Founded in 2017 as a spin-off from the renowned RWTH Aachen University, envelio is quickly becoming a name to reckon with in the clean tech arena. Based in Cologne, the startup focuses on ...

5 Exciting Tech Start-ups from Germany You Need to Watch

envelio

If you are going deep into cloud security, you may have seen the term “Envelope encryption”. Let’s explain the concept and why it is useful. Cryptography Like I Am 5 In the wild case you have never heard of cryptography, it is a set of tools and techniques to transform a message (usually of high importance) in a way that only authorized people can read it. The encrypted message (ciphertext) looks like plain-everyday text, but it is gibberish to an external. Moreover, encryption/decryption can happen either using a key that is known to all legitimate parties (symmetric encryption) or when the users share their keys (asymmetric encryption). In the latter case, each user has: A public key (which they share with anyone they want). A private key (known only to them). Envelope Encryption 101 Envelope encryption is a technique for encrypting data. It involves using a combination of symmetric and asymmetric encryption algorithms. First, the basic idea is to use symmetric encryption to encrypt the data itself. Then to use asymmetric encryption to encrypt the symmetric key used for the encryption. It is the same idea (not the same technique, though) SSL uses to exchange keys between the client and the server. If we could lay down the envelope encryption algorithm, the steps would be: A random symmetric key is generated to encrypt the data. The data is encrypted using the symmetric key generated above. The symmetric key is encrypted using the recipient’s public key (asymmetric encryption). The encrypted symmetric key and encrypted data are sent to the recipient. The recipient uses their private key to decrypt the encrypted symmetric key. The decrypted symmetric key is then used to decrypt the encrypted data. So this “double encryption” in steps 2 and 3 generates a conceptual envelope, thus the name. What About Performance? Symmetric encryption is faster (compared to asymmetric), so we have tangible benefits here. And wouldn’t it be easier and faster, if we didn’t encrypt at all, I hear you saying? Of course, as with all things in computer science, we have a trade-off here. In that case, you gain security points, you lose milliseconds. It is up to you and what you are trying to protect against. For example, what can happen if an attacker gains access to the encrypted key? They would also need to obtain the private key to decrypt it. And this multiplies their effort significantly. Services Where can you find envelope encryption? The most known tools are the key management systems in the public cloud providers. The most popular (and I believe the first in that scale) is AWS KMS Conclusion I hope this post helped you understand better the concept of envelope encryption. If you came here by accident, I want to say hi and I hope you enjoyed my writing anyway.

What the Heck Is Envelope Encryption in Cloud Security? 

Recently at work, I came across an interesting method to handle encryption at scale called envelope encryption. First of all, it increases security and helps you ease out the management of encryption keys. But it's also a highly recommended pattern by PCI-DSS (Security Standard for Credit Card Processing) and results in much stronger data privacy and data protection of Personally Identifiable Information (PII). When we think of data there are 3 places we can think of encrypting data: At Rest - On hardware storage devices like on a disk or in your devices In Transit - In moving data between different locations like server to server through API calls In Use - While it's being used by a server (New concept and still being researched) We will be dealing primarily with encryption at rest and envelope encryption is a popular pattern recommended for it. So What is Envelope Encryption? 🤔 This is a type of encryption that involves encrypting your data with a Data Encryption Key, then encrypting the Data Encryption Key (DEK) with a Customer Master Keys (CMK). You then store both the encrypted data and the encrypted DEK alongside each other in the database. This practice of using a wrapping key to encrypt data keys is known as envelope encryption. Like mentioned there are 2 keys you need to understand first before we see how the encryption process takes place: Customer Master Key (CMK) Data Encryption Key (DEK) Customer Master Keys/Root Keys/Key Encryption Keys (CMK) These are symmetric keys used to encrypt, decrypt, and re-encrypt data. It can also generate Data Encryption Keys that you can use outside of the KMS system. They follow the below: Access to these must be restricted to the least endpoints Access to these should be secured through ACL These keys must be stored in a location that is secure like a KMS of a Hardware Security Module (to comply with ) FIPS 140-2 In systems like Google Cloud Key Management Service, you have a hierarchy of keys as seen below with more information to be found . here Data Encryption Keys (DEK) Data keys are encryption keys you can use to encrypt data, including large amounts of data and other data encryption keys. Unlike CMK's, which can't be downloaded, data keys are returned to you for use outside of the KMS. Some of the best practices for DEKs: Generate DEKs locally When stored, always ensure DEKs are encrypted at rest For easy access, store the DEK near the data that it encrypts Generate a new DEK every time you write the data. This means you don't need to rotate the DEKs Do not use the same DEK to encrypt data from two different users Use a strong algorithm such as 256-bit Advanced Encryption Standard (AES) Encryption Process API request is sent to KMS to generate Data key using CMK KMS returns a response with Plain Data key and Encrypted Data key (using CMK) Data is encrypted using Plain Data key Plain Data key is removed from memory Encrypted Data and Encrypted Data Key is packaged together as an envelope and stored Decryption Process Encrypted Data key is extracted from the envelope API request is sent to KMS using Encrypted Data key which has information about CMK to be used in KMS for decryption KMS returns a response with Plain Data Key Encrypted Data is decrypted using Plain Data key Plain Data Key is removed from memory How is Envelope Encryption Different From Other Encryption Patterns? 🤔 Every service you build requires encryption at some point. This could be passwords or PII in a database, credentials for an external service, or even files in a filesystem. Configuration Files You can easily handle some of these situations with a configuration file but they pose their own security risks like: Proper planning is needed to keep the data secure Multiple formats are present e.g - YAML, JSON and XML to name a few Exact storage locations may be hard-coded in the app, making deployment potentially problematic Parsing of the config files can be problematic. Symmetric Encryption You can encrypt data using a symmetric key but they suffer from a major issue which is Key Management. You need to find a way to get the key to the party with whom you are sharing data and if someone gets their hands on a symmetric key, they can decrypt everything encrypted with that key. Asymmetric Encryption You can encrypt data using Asymmetric Encryption which is considered as a standard now a days. Some of the cons of it are: It is a slow process which makes its not suitable for decrypting bulk messages When you lose your private key, your received messages will not be decrypted If your private key is identified by an attacker, all of your messages can be read by him/her Envelope Encryption Some of the benefits offered by it are: - The data is encrypted using a DEK which follows symmetric encryption. The DEK is encrypted by a CMK which follows asymmetric encryption. By using asymmetric encryption, encrypted DEKs can be shared and unencrypted only by those with access to the CMK, mitigating the key exchange problem of symmetric algorithms. A combination of benefits from symmetric and asymmetric encryption - Multiple DEKs can be encrypted under a singular root key and ease the management of keys in a KMS. You can also do more secure key maintenance by rotating your root keys, instead of rotating and re-encrypting all of your DEKs. Easier key management - Because we encrypt the data key with the CMK, we don't have to worry about storing the encrypted data key. Thus, we can safely store the encrypted data key alongside the encrypted data. Data key protection Key Management Systems & Why it Works at Scale? 🤔 The biggest reason for Envelope Encryption and KMSs working at scale is . Like we mentioned before Asymmetric Encryptions are typically slow and Symmetric Encryptions are very fast but the management of keys is the issue. Performance So in Envelope Encryption for a large quantity of data, you quickly encrypt it using symmetric encryption using a random key. Then just the key is encrypted using asymmetric encryption. This gives the benefits of asymmetric encryption, with the performance of symmetric encryption. Key Management Systems like AWS KMS, Azure Key Vault, and Google Cloud Key Management Service gives you a fully managed service to store and manage encryption keys. These use envelope encryption internally, and they’re used by default in a lot of services that support encryption in cloud infrastructure providers like AWS, GCP, Azure, and others. An ideal key management system should be highly available, it should control access to the master key(s), it should audit the key(s) usage, and finally, it should manage key(s) lifecycle. Thus by having the above characteristics and by using envelope encryption internally, Key Management Systems are ideal to handle encryption at scale. Summary Envelope Encryption is one of the most trusted application security design patterns used at scale. It is the default encryption method used in services like AWS S3, GCP, and others. Hopefully, this enables you to understand how you can encrypt/decrypt a large amount of data using the envelope encryption method at scale in a more trusted setup. Thanks for reading! I really hope that you find this article useful. I invite you to participate in the discussion in the comments below, I'm always interested to know your thoughts and happy to answer any questions you might have in your mind. If you think this post was useful, please like the post to help promote this piece to others. Thanks for reading! :) P.S Do feel free to connect with me on or LinkedIn Twitter Appendix This article leans heavily on the following material: Google Cloud Data Encryption - Jayendra's Cloud Certification Blog - AWS KMS concepts - AWS AWS KMS and Envelope Encryption - Manish Pandit Cloud Architecture Pattern: Envelope Encryption (or Digital Envelope) with Public Cloud Providers Part 1 - Nilay Parikh AWS KMS Envelope Encryption - Chirag Modi Protecting data with envelope encryption - IBM Envelope encryption - GCP Encryption at rest in Google Cloud - GCP Managing encryption of data in the cloud (Google Cloud Next '17) - Maya Kaczorowski

