaluation system. The system meticulously analyzes the terms and conditions of each partner exchange, paying particular attention to their AML policies and procedures regarding suspected fraudulent transactions. Flashift's AI identifies exchanges that explicitly forego the right to retain user assets in cases of suspected fraud. These exchanges are then assigned the "No Hold" tag, indicating to users that their funds will not be frozen or restricted under such circumstances. Benefits of Choosing No Hold Exchanges For users prioritizing the unhindered availability of their funds, opting for exchanges bearing the "No Hold" tag offers several advantages: Reduced Risk of Asset Retention: Users can minimize the likelihood of their assets being frozen or restricted due to AML investigations or other regulatory procedures.
Enhanced Peace of Mind: The "No Hold" tag provides a sense of security and assurance regarding the accessibility of users' funds.
Empowered Decision-Making: Users can make informed choices about their exchange partners, prioritizing platforms that respect their ownership rights and minimize potential asset holds. Best in KYC Tag Flashift's AI system continuously monitors and analyzes user reviews, comments, and feedback from various platforms, including Trustpilot, to gain insights into user experiences with partner exchanges. By meticulously categorizing and evaluating this vast amount of data, the AI system generates a comprehensive understanding of user sentiment towards each exchange. Based on this in-depth analysis, Flashift's AI assigns the "Best in KYC" tag to exchanges that consistently demonstrate high levels of user satisfaction and positive feedback. When a user selects a trading pair on the Flashift platform, the AI automatically identifies the "Best in User Satisfaction" exchange, enabling users to prioritize platforms that prioritize user experience and satisfaction. Best Rate Tag As the name suggests, the "Best Rate" tag highlights the exchange offering the most favorable exchange rate for a given trading pair at the precise moment of transaction. This real-time evaluation ensures that users are presented with the most advantageous exchange option, minimizing transaction costs and maximizing their trading potential. Recommended Tag At the forefront of Flashift's exchange evaluation system lies the Recommended tag, a culmination of AI-powered insights and data-driven analysis that guides users towards the exchange that best suits their trading needs. Flashift's AI meticulously gathers and evaluates a comprehensive range of exchange metrics, encompassing factors such as competitive exchange rates, transaction speeds, user satisfaction levels, and KYC compliance. By meticulously analyzing this vast array of data, the AI identifies the exchange that stands out as the optimal choice for each user. To further refine the evaluation process, Flashift's AI constructs a detailed table of key exchange metrics. Employing proprietary algorithms and formulas, the AI assigns scores to each exchange, ultimately identifying the top performer. This data-driven approach ensures that the Recommended tag is not swayed by personal opinions or subjective biases, but rather reflects a well-informed assessment based on objective criteria. With its expertise honed through extensive data analysis, Flashift's AI has evolved into a seasoned cryptocurrency exchange advisor. Just as an experienced consultant would guide a client towards the most suitable option, Flashift's Recommended tag presents users with an exchange that aligns perfectly with their trading goals and preferences. By prioritizing the Recommended tag, users can eliminate the need for time-consuming research and subjective judgment, confident in the knowledge that they are selecting the exchange that offers the best combination of security, speed, user satisfaction, and KYC compliance. The Recommended tag serves as a beacon, guiding users towards a seamless and rewarding trading experience. Flashift’s Recommended tag stands as a testament to the company's commitment to user-centricity and innovation. By harnessing the power of AI and data-driven insights, Flashift empowers users to make informed decisions, ensuring that they select the exchange that aligns perfectly with their trading needs and aspirations. With the Recommended tag as their guide, users can embark on a secure, efficient, and ultimately rewarding cryptocurrency trading journey. Choose Flashift for a streamlined, secure, and intelligent approach to cryptocurrency swaps. Flashift Exchange | How Flashift works | Medium This story was distributed as a release by Btcwire under HackerNoon’s Business Blogging Program. Learn more about the program here. Flashift emerges as a premier platform, offering cutting-edge solutions to the blockchain community and digital currency enthusiasts. Through Flashift, users can effortlessly swap a wide range of cryptocurrencies across multiple blockchains. Leveraging strategic alliances with leading non-custodial exchanges that facilitate cryptocurrency swaps, Flashift identifies and delivers optimal market prices directly to its users. As a sophisticated aggregator, Flashift meticulously scours available offers, organizes them, and presents the best rates within a streamlined interface. Flashift Flashift One of the key features that distinguishes Flashift from other competitors is its use of artificial intelligence. Powered by its powerful AI, Flashift monitors all of its partner exchanges and identifies the best ones for you at the moment of transaction using special tags. 4 tags are assigned to exchanges: Recommended, Best Rate, Best in KYC, and No Hold to help you make the best choice based on your needs. Below is an introduction to these 4 tags: No Hold Tag In the realm of cryptocurrency exchanges, adhering to international Anti-Money Laundering (AML) regulations is paramount. These regulations aim to combat illicit financial activities and safeguard the integrity of the financial system. However, AML compliance can sometimes lead to complexities for exchange users, particularly in