<strong>Scaler</strong> Academy said its operational revenue rose to Rs 316.7 crore in the financial year ended March 2023, from Rs 65.6 crore a year earlier. Total expenses at the firm went up to Rs 654.6 crore, significantly pushed up by employee expenses of Rs 322.07 crore and other expenses of Rs 316.93 crore.

sacler academy fy23 revenue: Scaler Academy posts five-fold rise in FY23 revenue to Rs 317 crore - The Economic Times

Scaler

scaler

Authors:
(1) Yuwei Guo, The Chinese University of Hong Kong;
(2) Ceyuan Yang, Shanghai Artificial Intelligence Laboratory with Corresponding Author;
(3) Anyi Rao, Stanford University;
(4) Zhengyang Liang, Shanghai Artificial Intelligence Laboratory;
(5) Yaohui Wang, Shanghai Artificial Intelligence Laboratory;
(6) Yu Qiao, Shanghai Artificial Intelligence Laboratory;
(7) Maneesh Agrawala, Stanford University;
(8) Dahua Lin, Shanghai Artificial Intelligence Laboratory;
(9) Bo Dai, The Chinese University of Hong Kong and The Chinese University of Hong Kong. Table of Links Abstract and 1 Introduction 2 Work Related 3 Preliminary AnimateDiff 4.1 Alleviate Negative Effects from Training Data with Domain Adapter 4.2 Learn Motion Priors with Motion Module 4.3 Adapt to New Motion Patterns with MotionLora 4.4 AnimateDiff in Practice 5 Experiments and 5.1 Qualitative Results 5.2 Qualitative Comparison 5.3 Ablative Study 5.4 Controllable Generation 6 Conclusion 7 Ethics Statement 8 Reproducibility Statement, Acknowledgement and References 5.3 ABLATIVE STUDY Domain adapter. To investigate the impact of the domain adapter in AnimateDiff, we conducted a study by adjusting the scaler in the adapter layers during inference, ranging from 1 (full impact) to 0 (complete removal). As illustrated in Figure 6, as the scaler of the adapter decreases, there is an improvement in overall visual quality, accompanied by a reduction in the visual content distribution learned from the video dataset (the watermark in the case of WebVid (Bain et al., 2021)). These results indicate the successful role of the domain adapter in enhancing the visual quality of AnimateDiff by alleviating the motion module from learning the visual distribution gap. Motion module design. We compare our motion module design of the temporal Transformer with its full convolution counterpart, which is motivated by the fact that both designs are widely employed in recent works on video generation. We replace the temporal attention with 1D temporal convolution and ensured that the two model parameters were closely aligned. As depicted in supplementary materials, the convolutional motion module aligns all frames to be identical but does not incorporate any motion compared to the Transformer architecture. Efficiency of MotionLoRA. The efficiency of MotionLoRA in AnimateDiff was examined in terms of parameter efficiency and data efficiency. Parameter efficiency is crucial for efficient model training and sharing among users, while data efficiency is essential for real-world applications where collecting an adequate number of reference videos for specific motion patterns may be challenging. To investigate these aspects, we trained multiple MotionLoRA models with varying parameter scales and reference video quantities. In Fig. 7, the first two samples demonstrate that MotionLoRA is capable of learning new camera motions (e.g., zoom-in) with a small parameter scale while maintaining comparable motion quality. Furthermore, even with a modest number of reference videos (e.g., N = 50), the model successfully learns the desired motion patterns. However, when the number of reference videos is excessively limited (e.g., N = 5), significant degradation in quality is observed, suggesting that MotionLoRA encounters difficulties in learning shared motion patterns and instead relies on capturing texture information from the reference videos. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Yuwei Guo, The Chinese University of Hong Kong; (2) Ceyuan Yang, Shanghai Artificial Intelligence Laboratory with Corresponding Author; (3) Anyi Rao, Stanford University; (4) Zhengyang Liang, Shanghai Artificial Intelligence Laboratory; (5) Yaohui Wang, Shanghai Artificial Intelligence Laboratory; (6) Yu Qiao, Shanghai Artificial Intelligence Laboratory; (7) Maneesh Agrawala, Stanford University; (8) Dahua Lin, Shanghai Artificial Intelligence Laboratory; (9) Bo Dai, The Chinese University of Hong Kong and The Chinese University of Hong Kong. Authors: Authors: (1) Yuwei Guo, The Chinese University of Hong Kong; (2) Ceyuan Yang, Shanghai Artificial Intelligence Laboratory with Corresponding Author; (3) Anyi Rao, Stanford University; (4) Zhengyang Liang, Shanghai Artificial Intelligence Laboratory; (5) Yaohui Wang, Shanghai Artificial Intelligence Laboratory; (6) Yu Qiao, Shanghai Artificial Intelligence Laboratory; (7) Maneesh Agrawala, Stanford University; (8) Dahua Lin, Shanghai Artificial Intelligence Laboratory; (9) Bo Dai, The Chinese University of Hong Kong and The Chinese University of Hong Kong. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Work Related 2 Work Related 3 Preliminary 3 Preliminary AnimateDiff AnimateDiff 4.1 Alleviate Negative Effects from Training Data with Domain Adapter 4.1 Alleviate Negative Effects from Training Data with Domain Adapter 4.2 Learn Motion Priors with Motion Module 4.2 Learn Motion Priors with Motion Module 4.3 Adapt to New Motion Patterns with MotionLora 4.3 Adapt to New Motion Patterns with MotionLora 4.4 AnimateDiff in Practice 4.4 AnimateDiff in Practice 5 Experiments and 5.1 Qualitative Results 5 Experiments and 5.1 Qualitative Results 5.2 Qualitative Comparison 5.2 Qualitative Comparison 5.3 Ablative Study 5.3 Ablative Study 5.4 Controllable Generation 5.4 Controllable Generation 6 Conclusion 6 Conclusion 7 Ethics Statement 7 Ethics Statement 8 Reproducibility Statement, Acknowledgement and References 8 Reproducibility Statement, Acknowledgement and References 5.3 ABLATIVE STUDY Domain adapter. To investigate the impact of the domain adapter in AnimateDiff, we conducted a study by adjusting the scaler in the adapter layers during inference, ranging from 1 (full impact) to 0 (complete removal). As illustrated in Figure 6, as the scaler of the adapter decreases, there is an improvement in overall visual quality, accompanied by a reduction in the visual content distribution learned from the video dataset (the watermark in the case of WebVid (Bain et al., 2021)). These results indicate the successful role of the domain adapter in enhancing the visual quality of AnimateDiff by alleviating the motion module from learning the visual distribution gap. Domain adapter. Motion module design. We compare our motion module design of the temporal Transformer with its full convolution counterpart, which is motivated by the fact that both designs are widely employed in recent works on video generation. We replace the temporal attention with 1D temporal convolution and ensured that the two model parameters were closely aligned. As depicted in supplementary materials, the convolutional motion module aligns all frames to be identical but does not incorporate any motion compared to the Transformer architecture. Motion module design. Efficiency of MotionLoRA. The efficiency of MotionLoRA in AnimateDiff was examined in terms of parameter efficiency and data efficiency. Parameter efficiency is crucial for efficient model training and sharing among users, while data efficiency is essential for real-world applications where collecting an adequate number of reference videos for specific motion patterns may be challenging. To investigate these aspects, we trained multiple MotionLoRA models with varying parameter scales and reference video quantities. In Fig. 7, the first two samples demonstrate that MotionLoRA is capable of learning new camera motions (e.g., zoom-in) with a small parameter scale while maintaining comparable motion quality. Furthermore, even with a modest number of reference videos (e.g., N = 50), the model successfully learns the desired motion patterns. However, when the number of reference videos is excessively limited (e.g., N = 5), significant degradation in quality is observed, suggesting that MotionLoRA encounters difficulties in learning shared motion patterns and instead relies on capturing texture information from the reference videos. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Ablative Study on Domain Adapter, Motion Module Design, and MotionLoRA Efficiency