Managing Encryption at Scale with Envelope Encryption

The Research and Innovation team led by Diego Masini held a webinar on 26th April 2021 to introduce the Project to the RSK community. RIF Enveloping The RIF Enveloping team comprises of: Diego Masini Raul Laprida Julian Len The main motivation behind RIF Enveloping is to allow users to pay for transaction fees using their preferred tokens, driving down the barriers to DeFi adoption by simplifying the UX. RIF Enveloping makes it easier to onboard new users by removing the need to get blockchain native currency to operate, allowing users to work only with the tokens they prefer. Features: Users will be able to find a sponsor willing to pay for their transaction fees in exchange for tokens on the network. Send a signed message to the sponsor with transaction details, along with a payment in tokens. The sponsor will need to be incentivized to ensure that transactions sent to the sponsor are submitted to the network, and also incentives for the users to ensure that the user does not try to cheat the sponsor of his payment. A potential solution for this type of problem is the . Due to the drawbacks of the GSN, it couldn’t be adapted to the RSK Network. Gas Station Network (GSN) GSN Drawbacks Existing contracts need to be adapted to use GSN The solution is expensive for users To solve these problems mentioned above, a redesign of GSN was made, this is known as Enveloping. Improvements Provide users with an address ready for enveloping (Smart Wallet), controlled by the respective user’s . EOA Changed the architecture to remove unnecessary contract interactions, and optimize remaining contracts. RIF Enveloping Design RIF Enveloping is designed to provide users with enveloping-ready addresses (Smart Wallet is deployed through a sponsor only when the user wants to submit a transaction. The smart wallet is designed in such a way as to ensure that the sponsor gets paid. Users in the RSK network can stop using Enveloping at any time once they get enough RBTC to operate on their own. How it Works Relay Transaction Flow In the image above, the user collects a request for all the information needed to perform the execution of the transaction plus all the information to send the payment to the sponsor. The user sends this information to the relay server, through a relayProvider (Web3 provider). The server then creates a transaction, using the information from the user as payload, then sends this transaction in an off-chain call to the relayVerifier, to verify that the transaction will be successful and well-formed. Once the server determines the request is valid, it then sends the request to the relayHub, which is an on-chain contract using a relay worker account, once the relayHub receives this transaction, it opens it and gets the payload, then submits it to the Smart Wallet proxy (which pays the token contract), reimbursing the money spent by the relayWorker using tokens, and then executes the destination contracts. A similar flow is followed when the user wants to deploy a transaction. Deploy Transaction Flow The user creates the requests with all the information for deployment, plus the payment. The RelayServer verifies that the deployment is going to be successful, using a deployVerifier, then if the request is valid, it sends the request to the relayHub contract using the relayWorker but now instead of interacting directly with the smart wallet because it doesn’t exist, the relayHub interacts with the SmartWalletFactory. Proxies were used for the smart wallets to further reduce the cost of the solution. The smart wallet factory deploys the actual smart wallet and pays the relay worker in tokens. Raul Laprida then demoed a live solution to display these features. Getting Started Want to integrate enveloping in your dApp project? Visit the and also see the . RIF Enveloping Documentation Enveloping User Guide See the for an explanation of key terms. Glossary Section Also published on: https://www.rifos.org/blog/rif-enveloping-webinar-key-takeaways