situations involving suspected fraudulent transactions. The Challenge of Asset Holds Many exchanges implement AML policies that empower them to restrict or freeze user assets in case of suspected fraudulent transactions. While this approach serves to deter illicit activities, it can also cause inconvenience for legitimate users, potentially hindering their access to their funds. Flashift's AI-Powered Solution Flashift's AI-Powered Solution Flashift, a leading cryptocurrency exchange aggregator, has addressed this challenge through its sophisticated AI-driven evaluation system. The system meticulously analyzes the terms and conditions of each partner exchange, paying particular attention to their AML policies and procedures regarding suspected fraudulent transactions. Flashift's AI identifies exchanges that explicitly forego the right to retain user assets in cases of suspected fraud. These exchanges are then assigned the "No Hold" tag, indicating to users that their funds will not be frozen or restricted under such circumstances. Benefits of Choosing No Hold Exchanges Benefits of Choosing No Hold Exchanges For users prioritizing the unhindered availability of their funds, opting for exchanges bearing the "No Hold" tag offers several advantages: Reduced Risk of Asset Retention: Users can minimize the likelihood of their assets being frozen or restricted due to AML investigations or other regulatory procedures. Enhanced Peace of Mind: The "No Hold" tag provides a sense of security and assurance regarding the accessibility of users' funds. Empowered Decision-Making: Users can make informed choices about their exchange partners, prioritizing platforms that respect their ownership rights and minimize potential asset holds. Reduced Risk of Asset Retention: Users can minimize the likelihood of their assets being frozen or restricted due to AML investigations or other regulatory procedures. Enhanced Peace of Mind: The "No Hold" tag provides a sense of security and assurance regarding the accessibility of users' funds. Empowered Decision-Making: Users can make informed choices about their exchange partners, prioritizing platforms that respect their ownership rights and minimize potential asset holds. Best in KYC Tag Flashift's AI system continuously monitors and analyzes user reviews, comments, and feedback from various platforms, including Trustpilot, to gain insights into user experiences with partner exchanges. By meticulously categorizing and evaluating this vast amount of data, the AI system generates a comprehensive understanding of user sentiment towards each exchange. Based on this in-depth analysis, Flashift's AI assigns the "Best in KYC" tag to exchanges that consistently demonstrate high levels of user satisfaction and positive feedback. When a user selects a trading pair on the Flashift platform, the AI automatically identifies the "Best in User Satisfaction" exchange, enabling users to prioritize platforms that prioritize user experience and satisfaction. Best Rate Tag Best Rate Tag As the name suggests, the "Best Rate" tag highlights the exchange offering the most favorable exchange rate for a given trading pair at the precise moment of transaction. This real-time evaluation ensures that users are presented with the most advantageous exchange option, minimizing transaction costs and maximizing their trading potential. Recommended Tag Recommended Tag At the forefront of Flashift's exchange evaluation system lies the Recommended tag, a culmination of AI-powered insights and data-driven analysis that guides users towards the exchange that best suits their trading needs. Flashift's AI meticulously gathers and evaluates a comprehensive range of exchange metrics, encompassing factors such as competitive exchange rates, transaction speeds, user satisfaction levels, and KYC compliance. By meticulously analyzing this vast array of data, the AI identifies the exchange that stands out as the optimal choice for each user. To further refine the evaluation process, Flashift's AI constructs a detailed table of key exchange metrics. Employing proprietary algorithms and formulas, the AI assigns scores to each exchange, ultimately identifying the top performer. This data-driven approach ensures that the Recommended tag is not swayed by personal opinions or subjective biases, but rather reflects a well-informed assessment based on objective criteria. With its expertise honed through extensive data analysis, Flashift's AI has evolved into a seasoned cryptocurrency exchange advisor. Just as an experienced consultant would guide a client towards the most suitable option, Flashift's Recommended tag presents users with an exchange that aligns perfectly with their trading goals and preferences. By prioritizing the Recommended tag, users can eliminate the need for time-consuming research and subjective judgment, confident in the knowledge that they are selecting the exchange that offers the best combination of security, speed, user satisfaction, and KYC compliance. The Recommended tag serves as a beacon, guiding users towards a seamless and rewarding trading experience. Flashift’s Recommended tag stands as a testament to the company's commitment to user-centricity and innovation. By harnessing the power of AI and data-driven insights, Flashift empowers users to make informed decisions, ensuring that they select the exchange that aligns perfectly with their trading needs and aspirations. With the Recommended tag as their guide, users can embark on a secure, efficient, and ultimately rewarding cryptocurrency trading journey. Choose Flashift for a streamlined, secure, and intelligent approach to cryptocurrency swaps. Flashift Exchange | How Flashift works | Medium Flashift Exchange | How Flashift works | Medium Flashift Exchange Flashift Exchange How Flashift works How Flashift works Medium Medium This story was distributed as a release by Btcwire under HackerNoon’s Business Blogging Program. Learn more about the program here. This story was distributed as a release by Btcwire under HackerNoon’s Business Blogging Program. Learn more about the program here . here here