Authors:
(1) Yuwei Guo, The Chinese University of Hong Kong;
(2) Ceyuan Yang, Shanghai Artificial Intelligence Laboratory with Corresponding Author;
(3) Anyi Rao, Stanford University;
(4) Zhengyang Liang, Shanghai Artificial Intelligence Laboratory;
(5) Yaohui Wang, Shanghai Artificial Intelligence Laboratory;
(6) Yu Qiao, Shanghai Artificial Intelligence Laboratory;
(7) Maneesh Agrawala, Stanford University;
(8) Dahua Lin, Shanghai Artificial Intelligence Laboratory;
(9) Bo Dai, The Chinese University of Hong Kong and The Chinese University of Hong Kong. Table of Links Abstract and 1 Introduction 2 Work Related 3 Preliminary AnimateDiff 4.1 Alleviate Negative Effects from Training Data with Domain Adapter 4.2 Learn Motion Priors with Motion Module 4.3 Adapt to New Motion Patterns with MotionLora 4.4 AnimateDiff in Practice 5 Experiments and 5.1 Qualitative Results 5.2 Qualitative Comparison 5.3 Ablative Study 5.4 Controllable Generation 6 Conclusion 7 Ethics Statement 8 Reproducibility Statement, Acknowledgement and References 4.4 ANIMATEDIFF IN PRACTICE We elaborate on the training and inference here and put the detailed configurations in supplementary materials. It’s worth noting that when training the domain adapter, the motion module, and the MotionLoRA, parameters outside the trainable part remain frozen. Inference. At inference time (Fig. 2), the personalized T2I model will first be inflated in the same way discussed in Section 4.2, then injected with the motion module for general animation generation, and the optional MotionLoRA for generating animation with personalized motion. As for the domain adapter, instead of simply dropping it during the inference time, in practice, we can also inject it into the personalized T2I model and adjust its contribution by changing the scaler α in Eq. (4). An ablation study on the value of α is conducted in experiments. Finally, the animation frames can be obtained by performing the reverse diffusion process and decoding the latent codes. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Yuwei Guo, The Chinese University of Hong Kong; (2) Ceyuan Yang, Shanghai Artificial Intelligence Laboratory with Corresponding Author; (3) Anyi Rao, Stanford University; (4) Zhengyang Liang, Shanghai Artificial Intelligence Laboratory; (5) Yaohui Wang, Shanghai Artificial Intelligence Laboratory; (6) Yu Qiao, Shanghai Artificial Intelligence Laboratory; (7) Maneesh Agrawala, Stanford University; (8) Dahua Lin, Shanghai Artificial Intelligence Laboratory; (9) Bo Dai, The Chinese University of Hong Kong and The Chinese University of Hong Kong. Authors: Authors: (1) Yuwei Guo, The Chinese University of Hong Kong; (2) Ceyuan Yang, Shanghai Artificial Intelligence Laboratory with Corresponding Author; (3) Anyi Rao, Stanford University; (4) Zhengyang Liang, Shanghai Artificial Intelligence Laboratory; (5) Yaohui Wang, Shanghai Artificial Intelligence Laboratory; (6) Yu Qiao, Shanghai Artificial Intelligence Laboratory; (7) Maneesh Agrawala, Stanford University; (8) Dahua Lin, Shanghai Artificial Intelligence Laboratory; (9) Bo Dai, The Chinese University of Hong Kong and The Chinese University of Hong Kong. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Work Related 2 Work Related 3 Preliminary 3 Preliminary AnimateDiff AnimateDiff 4.1 Alleviate Negative Effects from Training Data with Domain Adapter 4.1 Alleviate Negative Effects from Training Data with Domain Adapter 4.2 Learn Motion Priors with Motion Module 4.2 Learn Motion Priors with Motion Module 4.3 Adapt to New Motion Patterns with MotionLora 4.3 Adapt to New Motion Patterns with MotionLora 4.4 AnimateDiff in Practice 4.4 AnimateDiff in Practice 5 Experiments and 5.1 Qualitative Results 5 Experiments and 5.1 Qualitative Results 5.2 Qualitative Comparison 5.2 Qualitative Comparison 5.3 Ablative Study 5.3 Ablative Study 5.4 Controllable Generation 5.4 Controllable Generation 6 Conclusion 6 Conclusion 7 Ethics Statement 7 Ethics Statement 8 Reproducibility Statement, Acknowledgement and References 8 Reproducibility Statement, Acknowledgement and References 4.4 ANIMATEDIFF IN PRACTICE We elaborate on the training and inference here and put the detailed configurations in supplementary materials. It’s worth noting that when training the domain adapter, the motion module, and the MotionLoRA, parameters outside the trainable part remain frozen. Inference. At inference time (Fig. 2), the personalized T2I model will first be inflated in the same way discussed in Section 4.2, then injected with the motion module for general animation generation, and the optional MotionLoRA for generating animation with personalized motion. As for the domain adapter, instead of simply dropping it during the inference time, in practice, we can also inject it into the personalized T2I model and adjust its contribution by changing the scaler α in Eq. (4). An ablation study on the value of α is conducted in experiments. Finally, the animation frames can be obtained by performing the reverse diffusion process and decoding the latent codes. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

AnimateDiff in the Wild

Flighty App Can Now Predict Flight Delays: Here’s How You Can Also Do It Using Machine Learning

ফ্লাইটি অ্যাপ এখন ফ্লাইট বিলম্বের পূর্বাভাস দিতে পারে: এখানে আপনি কীভাবে মেশিন লার্নিং ব্যবহার করে এটি করতে পারেন

Die Flighty-App kann jetzt Flugverspätungen vorhersagen: So können Sie dies auch mithilfe von maschinellem Lernen tun

La aplicación Flighty ahora puede predecir retrasos en los vuelos: así es como puedes hacerlo tú también mediante aprendizaje automático

L’application Flighty peut désormais prédire les retards de vol : voici comment vous pouvez également le faire grâce au machine learning

फ्लाइटी ऐप अब उड़ान में देरी का अनुमान लगा सकता है: यहां बताया गया है कि आप मशीन लर्निंग का उपयोग करके ऐसा कैसे कर सकते हैं

Flighty アプリがフライトの遅延を予測できるようになりました。機械学習を使用してこれを行う方法は次のとおりです。

Flighty 앱이 이제 항공편 지연을 예측할 수 있습니다. 머신 러닝을 사용하여 예측하는 방법도 있습니다.