RIF Enveloping: How to Improve the UX of DeFi

Angular Expandable and Collapsible Sidebar With Icons: A Step-by-Step Guide

1. HackerNoon: What is your company in 2–5 words? Aaron Lee: We offer AI-driven call & chat answering to SMB. 2. Why is now the time for your company to exist? Providing real time 24/7 customer engagement, lead qualification and conversion has been the biggest pain point for SMB as they don’t have the time nor the resources to do it. Traditional call answering services can only handle simple tasks like take a message and gather contact information. With advancements in technology, particularly AI and automation, Smith.ai evolved to leverage AI to answer calls and chat with human-in-the-loop for escalation, expanding services to integrate with CRMs, provide scheduling assistance, payment collection, and deliver enhanced customer support tailored for businesses of all sizes, especially SMBs. 3. What do you love about your team, and why are you the ones to solve this problem? AI is fast moving but at the current stage, we don’t believe it can handle 100% of the calls and chats as human conversation is nuanced and complicated. Humans can easily detect caller’s emotion but it’s challenging if AI only looks at the transcription, e.g. “That’s great” could be either positive or sarcastic. We believe having a group of highly trained agents + highly tuned AI based on each individual business (think about it as Voice GPT for your business) is the best way to achieve the business outcomes: converting callers to customers. My background is in software engineering, computer graphics with GPU and was the former CTO of The Home Depot which allowed me to work with SMB on a daily basis for years. Our team is unique as we have AI expertise, deep domain knowledge, as well as huge appreciation to the pain points that SMB are facing. 4. If you weren’t building your startup, what would you be doing? I am very excited about entrepreneurship, helping other startups succeed, and have been angel investing for a long time as a side hobby. I probably would be more active in early stage investments if I wasn’t building my startup. 5. At the moment, how do you measure success? What are your metrics? We track customer engagement (number inbound/outbound calls or chats) as one of the core KPIs. Another metric we track is Quality, we treat any internal/external complaint/mistake/expectation mismatch as a concern, basically anything that is not satisfactory to our team or clients would be counted as one and we have been able to maintain a 99.7% correctness for a very long time. This metric has a direct correlation to our churn which is extremely low, which drives the flywheel of great reviews https://www.trustpilot.com/review/smith.ai and more referrals. 6. In a few sentences, what do you offer to whom? Smith.ai provides AI-driven virtual receptionist services, including call and chat answering, tailored for small and medium-sized businesses (SMBs). Our offerings encompass 24/7 live virtual receptionists, AI voice assistants, outreach campaigns, and AI-powered web chat, all designed to enhance customer engagement and support. We serve a wide range of industries, including law firms, home services, marketing agencies, medical and wellness, and many more. 7. What’s most exciting about your traction to date? The most exciting thing about Smith.ai is we have answered tens of millions of calls and chats with extremely high accuracy and satisfaction (99.7%). The most important thing is we help our SMB clients grow by converting callers to customers with the latest AI technologies which is impossible if they were to implement on their own. 8. Where do you think your growth will be next year? Given the uncertainty in inflation and macroeconomics, we believe the trend of doing more with less by leveraging AI and outsourcing will remain front and center for our clients. The fact that we can drive instant satisfaction and revenue growth at a much lower cost would be a huge tailwind for our exponential growth next year. 9. Tell us about your first paying customer and revenue expectations over the next year. We started the company back in October 2015 and I am excited to say that the majority of them are still with us today and have increased their spend with us 10X. We recently launched our AI-first Voice Assistant and it’s a total game changer. One client said “whoever put this together is a bloody genius.” - this is the magical moment when the product is so good that wows our clients. We believe we will have a significant jump in revenue next year as we continue to deploy this technology to 30M SMB in the US. 10. What’s your biggest threat? Our biggest threat is really execution. The market is huge and demand is clear. We have clients asking us for new features on a daily basis and it’s up to us to push the envelope of AI with humans fast enough to meet their needs. Our assessment of the current AI capabilities is that they may struggle with nuanced and complex conversations, especially in emotional or context-heavy interactions. We believe our hybrid human-AI approach is the best in order to bridge the gap. For appointment booking or sales conversion, we believe human touch offers the best business outcome. This startup founder interview template is based on HackerNoon Founder & CEO David Smooke’s ten questions for startup founders. Would you like to take a stab at answering some of these questions? The link for the template is HERE. 1. HackerNoon : What is your company in 2–5 words? HackerNoon Aaron Lee : We offer AI-driven call & chat answering to SMB. Aaron Lee Aaron Lee 2. Why is now the time for your company to exist? Providing real time 24/7 customer engagement, lead qualification and conversion has been the biggest pain point for SMB as they don’t have the time nor the resources to do it. Traditional call answering services can only handle simple tasks like take a message and gather contact information. With advancements in technology, particularly AI and automation, Smith.ai evolved to leverage AI to answer calls and chat with human-in-the-loop for escalation, expanding services to integrate with CRMs, provide scheduling assistance, payment collection, and deliver enhanced customer support tailored for businesses of all sizes, especially SMBs. Smith.ai 3. What do you love about your team, and why are you the ones to solve this problem? AI is fast moving but at the current stage, we don’t believe it can handle 100% of the calls and chats as human conversation is nuanced and complicated. Humans can easily detect caller’s emotion but it’s challenging if AI only looks at the transcription, e.g. “That’s great” could be either positive or sarcastic. We believe having a group of highly trained agents + highly tuned AI based on each individual business (think about it as Voice GPT for your business) is the best way to achieve the business outcomes: converting callers to customers. My background is in software engineering, computer graphics with GPU and was the former CTO of The Home Depot which allowed me to work with SMB on a daily basis for years. Our team is unique as we have AI expertise, deep domain knowledge, as well as huge appreciation to the pain points that SMB are facing. 4. If you weren’t building your startup, what would you be doing? I am very excited about entrepreneurship, helping other startups succeed, and have been angel investing for a long time as a side hobby. I probably would be more active in early stage investments if I wasn’t building my startup. 5. At the moment, how do you measure success? What are your metrics? We track customer engagement (number inbound/outbound calls or chats) as one of the core KPIs. Another metric we track is Quality, we treat any internal/external complaint/mistake/expectation mismatch as a concern, basically anything that is not satisfactory to our team or clients would be counted as one and we have been able to maintain a 99.7% correctness for a very long time. This metric has a direct correlation to our churn which is extremely low, which drives the flywheel of great reviews https://www.trustpilot.com/review/smith.ai and more referrals. https://www.trustpilot.com/review/smith.ai https://www.trustpilot.com/review/smith.ai 6. In a few sentences, what do you offer to whom? Smith.ai provides AI-driven virtual receptionist services, including call and chat answering, tailored for small and medium-sized businesses (SMBs). Our offerings encompass 24/7 live virtual receptionists, AI voice assistants, outreach campaigns, and AI-powered web chat, all designed to enhance customer engagement and support. We serve a wide range of industries, including law firms, home services, marketing agencies, medical and wellness, and many more. 7. What’s most exciting about your traction to date? The most exciting thing about Smith.ai is we have answered tens of millions of calls and chats with extremely high accuracy and satisfaction (99.7%). The most important thing is we help our SMB clients grow by converting callers to customers with the latest AI technologies which is impossible if they were to implement on their own. 8. Where do you think your growth will be next year? Given the uncertainty in inflation and macroeconomics, we believe the trend of doing more with less by leveraging AI and outsourcing will remain front and center for our clients. The fact that we can drive instant satisfaction and revenue growth at a much lower cost would be a huge tailwind for our exponential growth next year. 9. Tell us about your first paying customer and revenue expectations over the next year. We started the company back in October 2015 and I am excited to say that the majority of them are still with us today and have increased their spend with us 10X. We recently launched our AI-first Voice Assistant and it’s a total game changer. One client said “whoever put this together is a bloody genius.” - this is the magical moment when the product is so good that wows our clients. We believe we will have a significant jump in revenue next year as we continue to deploy this technology to 30M SMB in the US. 10. What’s your biggest threat? Our biggest threat is really execution. The market is huge and demand is clear. We have clients asking us for new features on a daily basis and it’s up to us to push the envelope of AI with humans fast enough to meet their needs. Our assessment of the current AI capabilities is that they may struggle with nuanced and complex conversations, especially in emotional or context-heavy interactions. We believe our hybrid human-AI approach is the best in order to bridge the gap. For appointment booking or sales conversion, we believe human touch offers the best business outcome. This startup founder interview template is based on HackerNoon Founder & CEO David Smooke’s ten questions for startup founders. This startup founder interview template is based on HackerNoon Founder & CEO David Smooke ’s ten questions for startup founders . HackerNoon Founder & CEO David Smooke ten questions for startup founders Would you like to take a stab at answering some of these questions? The link for the template is HERE. Would you like to take a stab at answering some of these questions? The link for the template is HERE. The link for the template is HERE.