Meet Flashift's New AI-Powered Platform For Seamless Cryptocurrency Swaps

Analyzing HIVE Flaws and Part-Prototype Models: A Comparative Study

Happiness Is Based on a Flawed Equation — Here's How to Fix It

সুখ একটি ত্রুটিপূর্ণ সমীকরণের উপর ভিত্তি করে - এটি কীভাবে ঠিক করা যায় তা এখানে

Glück basiert auf einer fehlerhaften Gleichung – so lässt sie sich beheben

La felicidad se basa en una ecuación defectuosa: aquí se explica cómo solucionarla

Le bonheur est basé sur une équation erronée : voici comment y remédier

खुशी एक दोषपूर्ण समीकरण पर आधारित है - इसे ठीक करने का तरीका यहां बताया गया है

幸福は間違った方程式に基づいている — これを修正する方法はこれだ

행복은 잘못된 방정식에 기초합니다. 문제를 해결하는 방법은 다음과 같습니다.

A felicidade é baseada em uma equação falha – veja como consertar

Счастье основано на ошибочном уравнении — вот как это исправить

Mutluluk Kusurlu Bir Denkleme Dayalı - İşte Nasıl Düzeltileceği

Hạnh phúc dựa trên một phương trình sai lầm - Đây là cách khắc phục

幸福建立在一个有缺陷的等式上——下面是如何修复它

Dortmund, Germany, May 9th, 2024/Chainwire/--CONF3RENCE 2024 and BLOCKCHANCE are thrilled to announce their union, marking the inception of a flagship Web3 event in the Ruhr area, Europe's economic powerhouse. BLOCKCHANCE has already demonstrated the possibilities of the German Web3 ecosystem. Together, aiming to build a bridge between traditional economy and Web3, this union propels CONF3RENCE to a leading position on a global level. Scheduled for next week, May 15th and 16th at Signal Iduna Park in Dortmund, Germany's largest stadium, CONF3RENCE aims to bridge the traditional economy with the burgeoning Web3 and AI sectors. The event is set to be a pivotal moment for the blockchain community, featuring 150+ speakers and hosting more than 100 partners & exhibitors from around the globe. Highlights of CONF3RENCE 2024: Top-notch speakers from traditional enterprises, finance, science, and politics; European Central Bank, NVIDIA, Animoca Brands, IBM, Litecoin and many more
Exciting panel discussions and workshops on the most critical topics in the blockchain and AI industries
Exclusive networking opportunities for making new contacts and exchanging ideas with experts.
8 million EUR in prizes for our Startup Award, an innovative hackathon, live demos and further exciting interactive elements
Partnerships with leading industry and finance associations, significant blockchain associations and business clubs enable an unprecedented quality of visitors and opportunities
Large exhibition area for presenting products and business opportunities to promote innovation and business relationships CONF3RENCE is more than an event; it's a mission to unify stakeholders across sectors to foster the implementation of blockchain and AI technologies, showcasing existing applications and exploring new synergies. With its comprehensive program and dedication to innovation, CONF3RENCE is poised to catalyze the growth of blockchain technology and shape the future of digital economies. Join us in Dortmund for an unprecedented exploration of the potential that lies at the intersection of the traditional economy and the dynamic Web3 and AI industries. CONF3RENCE is your opportunity to be part of the conversation shaping the future of technology and society. About CONF3RENCE CONF3RENCE is a leading global forum dedicated to the exploration and promotion of blockchain and AI technologies. Situated in the heart of Europe's economic powerhouse, the Ruhr area, CONF3RENCE serves as a pivotal platform for bridging the traditional economy with the innovative realms of Web3 and AI. To learn more, visit the website, or X (Twitter), or LinkedIn. Contact Sascha Roehrer SOLV3 GmbH sascha@solv3.eu This story was distributed as a release by Chainwire under HackerNoon’s Business Blogging Program. Learn more about the program here. Dortmund, Germany, May 9th, 2024/Chainwire/--CONF3RENCE 2024 and BLOCKCHANCE are thrilled to announce their union, marking the inception of a flagship Web3 event in the Ruhr area, Europe's economic powerhouse. BLOCKCHANCE has already demonstrated the possibilities of the German Web3 ecosystem. Together, aiming to build a bridge between traditional economy and Web3, this union propels CONF3RENCE to a leading position on a global level. Scheduled for next week, May 15th and 16th at Signal Iduna Park in Dortmund, Germany's