O aplicativo Flighty agora pode prever atrasos de voos: veja como você também pode fazer isso usando aprendizado de máquina

Приложение Flighty теперь может предсказывать задержки рейсов: вот как это можно сделать с помощью машинного обучения

Flighty Uygulaması Artık Uçuş Gecikmelerini Tahmin Edebiliyor: Makine Öğrenimini Kullanarak Bunu Nasıl Yapabileceğinizi Öğrenin

Ứng dụng Flighty hiện có thể dự đoán sự chậm trễ của chuyến bay: Đây là cách bạn cũng có thể thực hiện bằng cách sử dụng máy học

Flighty 应用程序现在可以预测航班延误：您也可以使用机器学习来实现这一点

Score Big With Football Analytics: Design Your Country's Perfect National Team with Machine Learning

Plunging Into Data: Unraveling Time Series Patterns

Depending on who you asked about the outlook of the crypto industry going into 2024, you would have heard a response that ranged between cautious optimism and more of the same from 2023. The prevailing bearish macroeconomic conditions of heightened inflation and elevated interest rates tempered expectations.Even with the potential for a final Bitcoin ETF approval in the U.S. and the next Bitcoin halving, many thought the headwinds were too strong and that the explosion in popularity of AI would siphon off whatever VC money was still available.As we enter Q2 of 2024, most people are shocked at how strong and resilient the industry is proving to be. The popularity of the Bitcoin ETFs with institutional investors and the calming of the Federal Reserve’s language has spurred renewed interest in the highly volatile asset class. But what can we learn from the mistakes of the last bull run, and how can we overcome the reputational damage that crypto scandals did while navigating this new regulatory landscape? How we got here It seems like a distant memory, but 2021 was a banner year for the industry. Most crypto projects were hitting all-time highs, staking and lending protocols like the Anchor Protocoloffered over 20% returns, and charismatic founders like Do Kwon, SBF, and Alex Maschinsky were being hailed by the mainstream media as the next coming of Steve Jobs.Yet, there were warning signs that the market sentiment was about to shift and that these new wunderkinder were nothing more than grifters and charlatans preying on overhyped market sentiment. Even before the bull run in 2021, analysts were highlighting the potential pitfalls of some of the most popular projects of the last cycle. In April 2018,  former Head of Risk at MakerDAO and Research Analyst at Scaler Cyrus Younessi correctly identified the critical failure point in Terra Luna: “If Terra were to fall and break peg, then it would depend on Luna to Save Terra. But Luna would fall as investors would panic, and then Terra would continue to fall, and then they would just each keep contributing to each other’s demise.” Collapses in 2022, bear market 2023 The euphoria that was felt across the market in 2021 would come to an abrupt end in 2022 from the fall of the first of many dominoes with the collapse of Terra Luna. Terra ecosystem In one of the most remarkable collapses of this period, the Terra ecosystem’s flagship protocol, LUNA, and algorithmic stablecoin (UST) lost its peg to the U.S. dollar, causing the price to crash to zero in a matter of days. LUNA’s fall from grace was made all the more spectacular because of how arrogant its founder Do Kwon was about promoting its strength and stability. Celcius Throughout much of 2021, Celsius was considered a premium centralized lending platform. Its charismatic founder, Alex Maschinsky, aggressively promoted its more than 20% return on investment, and it became the de facto custodian for retail and institutional investors. However, the model became unsustainable as the market shifted and the underlying assets began to lose value. It was revealed that Celcius was using Ponzi economics to prop up the platform. It was borrowing new deposits to make up for the lack of funds to pay out the high returns, and the whole platform collapsed under the weight, taking with it investors' deposits. FTX No project was as popular or successful during the 2021 bull run as the centralized exchange FTX. It rose to prominence because of its founder, Sam Bankman Fried’s reputation as a dominant arbitrage trader in his first company, Alameda Research. He was an enigma that the media fawned over as he embraced the lifestyle of a humble billionaire by spouting the benefits of the philosophical lifestyle of effective altruism. FTX became the number one exchange globally, signing massive celebrity endorsements from Larry David, Tom Brady, and Steph Curry. FTX had entrenched itself so deeply into the industry by requiring projects it supported to hold their assets on the platform that when it collapsed, it impacted just about everyone in the space. The once dominant exchange was ultimately undermined by Binance’s founder CZ, after he publicly sold Binance’s entire stake in the FTX utility token FTT after it was revealed SBF had been using customers' deposits to cover Alameda’s losses and using FTT to make up the difference on the balance sheet. Shifting Regulatory Landscape 2022 plunged the crypto industry into a deep bear market that we are only starting to emerge from today. The industry had largely been left to self-regulate, with the SEC only intervening in cases of clear criminality. Yet, the experiences of 2022 left many people heavily criticizing the SEC’s approach, specifically Chairman Gary Gensler and the close relationship he had with SBF and Alameda Research’s CEO Caroline Ellison.In a change of tone, the SEC and Chairman Gensler began pursuing legal action against anyone and everyone it could in an aggressive shift to a regulation policy by enforcement. The shift impacted individual projects like Ripple, Library, and Tornado Cash. The shift also meant constant deferral delays on SEC regulatory decisions, specifically in relation to Bitcoin ETF approvals, as well as targeting good faith actors like centralized exchanges__Kraken__ and Coinbase. Kraken and Coinbase had been asking for clear regulations for years to operate within. For their troubles of trying to remain compliant, the SEC filed lawsuits against each for operating unregistered security exchanges To end 2023, the SEC went trophy hunting and emerged with the arrest of the CEO of the world’s largest crypto exchange (Binance), CZ. CZ pled guilty to anti-money laundering charges and now faces over ten years in prison, while the exchange was fined $ 4 billion. The shift in policy by the SEC represented a direct attack on the industry. Still, it ultimately did more damage by undermining these firms' ability to operate in the U.S. by creating confusion and an environment of ever-shifting regulatory goalposts. Ongoing hacks and heists Despite the SEC's new approach to the industry, heists and hacks throughout the industry remained rampant throughout the course of 2023, totaling over $ 1.7 billion. While 2024 is again proving to be a hotbed for crypto hacks and scams, the first three months totaling over $437 million. With the looming Bitcoin halving and the historical trend of it spurring the next bull run, it is relatively safe to say that this problem will get worse before it gets better. Outlook for 2024 So, with all the negative sentiment built up in the industry from the last cycle. How can we emerge and market the industry in a positive light?If we learned anything from the last cycle it's that the product needs to be centerstage rather than any charismatic leader. It is easy to get swept up in the excitement when the market moves, but we need to remember the fundamentals of DYOR. Accountability for those bad actors of the last cycle is also key. We are beginning to see how the criminal activity of those wunderkind founders was not because they were in crypto but because they were criminals. SBF and Maschinsky committed embezzlement and financial crimes because they took advantage of user funds, not because of anything inherent with the technology. They are more akin to Bernie Madoff than Satoshi Nakamoto. Explaining the difference will be key. We should also take comfort in the fact that despite Chairman Gensler’s disdain for the industry, he will be limited to his term, other chairmen and women in the SEC have dissenting opinions (Hester Pierce) to the SEC, and his legal challenges to reign in the industry often fail. Depending on who you asked about the outlook of the crypto industry going into 2024, you would have heard a response that ranged between cautious optimism and