Blending AI and Human Expertise: How Smith.ai CEO/Founder is Revolutionizing SMB Customer Engagement

Omnity Eliminates the Chasm Between Bitcoin and Other Networks

Por qué Omnity Hub es la solución más segura y confiable para BTCFi

Pourquoi Omnity Hub est la solution la plus sûre et la plus fiable pour BTCFi

क्यों ओम्निटी हब BTCFi के लिए सबसे सुरक्षित और भरोसेमंद समाधान है

Omnity HubがBTCFiにとって最も安全で信頼できないソリューションである理由

ເປັນຫຍັງ Omnity Hub ເປັນການແກ້ໄຂທີ່ປອດໄພ ແລະບໍ່ມີຄວາມເຊື່ອໝັ້ນທີ່ສຸດສຳລັບ BTCFi

किन Omnity Hub BTCFi को लागि सबैभन्दा सुरक्षित र ट्रस्टलेस समाधान हो

ولې د Omnity Hub د BTCFi لپاره خورا خوندي او بې باوره حل دی

Imaraykutaq Omnity Hub Aswan Seguro & Mana Confiable Solución BTCFi kaqpaq

Почему Omnity Hub — самое безопасное и надежное решение для BTCFi

Impamvu Omnity Hub nigisubizo cyizewe cyane kandi kitizewe kuri BTCFi

Bakit Ang Omnity Hub ang Pinaka-Secure at Hindi Pinagkakatiwalaang Solusyon para sa BTCFi