largest stadium, CONF3RENCE aims to bridge the traditional economy with the burgeoning Web3 and AI sectors. The event is set to be a pivotal moment for the blockchain community, featuring 150+ speakers and hosting more than 100 partners & exhibitors from around the globe. Highlights of CONF3RENCE 2024: Top-notch speakers from traditional enterprises, finance, science, and politics; European Central Bank, NVIDIA, Animoca Brands, IBM, Litecoin and many more Exciting panel discussions and workshops on the most critical topics in the blockchain and AI industries Exclusive networking opportunities for making new contacts and exchanging ideas with experts. 8 million EUR in prizes for our Startup Award, an innovative hackathon, live demos and further exciting interactive elements Partnerships with leading industry and finance associations, significant blockchain associations and business clubs enable an unprecedented quality of visitors and opportunities Large exhibition area for presenting products and business opportunities to promote innovation and business relationships Top-notch speakers from traditional enterprises, finance, science, and politics; European Central Bank, NVIDIA, Animoca Brands, IBM, Litecoin and many more Exciting panel discussions and workshops on the most critical topics in the blockchain and AI industries Exclusive networking opportunities for making new contacts and exchanging ideas with experts. 8 million EUR in prizes for our Startup Award, an innovative hackathon, live demos and further exciting interactive elements Partnerships with leading industry and finance associations, significant blockchain associations and business clubs enable an unprecedented quality of visitors and opportunities Large exhibition area for presenting products and business opportunities to promote innovation and business relationships CONF3RENCE is more than an event; it's a mission to unify stakeholders across sectors to foster the implementation of blockchain and AI technologies, showcasing existing applications and exploring new synergies. With its comprehensive program and dedication to innovation, CONF3RENCE is poised to catalyze the growth of blockchain technology and shape the future of digital economies. Join us in Dortmund for an unprecedented exploration of the potential that lies at the intersection of the traditional economy and the dynamic Web3 and AI industries. CONF3RENCE is your opportunity to be part of the conversation shaping the future of technology and society. About CONF3RENCE CONF3RENCE is a leading global forum dedicated to the exploration and promotion of blockchain and AI technologies. Situated in the heart of Europe's economic powerhouse, the Ruhr area, CONF3RENCE serves as a pivotal platform for bridging the traditional economy with the innovative realms of Web3 and AI. To learn more, visit the website , or X (Twitter) , or LinkedIn . website website X (Twitter) X (Twitter) LinkedIn LinkedIn Contact Sascha Roehrer SOLV3 GmbH sascha@solv3.eu This story was distributed as a release by Chainwire under HackerNoon’s Business Blogging Program. Learn more about the program here. This story was distributed as a release by Chainwire under HackerNoon’s Business Blogging Program. Learn more about the program here . here here

CONF3RENCE 2024 The Web3 Flagship Event In Germany Is Just Around The Corner

CONF3RENCE 2024 জার্মানিতে ওয়েব3 ফ্ল্যাগশিপ ইভেন্টটি একেবারে কোণায় রয়েছে

CONF3RENCE 2024 Das Web3-Flaggschiff-Event in Deutschland steht vor der Tür

CONF3RENCE 2024 El evento insignia de Web3 en Alemania está a la vuelta de la esquina

CONF3RENCE 2024 L'événement phare du Web3 en Allemagne approche à grands pas

जर्मनी में वेब3 फ्लैगशिप इवेंट कॉन्फ़्रेंस 2024 बस आने ही वाला है

CONF3RENCE 2024 ドイツにおける Web3 のフラッグシップ イベントが間もなく開催されます

CONF3RENCE 2024 독일의 Web3 플래그십 이벤트가 곧 다가옵니다.

CONF3RENCE 2024 O principal evento da Web3 na Alemanha está chegando

CONF3RENCE 2024 Флагманское мероприятие Web3 в Германии уже не за горами

CONF3RENCE 2024 Almanya'daki Web3 Amiral Gemisi Etkinliği Çok Yakında

CONF3RENCE 2024 Sự kiện hàng đầu của Web3 ở Đức sắp diễn ra

德国 Web3 旗舰盛会 CONF3RENCE 2024 即将召开

Colliding Magnetospheres in The Young High-Eccentricity Binary DQ Tau: Flare Analyses

Stories, News & Mentions about FLKA

Flka’s Stories on HackerNoon