more of the same from 2023. The prevailing bearish macroeconomic conditions of heightened inflation and elevated interest rates tempered expectations. Even with the potential for a final Bitcoin ETF approval in the U.S. and the next Bitcoin halving, many thought the headwinds were too strong and that the explosion in popularity of AI would siphon off whatever VC money was still available. As we enter Q2 of 2024, most people are shocked at how strong and resilient the industry is proving to be. The popularity of the Bitcoin ETFs with institutional investors and the calming of the Federal Reserve’s language has spurred renewed interest in the highly volatile asset class. But what can we learn from the mistakes of the last bull run, and how can we overcome the reputational damage that crypto scandals did while navigating this new regulatory landscape? How we got here It seems like a distant memory, but 2021 was a banner year for the industry. Most crypto projects were hitting all-time highs, staking and lending protocols like the Anchor Protocol offered over 20% returns, and charismatic founders like Do Kwon, SBF, and Alex Maschinsky were being hailed by the mainstream media as the next coming of Steve Jobs. Yet, there were warning signs that the market sentiment was about to shift and that these new wunderkinder were nothing more than grifters and charlatans preying on overhyped market sentiment. Anchor Protocol Anchor Protocol Even before the bull run in 2021, analysts were highlighting the potential pitfalls of some of the most popular projects of the last cycle. In April 2018,  former Head of Risk at MakerDAO and Research Analyst at Scaler Cyrus Younessi correctly identified the critical failure point in Terra Luna: correctly identified correctly identified “If Terra were to fall and break peg, then it would depend on Luna to Save Terra. But Luna would fall as investors would panic, and then Terra would continue to fall, and then they would just each keep contributing to each other’s demise.” “If Terra were to fall and break peg, then it would depend on Luna to Save Terra. But Luna would fall as investors would panic, and then Terra would continue to fall, and then they would just each keep contributing to each other’s demise.” “If Terra were to fall and break peg, then it would depend on Luna to Save Terra. But Luna would fall as investors would panic, and then Terra would continue to fall, and then they would just each keep contributing to each other’s demise.” Collapses in 2022, bear market 2023 The euphoria that was felt across the market in 2021 would come to an abrupt end in 2022 from the fall of the first of many dominoes with the collapse of Terra Luna. Terra ecosystem In one of the most remarkable collapses of this period, the Terra ecosystem’s flagship protocol, LUNA, and algorithmic stablecoin (UST) lost its peg to the U.S. dollar, causing the price to crash to zero in a matter of days. LUNA’s fall from grace was made all the more spectacular because of how arrogant its founder Do Kwon was about promoting its strength and stability. price to crash to zero price to crash to zero Celcius Throughout much of 2021, Celsius was considered a premium centralized lending platform. Its charismatic founder, Alex Maschinsky, aggressively promoted its more than 20% return on investment, and it became the de facto custodian for retail and institutional investors. However, the model became unsustainable as the market shifted and the underlying assets began to lose value. It was revealed that Celcius was using Ponzi economics to prop up the platform. It was borrowing new deposits to make up for the lack of funds to pay out the high returns, and the whole platform collapsed under the weight , taking with it investors' deposits. collapsed under the weight collapsed under the weight FTX No project was as popular or successful during the 2021 bull run as the centralized exchange FTX. It rose to prominence because of its founder, Sam Bankman Fried’s reputation as a dominant arbitrage trader in his first company, Alameda Research. He was an enigma that the media fawned over as he embraced the lifestyle of a humble billionaire by spouting the benefits of the philosophical lifestyle of effective altruism. FTX became the number one exchange globally, signing massive celebrity endorsements from Larry David, Tom Brady, and Steph Curry. FTX had entrenched itself so deeply into the industry by requiring projects it supported to hold their assets on the platform that when it collapsed , it impacted just about everyone in the space. The once dominant exchange was ultimately undermined by Binance’s founder CZ, after he publicly sold Binance’s entire stake in the FTX utility token FTT after it was revealed SBF had been using customers' deposits to cover Alameda’s losses and using FTT to make up the difference on the balance sheet. when it collapsed when it collapsed Shifting Regulatory Landscape 2022 plunged the crypto industry into a deep bear market that we are only starting to emerge from today. The industry had largely been left to self-regulate, with the SEC only intervening in cases of clear criminality. Yet, the experiences of 2022 left many people heavily criticizing the SEC’s approach, specifically Chairman Gary Gensler and the close relationship he had with SBF and Alameda Research’s CEO Caroline Ellison. In a change of tone, the SEC and Chairman Gensler began pursuing legal action against anyone and everyone it could in an aggressive shift to a regulation policy by enforcement. The shift impacted individual projects like Ripple, Library, and Tornado Cash. The shift also meant constant deferral delays on SEC regulatory decisions, specifically in relation to Bitcoin ETF approvals, as well as targeting good faith actors like centralized exchanges__ Kraken __ and Coinbase . Kraken and Coinbase had been asking for clear regulations for years to operate within. For their troubles of trying to remain compliant, the SEC filed lawsuits against each for operating unregistered security exchanges Kraken Coinbase Coinbase filed lawsuits filed lawsuits operating unregistered security exchanges operating unregistered security exchanges To end 2023, the SEC went trophy hunting and emerged with the arrest of the CEO of the world’s largest crypto exchange (Binance), CZ. CZ pled guilty to anti-money laundering charges and now faces over ten years in prison, while the exchange was fined $ 4 billion. The shift in policy by the SEC represented a direct attack on the industry. Still, it ultimately did more damage by undermining these firms' ability to operate in the U.S. by creating confusion and an environment of ever-shifting regulatory goalposts. Ongoing hacks and heists Despite the SEC's new approach to the industry, heists and hacks throughout the industry remained rampant throughout the course of 2023, totaling over $ 1.7 billion. totaling over $ 1.7 billion. totaling over $ 1.7 billion. While 2024 is again proving to be a hotbed for crypto hacks and scams, the first three months totaling over $437 million . With the looming Bitcoin halving and the historical trend of it spurring the next bull run, it is relatively safe to say that this problem will get worse before it gets better. totaling over $437 million totaling over $437 million Outlook for 2024 So, with all the negative sentiment built up in the industry from the last cycle. How can we emerge and market the industry in a positive light? If we learned anything from the last cycle it's that the product needs to be centerstage rather than any charismatic leader. It is easy to get swept up in the excitement when the market moves, but we need to remember the fundamentals of DYOR. Accountability for those bad actors of the last cycle is also key. We are beginning to see how the criminal activity of those wunderkind founders was not because they were in crypto but because they were criminals. SBF and Maschinsky committed embezzlement and financial crimes because they took advantage of user funds, not because of anything inherent with the technology. They are more akin to Bernie Madoff than Satoshi Nakamoto. Explaining the difference will be key. We should also take comfort in the fact that despite Chairman Gensler’s disdain for the industry, he will be limited to his term, other chairmen and women in the SEC have dissenting opinions (Hester Pierce) to the SEC, and his legal challenges to reign in the industry often fail. dissenting opinions dissenting opinions often fail. often fail.