Чому Omnity Hub є найбезпечнішим і надійним рішенням для BTCFi

The massive, catastrophic flash flood that occurred in Spain on October 29 was so horrific and devastating that it left the world aghast at its destructive nature. Over 200 people died, dozens went missing, and whole towns were covered in mud and debris, especially in the eastern Valencia region. In the United States alone, in 2023, 25 weather or climate disasters [caused at least $1 billion in losses and the [deaths](https://www.ncei.noaa.gov/access/billions/#:~:text=In%202023%20(as%20of%20December,and%201%20winter%20storm%20event) of 482 people, according to the National Centers for Environmental Information (NCEI). According to the US Government Accountability Office (GAO), “Severe storms, hurricanes, floods, and wildfires kill hundreds of people and cause billions of dollars in damage in the U.S. each year.” In fact, [according](https://www.ncei.noaa.gov/access/billions/#:~:text=In%202023%20(as%20of%20December,and%201%20winter%20storm%20event) to the National Oceanic and Atmospheric Administration (NOAA), “the U.S. has sustained 400 weather and climate disasters since 1980 where overall damages/costs reached or exceeded $1 billion (including CPI adjustment to 2024). The total cost of these 400 events exceeds $2.785 trillion.” India suffers from a similar problem. The frequency of flooding is getting worse due to the rapid growth of cities and climate change. Numerous other countries, especially in Africa, have seen similar catastrophes caused by unprecedented flooding. Widespread flooding in West and Central Africa, according to the World Health Organization, “has affected more than 4 million people, caused extensive damage to homes, schools, health facilities and swaths of farmland, heightening the risk of water-borne diseases, food insecurity and malnutrition.” Human-induced climate change “is likely to have significantly worsened the situation”, according to scientists at the Red Cross Climate Centre. Many experts now agree that with the advancement of artificial intelligence, this technology can and should play a bigger role in predicting those events and saving lives. And Northeastern University faculty experts predict it will. In an interview last year, Auroop R. Ganguly, professor of civil and environmental engineering and director of AI4CaS (AI for Climate and Sustainability) said, “In the next 12 months, we are going to see more and more efforts where data-driven systems and artificial intelligence come together.” One example is the Earth-2 platform, which is able to simulate global weather patterns in great detail. Taiwan is already using a similar tool to precisely predict the path and impact of typhoons. AI4CaS serves as a focus area within Northeastern University’s Institute for Experiential AI. The GAO found that AI weather forecasting models can help predict these catastrophic weather events. Machine learning specifically “can help improve model accuracy, increase warning times, and reduce uncertainty” and while some of these types of models are currently in use, “there are challenges to developing and adopting them more widely.” Google recently released an AI model that it claims can generate accurate weather forecasts at scale — while being cheaper than conventional physics-based forecasting. The "Scalable Ensemble Envelope Diffusion Sampler" (SEEDS) model is designed similarly to popular large language models (LLMs) like ChatGPT. SEEDS generates many ensembles — or multiple weather scenarios — much quicker and cheaper than traditional predicting models can. The team described its findings in a paper published March 29 in the journal Science Advances. In yet another effort to get ahead of life-threatening weather patterns, the University of Illinois and IBM have joined forces to work together on an AI model capable of anticipating heavy rainfall and flash floods in the Appalachian Mountains. Such models can triple or quadruple the warning time currently available to meteorologists and scientists. In August, NVIDIA Research, announced that it has developed a new AI model for predicting extreme weather events, called StormCast, part of a larger effort by NVIDIA to improve climate research through AI. These new AI models are financially beneficial since they can process information far quicker than older models. As extreme weather events, like flooding, become more frequent and intense, it is becoming increasingly important to be able to predict, understand, and prepare for these events. AI models will assist in predicting weather catastrophes and will help reduce not just major damage, but will also save thousands of lives. Such AI technologies will help scientists make faster and more accurate predictions, which can ultimately save lives and reduce the cost of disasters. In September, Smithsonian Magazine interviewed computer scientist and meteorologist Amy McGovern. She said that while AI will not replace human meteorologists in the near future, “We’re developing things to give them more options so that they can focus on turning the forecasts into actionable information for people to make better decisions.” “Humans are a lot better at that than AI is right now,” she said. The massive, catastrophic flash flood that occurred in Spain on October 29 was so horrific and devastating that it left the world aghast at its destructive nature. Over 200 people died, dozens went missing, and whole towns were covered in mud and debris, especially in the eastern Valencia region. In the United States alone, in 2023, 25 weather or climate disasters [caused at least $1 billion in losses and the [deaths](https://www.ncei.noaa.gov/access/billions/#:~:text=In%202023%20(as%20of%20December,and%201%20winter%20storm%20event) of 482 people, according to the National Centers for Environmental Information (NCEI). According to the US Government Accountability Office (GAO), “Severe storms, hurricanes, floods, and wildfires kill hundreds of people and cause billions of dollars in damage in the U.S. each year.” kill hundreds of people In fact, [according](https://www.ncei.noaa.gov/access/billions/#:~:text=In%202023%20(as%20of%20December,and%201%20winter%20storm%20event) to the National Oceanic and Atmospheric Administration (NOAA), “the U.S. has sustained 400 weather and climate disasters since 1980 where overall damages/costs reached or exceeded $1 billion (including CPI adjustment to 2024). The total cost of these 400 events exceeds $2.785 trillion .” The total cost of these 400 events exceeds $2.785 trillion India suffers from a similar problem. The frequency of flooding is getting worse due to the rapid growth of cities and climate change. The frequency of flooding is getting worse Numerous other countries, especially in Africa, have seen similar catastrophes caused by unprecedented flooding. Widespread flooding in West and Central Africa , according to the World Health Organization, “has affected more than 4 million people, caused extensive damage to homes, schools, health facilities and swaths of farmland, heightening the risk of water-borne diseases, food insecurity and malnutrition.” Widespread flooding in West and Central Africa Human-induced climate change “is likely to have significantly worsened the situation”, according to scientists at the Red Cross Climate Centre . Red Cross Climate Centre Many experts now agree that with the advancement of artificial intelligence, this technology can and should play a bigger role in predicting those events and saving lives. And Northeastern University faculty experts predict it will. In an interview last year, Auroop R. Ganguly, professor of civil and environmental engineering and director of AI4CaS (AI for Climate and Sustainability) said, “In the next 12 months, we are going to see more and more efforts where data-driven systems and artificial intelligence come together.” One example is the Earth-2 platform , which is able to simulate global weather patterns in great detail. Earth-2 platform Taiwan is already using a similar tool to precisely predict the path and impact of typhoons. AI4CaS serves as a focus area within Northeastern University’s Institute for Experiential AI. The GAO found that AI weather forecasting models can help predict these catastrophic weather events. Machine learning specifically “can help improve model accuracy, increase warning times, and reduce uncertainty” and while some of these types of models are currently in use, “there are challenges to developing and adopting them more widely.” Google recently released an AI model that it claims can generate accurate weather forecasts at scale — while being cheaper than conventional physics-based forecasting. The " Scalable Ensemble Envelope Diffusion Sampler " (SEEDS) model is designed similarly to popular large language models (LLMs) like ChatGPT. Scalable Ensemble Envelope Diffusion Sampler SEEDS generates many ensembles — or multiple weather scenarios — much quicker and cheaper than traditional predicting models can. The team described its findings in a paper published March 29 in the journal Science Advances . Science Advances In yet another effort to get ahead of life-threatening weather patterns, the University of Illinois and IBM have joined forces to work together on an AI model capable of anticipating heavy rainfall and flash floods in the Appalachian Mountains. joined forces Such models can triple or quadruple the warning time currently available to meteorologists and scientists. In August, NVIDIA Research, announced that it has developed a new AI model for predicting extreme weather events, called StormCast, part of a larger effort by NVIDIA to improve climate research through AI. announced These new AI models are financially beneficial since they can process information far quicker than older models. As extreme weather events, like flooding, become more frequent and intense, it is becoming increasingly important to be able to predict, understand, and prepare for these events. AI models will assist in predicting weather catastrophes and will help reduce not just major damage, but will also save thousands of lives. Such AI technologies will help scientists make faster and more accurate predictions, which can ultimately save lives and reduce the cost of disasters. In September, Smithsonian Magazine interviewed computer scientist and meteorologist Amy McGovern. She said that while AI will not replace human meteorologists in the near future, “We’re developing things to give them more options so that they can focus on turning the forecasts into actionable information for people to make better decisions.” Smithsonian Magazine interviewed “Humans are a lot better at that than AI is right now,” she said.

Will AI Ever Be Able to Predict Disasters and Save Lives?