Resilient Strategies for Crypto Marketing Success in 2024: Overcoming Regulations and Scandals

2024 সালে ক্রিপ্টো বিপণন সাফল্যের জন্য স্থিতিস্থাপক কৌশল: প্রবিধান এবং কেলেঙ্কারি কাটিয়ে ওঠা

Belastbare Strategien für erfolgreiches Krypto-Marketing im Jahr 2024: Regulierungen und Skandale überwinden

Estrategias resilientes para el éxito del cripto marketing en 2024: superación de regulaciones y escándalos

Stratégies résilientes pour réussir le marketing cryptographique en 2024 : surmonter les réglementations et les scandales

2024 में क्रिप्टो मार्केटिंग की सफलता के लिए लचीली रणनीतियाँ: विनियमनों और घोटालों पर काबू पाना

2024 年に暗号通貨マーケティングを成功させるための強靭な戦略: 規制とスキャンダルを克服する

2024년 암호화폐 마케팅 성공을 위한 탄력적 전략: 규제 및 스캔들을 극복

Estratégias resilientes para o sucesso do marketing criptográfico em 2024: superando regulamentações e escândalos

Устойчивые стратегии успеха криптомаркетинга в 2024 году: преодоление регулирования и скандалов

2024'te Kripto Pazarlama Başarısı için Dayanıklı Stratejiler: Düzenlemelerin ve Skandalların Üstesinden Gelmek

Các chiến lược kiên cường để tiếp thị tiền điện tử thành công vào năm 2024: Vượt qua các quy định và vụ bê bối

2024 年加密货币营销成功的弹性策略：克服监管和丑闻

Optimizing Data Center Efficiency: A Deep Dive into Freecooling Techniques

Optimaliseer datasentrumdoeltreffendheid: 'n diep duik in vryverkoelingstegnieke

የውሂብ ማእከልን ቅልጥፍናን ማሳደግ፡ ወደ ፍሪፍሪንግ ቴክኒኮች ጥልቅ ዘልቆ መግባት

تحسين كفاءة مركز البيانات: نظرة متعمقة على تقنيات التبريد الحر

Centro de Datos ukan Eficiencia ukar askinjam uñjaña: Técnicas de Freecooling ukar manqhan ch’allt’aña

Məlumat Mərkəzinin Səmərəliliyinin Optimallaşdırılması: Sərbəst Soyutma Texnikalarına Dərin Bir Dalış

Аптымізацыя эфектыўнасці цэнтра апрацоўкі дадзеных: глыбокае паглыбленне ў метады свабоднага астуджэння

Оптимизиране на ефективността на центъра за данни: Задълбочено потапяне в техниките за свободно охлаждане

Optimiziranje efikasnosti podatkovnog centra: Duboko uronjenje u tehnike slobodnog hlađenja

Optimització de l'eficiència del centre de dades: una immersió profunda en les tècniques de refrigeració lliure

Optimalizace efektivity datového centra: Hluboký ponor do technik volného chlazení

Optimering af datacentereffektivitet: Et dybt dyk i frikølingsteknikker

Βελτιστοποίηση της αποτελεσματικότητας του κέντρου δεδομένων: Μια βαθιά κατάδυση στις τεχνικές ελεύθερης ψύξης

Datu zentroen eraginkortasuna optimizatzea: Freecooling tekniketan murgiltze sakona

بهینه سازی کارایی مرکز داده: شیرجه عمیق در تکنیک های خنک کننده آزاد

Palvelinkeskuksen tehokkuuden optimointi: Sukella syvään vapaajäähdytystekniikoihin

Optimización da eficiencia do centro de datos: un mergullo profundo nas técnicas de freecooling

אופטימיזציה של יעילות מרכז הנתונים: צלילה עמוקה לתוך טכניקות קירור חופשי

Optimiziranje učinkovitosti podatkovnog centra: Duboko poniranje u tehnike slobodnog hlađenja

Optimize Efikasite Sant Done: Yon Plonje Pwofon nan Teknik Freecooling

Az adatközpontok hatékonyságának optimalizálása: mélyrepülés a szabadhűtési technikákba

Տվյալների կենտրոնի արդյունավետության օպտիմիզացում. խորը սուզում դեպի ազատ հովացման տեխնիկա

Mengoptimalkan Efisiensi Pusat Data: Analisis Mendalam tentang Teknik Freecooling

Ottimizzazione dell'efficienza del data center: un'analisi approfondita delle tecniche di freecooling

אופטימיזציה של יעילות מרכז הנתונים: צלילה עמוקה לטכניקות קירור חופשי

მონაცემთა ცენტრის ეფექტურობის ოპტიმიზაცია: ღრმა ჩაძირვა თავისუფალი გაგრილების ტექნიკაში

Деректер орталығының тиімділігін оңтайландыру: Freecooling әдістеріне терең бойлау

ការបង្កើនប្រសិទ្ធភាពមជ្ឈមណ្ឌលទិន្នន័យ៖ ការជ្រមុជទឹកជ្រៅទៅក្នុងបច្ចេកទេស Freecooling

Маалымат борборунун эффективдүүлүгүн оптималдаштыруу: Freecooling техникаларына терең сүңгүп чыгуу

Optimisation ya efficacité ya Centre ya ba données: Plongée profonde na ba techniques ya freecooling

ການເພີ່ມປະສິດທິພາບຂອງສູນຂໍ້ມູນ: ການລົງເລິກໃນເທັກນິກ Freecooling

Duomenų centro efektyvumo optimizavimas: giliai pasinerkite į nemokamo aušinimo būdus

Datu centra efektivitātes optimizēšana: dziļa iedziļināšanās brīvās dzesēšanas metodēs

Fanamafisana ny fahaiza-manaon'ny Foiben'ny angona: Fidirana lalina amin'ny teknikan'ny coolcooling

Оптимизирање на ефикасноста на центарот за податоци: длабоко нурнување во техниките на слободно ладење

Өгөгдлийн төвийн үр ашгийг оновчтой болгох нь: Freecooling техникийг гүнзгий судлах

Mengoptimumkan Kecekapan Pusat Data: Menyelam Dalam Teknik Penyejukan Percuma

डाटा सेन्टर दक्षता अनुकूलन: फ्रीकुलिङ प्रविधिहरूमा गहिरो डुब्नुहोस्

Optimaliseren van de efficiëntie van datacenters: een diepgaande duik in freecooling-technieken

Optimizing Data Center Bokgoni: A Tebileng Dive ka Freecooling Thekeniki ya

Optymalizacja wydajności centrum danych: dogłębne spojrzenie na techniki freecoolingu

د ډیټا سنټر موثریت اصلاح کول: د فریکولینګ تخنیکونو کې ژور ډوب

Willaypa Wasinpa allin ruwayninta allinchay: Huk ukhu buceo Freecooling Técnicas kaqpi

Optimizarea eficienței centrului de date: o scufundare profundă în tehnicile de răcire liberă

Gutezimbere Data Centre ikora neza: Kwibira cyane muburyo bwa Freecooling

දත්ත මධ්‍යස්ථාන කාර්යක්ෂමතාව ප්‍රශස්ත කිරීම: නිදහස් සිසිලන ශිල්පීය ක්‍රමවල ගැඹුරට කිමිදීම

Optimalizácia efektivity dátového centra: Hlboký ponor do techník voľného chladenia

Kugonesa Data Center Kubudirira: Kunyura Kwakadzika muFreecooling Techniques

Hagaajinta Hufnaanta Xarunta Xogta: Qulqul qoto dheer ee Farsamooyinka Qaboojinta Xorta ah

Optimizimi i efikasitetit të qendrës së të dhënave: Një zhytje e thellë në teknikat e ftohjes së lirë

Оптимизација ефикасности центра података: дубоко уроњење у технике слободног хлађења

Optimera datacentereffektiviteten: En djupdykning i frikylningstekniker

Kuboresha Ufanisi wa Kituo cha Data: Kuzama kwa Kina katika Mbinu za Kupoa Huru

டேட்டா சென்டர் செயல்திறனை மேம்படுத்துதல்: ஃப்ரீகூலிங் டெக்னிக்குகளில் ஒரு ஆழமான டைவ்

Оптимизатсияи самаранокии маркази додаҳо: Муборизаи амиқ ба усулҳои Freecooling

การเพิ่มประสิทธิภาพศูนย์ข้อมูล: การเจาะลึกเทคนิคการทำความเย็นแบบอิสระ

ምምሕያሽ ብቕዓት ማእከል ዳታ፡ ዓሚቕ ምጥላቕ ናብ ሜላታት ፍሪኩሊንግ

Maglumat merkeziniň netijeliligini optimizirlemek: Freecooling usullaryna çuňňur çümmek

Pag-optimize sa Kahusayan ng Data Center: Isang Malalim na Pagsusuri sa Mga Teknik sa Freecooling

Ku Antswisa Vukorhokeri bya Senthara ya Data: Ku Nwela loku dzikeke eka Tithekiniki ta Freecooling

Оптимізація ефективності центру обробки даних: глибоке занурення в методи вільного охолодження

ڈیٹا سینٹر کی کارکردگی کو بہتر بنانا: فری کولنگ تکنیکوں میں ایک گہرا غوطہ

Ma'lumotlar markazining samaradorligini optimallashtirish: Freecooling texnikasiga chuqur sho'ng'ish

Ukuphucula ukusebenza kakuhle kweZiko leeNkcukacha: Ukuntywila ngokuNxibelela kwiiTekhnoloji zokupholisa ngokukhululekileyo

Ukuthuthukisa Ukusebenza Kwesikhungo Sedatha: Ukungena Okujulile Kumasu Okupholisa Okukhululekile

The Evolution of Data Center Cooling: From Air-Based Methods to Free Cooling

Die evolusie van datasentrumverkoeling: van luggebaseerde metodes tot gratis verkoeling

የውሂብ ማእከል ማቀዝቀዝ ዝግመተ ለውጥ፡- ከአየር-ተኮር ዘዴዎች ወደ ነፃ ማቀዝቀዝ

تطور تبريد مراكز البيانات: من الأساليب المعتمدة على الهواء إلى التبريد المجاني

Uka Evolución del Centro de Datos de Cooling: Métodos Basados en Aire ukanakat Libre Cooling ukar puriñkama

Məlumat Mərkəzinin Soyutmasının Təkamülü: Hava Əsaslı Metodlardan Sərbəst Soyutmaya

Эвалюцыя астуджэння цэнтра апрацоўкі дадзеных: ад паветраных метадаў да вольнага астуджэння

Еволюцията на охлаждането на центъра за данни: от въздушни методи до свободно охлаждане

Evolucija hlađenja podatkovnih centara: od metoda na bazi zraka do slobodnog hlađenja

L'evolució de la refrigeració del centre de dades: dels mètodes basats en l'aire a la refrigeració lliure

Evoluce chlazení datových center: od vzduchových metod k volnému chlazení

Udviklingen af datacenterkøling: Fra luftbaserede metoder til frikøling

Η Εξέλιξη της Ψύξης του Κέντρου Δεδομένων: Από τις Μεθόδους Με βάση τον Αέρα στην Ελεύθερη Ψύξη

Datu-zentroen hoztearen bilakaera: airean oinarritutako metodoetatik free hozteraino

سیر تکامل خنک‌سازی مرکز داده: از روش‌های مبتنی بر هوا تا خنک‌سازی رایگان

Palvelinkeskusten jäähdytyksen kehitys: ilmapohjaisista menetelmistä ilmaiseen jäähdytykseen

A evolución do arrefriamento do centro de datos: dos métodos baseados en aire ao free cooling

האבולוציה של קירור מרכזי נתונים: משיטות מבוססות אוויר לקירור חופשי

Evolucija hlađenja podatkovnog centra: od zračnih metoda do slobodnog hlađenja

Evolisyon nan refwadisman sant done: Soti nan metòd ki baze sou lè rive nan refwadisman gratis

Az adatközponti hűtés fejlődése: a levegő alapú módszerektől a szabad hűtésig

Տվյալների կենտրոնի սառեցման էվոլյուցիան. օդի վրա հիմնված մեթոդներից մինչև անվճար սառեցում

Evolusi Pendinginan Pusat Data: Dari Metode Berbasis Udara ke Pendinginan Gratis

L'evoluzione del raffreddamento dei data center: dai metodi basati sull'aria al raffreddamento gratuito

მონაცემთა ცენტრის გაგრილების ევოლუცია: ჰაერზე დაფუძნებული მეთოდებიდან თავისუფალ გაგრილებამდე

Деректер орталығын салқындату эволюциясы: ауа негізіндегі әдістерден еркін салқындату