Computing has changed in recent years, with the demand for high-performance computing (HPC) rising every single day. Many advancements were made, with most notable breakthroughs in the form of artificial intelligence and supercomputing. Meanwhile, Linux has created an ecosystem that is super adept at supercomputing efficiency, offering a robust and flexible platform for engineers. Without a doubt, the capabilities of the Linux system have changed the landscape of techdev with its innovations. Why Linux is the Operating System of Choice for Supercomputers There are many benefits of Linux and there is more than one reason why Linux is the first choice for many engineers and researchers. The platform it offers is highly scalable and flexible, including tools to optimize for HPC tasks, such as automatic vectorization in GCC (GNU Compiler Collection), OpenMP (Open Multi-Processing), and OpenMPI (Open Message Passing Interface), all of which aid in the optimization of HPC applications. Another benefit of Linux is that, regardless of programming language, Linux has a large set of capabilities to support them, which makes it much easier for developers to utilize the platform and implement solutions that are tailored to the applications as needed. The open-source nature of Linux encourages both growth and innovation by relying on the wisdom of the crowd - having hundreds of thousands of users all over the world contributing to its advancement. It also fosters a sense of community, and true enough, there is a vibrant community of developers that support one another’s research, collaborating in unison for the betterment of tech. In today’s landscape, 90% of the world’s most sophisticated supercomputers are powered by Linux. Moreover, for more than 20 years, Linux systems have filled more than half of the spots in the top 500 list of the most powerful machines in the world. Even back in 2018, all of them have become Linux powered, and, at that time, two new systems entering the list were IBM machines based on the POWER9 family of processors, with newly implemented supercomputer features designed by Gabriel Gomes, in the form of “compiler-toolchain software for high-precision, high-performance, floating-point arithmetic integrated into the core math library”, as he recalls. Now almost a decade later, Linux is still at the forefront, pushing the envelope for supercomputing and setting the standard for open-source platforms. Current HPC trends and cyber security Technology is constantly being upgraded. Developers and programmers never rest in their quest to find better solutions. With the integration of AI and machine learning, more doors have opened to researchers, and it is really up to the imagination how they apply the technology to create new and innovative solutions. In current trends, libraries like TensorFlow and PyTorch, which are optimized for Linux environments, are taking over because they are also open-sourced and rely on developers who work on underlying optimizations that enhance the performance of these machine learning algorithms. Optimization is not the only concern for HPC, though. The large amounts of data it consumes, as well as the A.I. models produced by the processing of this data are highly valuable assets and thus an increasingly common target for cybercrime and cyber security. Linux is again well positioned in this field, since having so many users using the platform means that vulnerabilities are quickly picked up and dealt with. Independently of the technique used - be it code review, static or dynamic analysis, or even automated A.I. solutions - detecting vulnerable paths before threats have a chance to manifest themselves is the goal. “A proactive approach is always better, as the saying goes, prevention is better than cure”, explains Gomes, who has been working with tech since graduating from University of Campinas and has since worked with IT hotshots like IBM and Linux distributor, SUSE. According to him, “It is important to continue fortifying the security of supercomputing systems for obvious reasons, particularly for maintaining uptime and preventing data breaches.” In this scenario, live patching is a new and valuable feature for supercomputing as it doesn’t interfere with uptime. During his time at SUSE, Gomes led the development of libpulp, a framework that allows the live patching of user-space processes, which can be used to safely apply changes to long-running computer programs. Since HPC requires performance, changes can also be reverted if their cost is deemed too high. This feature enhances the efficiency of patch management in production systems, and can be installed on any Linux distribution, explains Gomes, who is a Debian Developer in the Debian project. The Future of High-Performance Computing There is a growing reliance on supercomputing for sensitive research, which means that the performance and security of these systems will be the highest priority for future developments. The industry will continue to rely heavily on performance and security researchers that are able to focus on operating systems, compilers, libraries and patching mechanisms development for HPC. There will be increased collaboration across HPC domains, bringing together A.I., machine learning, and big data analytics. This also means that resource management will play a crucial role, as supercomputing consumes more resources than traditional computers. The development of sophisticated scheduling algorithms and workload management tools within the Linux ecosystem will further enhance resource availability and enable systems to hit peak performances. A driving force in the realm of high-performance computing, the Linux ecosystem will continue to excel and provide the technology needed to achieve greater efficiencies and capabilities, pushing forward the boundaries of what is possible in the future of computing. Computing has changed in recent years, with the demand for high-performance computing (HPC) rising every single day. Many advancements were made, with most notable breakthroughs in the form of artificial intelligence and supercomputing. Meanwhile, Linux has created an ecosystem that is super adept at supercomputing efficiency, offering a robust and flexible platform for engineers. Without a doubt, the capabilities of the Linux system have changed the landscape of techdev with its innovations. Why Linux is the Operating System of Choice for Supercomputers Why Linux is the Operating System of Choice for Supercomputers There are many benefits of Linux and there is more than one reason why Linux is the first choice for many engineers and researchers. The platform it offers is highly scalable and flexible, including tools to optimize for HPC tasks, such as automatic vectorization in GCC (GNU Compiler Collection), OpenMP (Open Multi-Processing), and OpenMPI (Open Message Passing Interface), all of which aid in the optimization of HPC applications . HPC applications HPC applications Another benefit of Linux is that, regardless of programming language, Linux has a large set of capabilities to support them, which makes it much easier for developers to utilize the platform and implement solutions that are tailored to the applications as needed. The open-source nature of Linux encourages both growth and innovation by relying on the wisdom of the crowd - having hundreds of thousands of users all over the world contributing to its advancement. It also fosters a sense of community, and true enough, there is a vibrant community of developers that support one another’s research, collaborating in unison for the betterment of tech. In today’s landscape, 90% of the world’s most sophisticated supercomputers are powered by Linux. Moreover, for more than 20 years, Linux systems have filled more than half of the spots in the top 500 list of the most powerful machines in the world. Even back in 2018, all of them have become Linux powered, and, at that time, two new systems entering the list were IBM machines based on the POWER9 family of processors, with newly implemented supercomputer features designed by Gabriel Gomes, in the form of “compiler-toolchain software for high-precision, high-performance, floating-point arithmetic integrated into the core math library”, as he recalls. Now almost a decade later, Linux is still at the forefront, pushing the envelope for supercomputing and setting the standard for open-source platforms. 90% of the world’s most sophisticated supercomputers top 500 list top 500 list Current HPC trends and cyber security Technology is constantly being upgraded. Developers and programmers never rest in their quest to find better solutions. With the integration of AI and machine learning, more doors have opened to researchers, and it is really up to the imagination how they apply the technology to create new and innovative solutions. In current trends, libraries like TensorFlow and PyTorch, which are optimized for Linux environments, are taking over because they are also open-sourced and rely on developers who work on underlying optimizations that enhance the performance of these machine learning algorithms. Optimization is not the only concern for HPC, though. The large amounts of data it consumes, as well as the A.I. models produced by the processing of this data are highly valuable assets and thus an increasingly common target for cybercrime and cyber security. Linux is again well positioned in this field, since having so many users using the platform means that vulnerabilities are quickly picked up and dealt with. Independently of the technique used - be it code review, static or dynamic analysis, or even automated A.I. solutions - detecting vulnerable paths before threats have a chance to manifest themselves is the goal. “A proactive approach is always better, as the saying goes, prevention is better than cure”, explains Gomes, who has been working with tech since graduating from University of Campinas and has since worked with IT hotshots like IBM and Linux distributor, SUSE. According to him, “It is important to continue fortifying the security of supercomputing systems for obvious reasons, particularly for maintaining uptime and preventing data breaches.” In this scenario, live patching is a new and valuable feature for supercomputing as it doesn’t interfere with uptime. During his time at SUSE, Gomes led the development of libpulp, a framework that allows the live patching of user-space processes, which can be used to safely apply changes to long-running computer programs. Since HPC requires performance, changes can also be reverted if their cost is deemed too high. This feature enhances the efficiency of patch management in production systems, and can be installed on any Linux distribution, explains Gomes, who is a Debian Developer in the Debian project. The Future of High-Performance Computing There is a growing reliance on supercomputing for sensitive research, which means that the performance and security of these systems will be the highest priority for future developments. The industry will continue to rely heavily on performance and security researchers that are able to focus on operating systems, compilers, libraries and patching mechanisms development for HPC. security of these systems security of these systems There will be increased collaboration across HPC domains, bringing together A.I., machine learning, and big data analytics. This also means that resource management will play a crucial role, as supercomputing consumes more resources than traditional computers. The development of sophisticated scheduling algorithms and workload management tools within the Linux ecosystem will further enhance resource availability and enable systems to hit peak performances. A driving force in the realm of high-performance computing, the Linux ecosystem will continue to excel and provide the technology needed to achieve greater efficiencies and capabilities, pushing forward the boundaries of what is possible in the future of computing.