ការវិវឌ្ឍន៍នៃការធ្វើឱ្យត្រជាក់នៃមជ្ឈមណ្ឌលទិន្នន័យ៖ ពីវិធីសាស្ត្រផ្អែកលើខ្យល់ទៅភាពត្រជាក់ដោយឥតគិតថ្លៃ

Маалымат борборунун муздатуусунун эволюциясы: абага негизделген ыкмалардан эркин муздатууга

Evolution ya Refroidissement ya Centre de Données: Kobanda na ba méthodes basées na mopepe tii na refroidissement gratuit

ການວິວັດທະນາການຂອງສູນຂໍ້ມູນຄວາມເຢັນ: ຈາກວິທີການທີ່ອີງໃສ່ອາກາດໄປສູ່ຄວາມເຢັນຟຣີ

Duomenų centrų aušinimo evoliucija: nuo oru naudojamų metodų iki nemokamo aušinimo

Datu centra dzesēšanas evolūcija: no gaisa metodēm līdz bezmaksas dzesēšanai

Ny fivoaran'ny fampangatsiahana ivotoerana data: avy amin'ny fomba mifototra amin'ny rivotra mankany amin'ny fangatsiahana maimaim-poana

Еволуцијата на ладењето на центарот за податоци: од методи базирани на воздух до бесплатно ладење

Дата төвийн хөргөлтийн хувьсал: Агаарт суурилсан аргуудаас чөлөөт хөргөлт хүртэл

Evolusi Penyejukan Pusat Data: Daripada Kaedah Berasaskan Udara kepada Penyejukan Percuma

डाटा सेन्टर कूलिङको विकास: वायुमा आधारित विधिहरूबाट नि:शुल्क कूलिङसम्म

De evolutie van datacenterkoeling: van luchtgebaseerde methoden naar vrije koeling

Tlhagelelo ya go Fodiša ga Setsi sa Datha: Go Tšwa go Mekgwa ye e Theilwego Moyeng go ya go go Fodiša ga Mahala

Ewolucja chłodzenia centrów danych: od metod opartych na powietrzu do chłodzenia swobodnego

د ډیټا سنټر یخولو تکامل: د هوا پر اساس میتودونو څخه وړیا یخولو ته

Willaypa chawpinpi q'uñichiypa wiñaynin: Wayrapi ruwasqa ruwaykunamanta mana qullqiyuq q'uñichiyman

Evoluția răcirii centrului de date: de la metode pe bază de aer la răcire liberă

Ubwihindurize bwa Data Centre Cooling: Kuva Muburyo bushingiye ku kirere kugeza ubukonje bwubusa

දත්ත මධ්‍යස්ථාන සිසිලනයේ පරිණාමය: වායු පාදක ක්‍රමවල සිට නොමිලේ සිසිලනය දක්වා

Evolúcia chladenia dátových centier: od metód založených na vzduchu po voľné chladenie

Iko Evolution yeData Center Kutonhodza: Kubva kuMhepo-Yakavakirwa Nzira kuenda Kumahara Kutonhodza

Kobcinta Qaboojinta Xarunta Xogta: Laga soo bilaabo Hababka Hawada Ku Salaysan ilaa Qaboojinta Bilaashka ah

Evolucioni i ftohjes së qendrës së të dhënave: Nga metodat e bazuara në ajër te ftohja falas

Еволуција хлађења центара података: од ваздушних метода до бесплатног хлађења

Utvecklingen av kylning av datacenter: från luftbaserade metoder till fri kylning

Mageuzi ya Kupoeza kwa Kituo cha Data: Kutoka Mbinu Zinazotegemea Hewa hadi Kupoeza Bila Malipo

தரவு மைய குளிரூட்டலின் பரிணாமம்: காற்று அடிப்படையிலான முறைகளிலிருந்து இலவச குளிரூட்டலுக்கு

Эволютсияи сардшавии маркази додаҳо: аз усулҳои ҳавоӣ то сардшавии ройгон

วิวัฒนาการของระบบทำความเย็นศูนย์ข้อมูล: จากวิธีการใช้ลมสู่ระบบทำความเย็นแบบอิสระ

ፍልቀት መዝሓሊ ማእከል ዳታ፡ ካብ ኣየር መሰረት ዝገበረ ሜላታት ናብ ነጻ ምዝሓል

Maglumat merkeziniň sowadyşynyň ewolýusiýasy: Howa esasly usullardan mugt sowatmaga çenli

Ang Ebolusyon ng Data Center Cooling: Mula sa Air-Based Methods hanggang sa Libreng Paglamig

Ku Hundzuka ka Ku Pfuxeta ka Data Center: Ku suka eka Maendlelo lama Sekeriweke eMoyeni ku ya eka ku Pfuxeta ka Mahala

Еволюція охолодження центрів обробки даних: від повітряних методів до вільного охолодження

ڈیٹا سینٹر کولنگ کا ارتقاء: ہوا پر مبنی طریقوں سے مفت کولنگ تک

Ma'lumotlar markazini sovutishning evolyutsiyasi: havoga asoslangan usullardan erkin sovutishgacha

I-Evolution yeZiko leeNkcukacha zokuPholisa: Ukusuka kwiiNdlela eziSekwe kuMoya ukuya kwi-Free Cooling

Ukuvela Kokupholisa Kwesikhungo Sedatha: Kusuka Ezindleleni Ezisekelwe Emoyeni Kuya Ekupholiseni Kwamahhala