Why Linux Leads the Charge in High-Performance Computing

Authors:
(1) Shih-Tang Su, University of Michigan, Ann Arbor (shihtang@umich.edu);
(2) Vijay G. Subramanian, University of Michigan, Ann Arbor and (vgsubram@umich.edu);
(3) Grant Schoenebeck, University of Michigan, Ann Arbor (schoeneb@umich.edu). Table of Links Abstract and 1. Introduction 2. Problem Formulation 2.1 Model of Binary-Outcome Experiments in Two-Phase Trials 3 Binary-outcome Experiments in Two-phase Trials and 3.1 Experiments with screenings 3.2 Assumptions and induced strategies 3.3 Constraints given by phase-II experiments 3.4 Persuasion ratio and the optimal signaling structure 3.5 Comparison with classical Bayesian persuasion strategies 4 Binary-outcome Experiments in Multi-phase trials and 4.1 Model of binary-outcome experiments in multi-phase trials 4.2 Determined versus sender-designed experiments 4.3 Multi-phase model and classical Bayesian persuasion and References 3.4 Persuasion ratio and the optimal signaling structure Careful readers may notice that if we multiply the persuasion ratio with the prior, the value will be the (maximum) expected utility the sender can achieve from the given set of strategies. Since the sender’s expected utility is monotone increasing in the prior p regardless of which set of strategies the sender adopts, the persuasion ratio under a given prior can be viewed as the relative utility gain this set of strategies can offer to the sender. Hence, given a specific prior, if a set of strategies has a higher persuasion ratio with respect to another set of strategies, the sender should use a strategy in the former instead of the latter. According to this discussion, we can draw a persuasion ratio curve for each set of strategies as the prior is varied in [0, 1]. Abusing notation, we represent the persuasion ratio curve by P R(S). It may appear that an optimization needs to be carried out for each value of the prior. However, structural insights presented in the following two lemmas considerably simplify the analysis. Properties presented in Lemma 3 narrow down the space where the sender needs to search for the optimal signaling strategy. This allows us to depict persuasion ratio curves P R(S) for some basic strategies. On top of that, Lemma 4 provides a systematic approach to derive persuasion ratio curves for all types of strategies. Lemma 3. Given a type of strategy S and a prior p, there exists a (sender’s) optimal strategy S ∈ S which satisfies one of the following two conditions: At least one IC requirement of the constituent induced strategies is tight;


There is a signal that will be sent with probability 1 under S. Before discussing the simplifications that Lemma 3 yields in terms of the key properties for solving the problem, we give an intuitive outline of the proof of Lemma 3. When the IC requirements of the two induced strategies are not tight and both signals are sent with non-zero probability in a strategy S, the sender can increase her expected utility by slightly raising the probability of the signal with a higher persuasion ratio (and adjust the probability of the other signal to respect the prior) to form a strategy S+. The sender can keep doing this ‘slight’ modification of her strategies until either one of the IC conditions is satisfied or the signal is sent with probability one. Lemma 5. The persuasion ratio curve of the optimal signaling strategy P R∗ (p) is the upper envelope of the different types of strategies’ persuasion curves. Further, the optimal signaling strategy (under a given prior) is the strategy that reaches the frontier of the persuasion ratio curve (at that prior). For a two-phase trial or to solve the last two phases of a trial with more than two phases studied in Section 4, the following corollary can further simplify the sender’s optimization procedure. With Corollary 1, the Equation (3) in Lemma 4 reduces to a linear equation. Hence, the comparison in Lemma 5 and the computation in Lemma 4 can be reduced to a comparison of the (unique) corresponding IC strategies (if one exists) under the interim belief listed in Corollary1 for different types of strategies. This paper is available on arxiv under CC 4.0 license. Authors: (1) Shih-Tang Su, University of Michigan, Ann Arbor (shihtang@umich.edu); (2) Vijay G. Subramanian, University of Michigan, Ann Arbor and (vgsubram@umich.edu); (3) Grant Schoenebeck, University of Michigan, Ann Arbor (schoeneb@umich.edu). Authors: Authors: (1) Shih-Tang Su, University of Michigan, Ann Arbor (shihtang@umich.edu); (2) Vijay G. Subramanian, University of Michigan, Ann Arbor and (vgsubram@umich.edu); (3) Grant Schoenebeck, University of Michigan, Ann Arbor (schoeneb@umich.edu). Table of Links Abstract and 1. Introduction Abstract and 1. Introduction 2. Problem Formulation 2. Problem Formulation 2.1 Model of Binary-Outcome Experiments in Two-Phase Trials 2.1 Model of Binary-Outcome Experiments in Two-Phase Trials 3 Binary-outcome Experiments in Two-phase Trials and 3.1 Experiments with screenings 3 Binary-outcome Experiments in Two-phase Trials and 3.1 Experiments with screenings 3.2 Assumptions and induced strategies 3.2 Assumptions and induced strategies 3.3 Constraints given by phase-II experiments 3.3 Constraints given by phase-II experiments 3.4 Persuasion ratio and the optimal signaling structure 3.4 Persuasion ratio and the optimal signaling structure 3.5 Comparison with classical Bayesian persuasion strategies 3.5 Comparison with classical Bayesian persuasion strategies 4 Binary-outcome Experiments in Multi-phase trials and 4.1 Model of binary-outcome experiments in multi-phase trials 4 Binary-outcome Experiments in Multi-phase trials and 4.1 Model of binary-outcome experiments in multi-phase trials 4.2 Determined versus sender-designed experiments 4.2 Determined versus sender-designed experiments 4.3 Multi-phase model and classical Bayesian persuasion and References 4.3 Multi-phase model and classical Bayesian persuasion and References 3.4 Persuasion ratio and the optimal signaling structure Careful readers may notice that if we multiply the persuasion ratio with the prior, the value will be the (maximum) expected utility the sender can achieve from the given set of strategies. Since the sender’s expected utility is monotone increasing in the prior p regardless of which set of strategies the sender adopts, the persuasion ratio under a given prior can be viewed as the relative utility gain this set of strategies can offer to the sender. Hence, given a specific prior, if a set of strategies has a higher persuasion ratio with respect to another set of strategies, the sender should use a strategy in the former instead of the latter. According to this discussion, we can draw a persuasion ratio curve for each set of strategies as the prior is varied in [0, 1]. Abusing notation, we represent the persuasion ratio curve by P R(S). It may appear that an optimization needs to be carried out for each value of the prior. However, structural insights presented in the following two lemmas considerably simplify the analysis. Properties presented in Lemma 3 narrow down the space where the sender needs to search for the optimal signaling strategy. This allows us to depict persuasion ratio curves P R(S) for some basic strategies. On top of that, Lemma 4 provides a systematic approach to derive persuasion ratio curves for all types of strategies. Lemma 3. Given a type of strategy S and a prior p, there exists a (sender’s) optimal strategy S ∈ S which satisfies one of the following two conditions: Lemma 3. Given a type of strategy S and a prior p, there exists a (sender’s) optimal strategy S ∈ S which satisfies one of the following two conditions: At least one IC requirement of the constituent induced strategies is tight; There is a signal that will be sent with probability 1 under S. At least one IC requirement of the constituent induced strategies is tight; At least one IC requirement of the constituent induced strategies is tight; At least one IC requirement of the constituent induced strategies is tight; There is a signal that will be sent with probability 1 under S. There is a signal that will be sent with probability 1 under S. There is a signal that will be sent with probability 1 under S. Before discussing the simplifications that Lemma 3 yields in terms of the key properties for solving the problem, we give an intuitive outline of the proof of Lemma 3. When the IC requirements of the two induced strategies are not tight and both signals are sent with non-zero probability in a strategy S, the sender can increase her expected utility by slightly raising the probability of the signal with a higher persuasion ratio (and adjust the probability of the other signal to respect the prior) to form a strategy S+. The sender can keep doing this ‘slight’ modification of her strategies until either one of the IC conditions is satisfied or the signal is sent with probability one. Lemma 5. The persuasion ratio curve of the optimal signaling strategy P R∗ (p) is the upper envelope of the different types of strategies’ persuasion curves. Further, the optimal signaling strategy (under a given prior) is the strategy that reaches the frontier of the persuasion ratio curve (at that prior). Lemma 5. The persuasion ratio curve of the optimal signaling strategy P R∗ (p) is the upper envelope of the different types of strategies’ persuasion curves. Further, the optimal signaling strategy (under a given prior) is the strategy that reaches the frontier of the persuasion ratio curve (at that prior). For a two-phase trial or to solve the last two phases of a trial with more than two phases studied in Section 4, the following corollary can further simplify the sender’s optimization procedure. With Corollary 1, the Equation (3) in Lemma 4 reduces to a linear equation. Hence, the comparison in Lemma 5 and the computation in Lemma 4 can be reduced to a comparison of the (unique) corresponding IC strategies (if one exists) under the interim belief listed in Corollary1 for different types of strategies. This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

How to Maximize Persuasion Ratios in Two-Phase Trials

Is ByteDance’s Lumi an AI Experiment Disguised as an AI Creator Community?

ByteDance's Lumi የ AI ሙከራ እንደ AI ፈጣሪ ማህበረሰብ ተደብቋል?

¿Es Lumi de ByteDance un experimento de IA disfrazado de comunidad de creadores de IA?

Èske Lumi ByteDance a se yon eksperyans AI degize kòm yon kominote kreyatè AI?

Արդյո՞ք ByteDance-ի Lumi-ն AI փորձ է, որը քողարկված է որպես AI Ստեղծողների համայնք:

ByteDance の Lumi は AI クリエイター コミュニティを装った AI 実験なのか?

ByteDance의 Lumi는 AI 크리에이터 커뮤니티로 위장한 AI 실험일까요?

ByteDance’s Lumi ezali Expérience ya AI oyo e déguisé lokola Communauté ya Créateur ya AI?

Lumi ຂອງ ByteDance ເປັນການທົດລອງ AI ປອມຕົວເປັນຊຸມຊົນຜູ້ສ້າງ AI ບໍ?

Adakah Lumi ByteDance merupakan Eksperimen AI yang Menyamar sebagai Komuniti Pencipta AI?

Je, Lumi ya ByteDance ni Jaribio la AI Limefichwa kama Jumuiya ya Watayarishi wa AI?

Оё Lumi byteDance як озмоиши AI аст, ки ҳамчун ҷомеаи Офарандаи AI пинҳон карда шудааст?

Ngaba i-ByteDance's Lumi liLingelo le-AI eliguqulelwe njengoluntu lwabadali be-AI?

Stories, News & Mentions about ENVELIO

Envelio’s Stories on HackerNoon

Envelio's News & Mentions