The future of AI chips is about more than NVIDIA: AMD, Intel, chiplets, upstarts, analog AI, optical computing, and AI chips designed by AI. The interest and investment in AI is skyrocketing, and generative AI is fueling it. Over one-third of CxOs have already embraced GenAI in their operations, with nearly half preparing to invest in it. reportedly What’s powering it all – AI chips – used to receive less attention. Up to the moment, OpenAI’s for a “wildly-ambitious” tech project to boost the world’s chip capacity. Sam Altman claimed he wants to raise up to $7 trillion and sensationalism aside, however, keeping an eye on AI chips means being aware of today’s blockers and tomorrow’s opportunities. Geopolitics According to a , the global AI chip market is expected to reach $89.6 Billion by 2029. The demand for AI chips has increased substantially over time. Growth in AI technology, rising demand for AI chips in consumer electronics, and AI chip innovation all contribute to this forecast. recent study by IMARC Few people have more insights to share on AI hardware than . In an extensive conversation, Pialis shared his insider’s perspective on the AI chip landscape, the transformative rise of chiplets, specialized hardware for training and inference, the future of AI chips and emerging directions like analog and optical computing, and much more. Tony Pialis, CEO and co-founder of Alphawave An AI Leader Behind the Scenes Pialis is a semiconductor serial entrepreneur. He has previously founded and successfully sold two startups: Snowbush Microelectronics and V Semiconductor Inc. V Semiconductor was acquired by Intel in 2012, where Pialis spent a few years as the Vice President of Analog and Mixed-Signal IP. Pialis co-founded AlphaWave in 2017 with the goal of becoming “the next great semiconductor company.” Alphawave went public in 2021 with an impressive market cap of $4.5 billion. The company’s core products include silicon IP, chips, custom silicon, and high-speed interconnect technology tailored for major hyper-scaler customers like Google, Microsoft, Amazon, and Meta. Alphawave is an AI leader behind the scenes, powering hyper scalers and founded on a couple of key premises that Pialis laid out. https://pod.co/orchestrate-all-the-things-podcast-connecting-the-dots-with-george-anadiotis/the-future-of-ai-hardware-leaders-dark-horses-and-rising-stars-featuring-tony-pialis-alphawave-ceo-co-founder?embedable=true , the major players in terms of scaling out data centers and computing these days are no longer the Ciscos of the world. It’s transitioned to Hyperscalers – the Googles, Microsofts, Amazons, and Metas. Hyperscalers have internal design capability, and they also build their own servers, networks, data centers, and campuses. First , the major challenge facing AI moving forward is not computing. Pialis’ view is that the ability to design and implement compute is there. The challenge is the connectivity needed to connect the compute in order to process all of the data. This is what AlphaWave is focusing on. Second The Explosion of Specialized AI Hardware While consumer apps like ChatGPT drove the initial hype cycle in 2023, enterprise adoption reports are mixed. However, according to Pialis, the AI semiconductor sector has seen an enormous explosion of investment and new in the second half of 2023 across industries and geographies. design starts Pialis cited major national initiatives around building domestic AI chip capabilities in the US, Canada, UK, France, Germany, Korea, Taiwan and Japan. After years of relying largely on NVIDIA, governments want to foster their own chip industries to reduce strategic dependence on any one vendor. While NVIDIA CEO Jensen Huang also says that , it does not seem like this also includes the hardware layer. every country needs Sovereign AI Pialis believes this surging demand is spurring not just startups but also tech giants to develop specialized training and inference hardware. In his view, not every organization could or should develop its own AI model, but that’s bound to change. “Over time, AI will inevitably progress to be something like a utility, where hyperscalers will provide access to all of the compute power, like electricity. I think it will be reasonably priced. Anyone will be able to use this utility to train, develop, optimize and deploy their own models. But there’s going to be a lot of profit taking between now and then, and who knows how long it’ll take to eventually reach that state”, said Pialis. Leaders, Dark Horses, and Rising Stars in the AI Chip Race Few people would disagree that NVIDIA is clearly the dominant force in AI accelerators today, and Pialis shares this view. He sees AMD as a strong second place and praised CEO Lisa Su for her leadership. Others such as Ben Lorica also see . favorable winds for AMD in the GenAI chip market But Pialis cautioned not to underestimate Intel, calling it a dark horse that could still prove highly formidable in this market. More specifically, Pialis spoke highly of . Intel’s Habana acqui-hired division led by David Dahan Having worked at Intel, Pialis has an insider view. Having , and kept track of their , we tend to agree. Dahan helped architect new Intel processors that have showcased performance exceeding NVIDIA’s latest GPUs on key benchmarks. covered Habana, met Dahan MLPerf results But performance is not everything. As Pialis also noted, NVIDIA’s software platform including CUDA gives it a huge advantage in the AI chip segment. The ecosystem effects are strong, as engineers and researchers develop optimized frameworks and models for NVIDIA’s architecture. However, there is potential for alternatives; for example, . Pialis opined that a rule of thumb for AI hardware companies is that for every hardware engineer, they need 2-3 software engineers. So even though there’s a lot of talk about NVIDIA and hardware, the vast majority of the investment actually goes into the software. building upon AMD’s hardware and software stack https://www.youtube.com/watch?v=VYx1rGieDmE&embedable=true This was confirmed by NVIDIA’s Dave Salvator. At the , Salvator said that NVIDIA has twice as many software engineers as hardware engineers. That’s not at all by accident, as he emphasized. last MLPerf results briefing for 2023 However, Pialis sees far more potential for challengers in . The reason is that standards are still emerging in that area. As OctoML’s Luis Ceze , there’s for LLM serving, for portable deployment, and and for multiplexed fined-tuned model inference. Inference, is in fact, a bigger market than training as Pialis noted. the inference accelerator market shared vLLM MLC-LLM LoRAX Punica “There’s so much focus on training, large models and the cost of training. But it’s deployment where we all benefit from, on the inference side. That needs significant scale. That needs a different solution. More chips will inevitably be sold on the inference side than the training side. I’m sure business plans also improve with that increase in volume”, Pialis said. Startups like and are attracting significant funding, and so are companies from countries like the UK, Korea, and China trying to reduce reliance on US firms. As far as hyperscalers go, Pialis sees Amazon and Google as being neck to neck, with Microsoft showing strong momentum, and Meta being behind and rumored to acquire one of the smaller startups. Groq Tenstorrent Join the Orchestrate All the Things Newsletter Stories about how Technology, Data, AI, and Media flow into each other shaping our lives. Analysis, Essays, Interviews, and News. Mid-to-long form, 1-3 times per month. The Chiplet Technological Revolution One of the biggest transformations underway in semiconductors, according to Pialis, is the shift to . Previously, progress meant integrating more and more functionality into a single monolithic chip. But transistor dimensions are now down to ~5 atoms wide, and even tiny defects can ruin a whole chip. chiplets Pialis shared an anecdote from his own experience to highlight this. As he related, when he visited OpenAI at some point, he saw a bunch of engineers on their knees praying in front of servers. Not because they were “ ”, but rather because they didn’t want the model they were training to crash. feeling the AGI Some of us may have first heard about chiplets in the context of the China–US trade war that spills over to technology. Chiplets are a key element in the strategy of both and the . China US For Pialis, “chiplets are another technological revolution to deal with the problem that God has thrown at us.” That may sound like a wild appeal to the supernatural, but it has more to do with the challenges of physics at nanoscale. “When we build transistors, which are the basic building blocks of any integrated circuit, we’re stacking atoms. When you’re stacking atoms, the laws of probability, the laws of averaging, fall apart, because now it’s two atoms rather than hundreds of atoms and electrons. So what you get is defects”, Pialis explained. Chiplets are an innovative solution to this nanoscale physics challenge. Rather than one giant chip, designs are split into smaller Lego-like chiplets that can be connected through advanced packaging. This modularity lets chipmakers avoid throwing out entire devices due to one small flaw. The benefits, Pialis said, are important for both makers and buyers. “Silicon is no longer the foundation of leading edge semiconductors; it’s the package. Silicon is now something that is a component on this packaging. There’s been a lot of talk about semiconductor supply chains. There’s a lot of silicon capacity now. Where there’s still very little to no capacity is on the packaging, specifically for designs built using chiplets”, Pialis noted. Chiplets as Lego Building Blocks However, there is a standout example of an AI hardware company that seems to be successful, despite going the opposite way: Cerebras. Cerebras is known for its wafer-scale hardware. Pialis was positive that Cerebras was going to encounter the same defects as everyone else. The difference with Cerebras is they have redundancy. A wafer is like a panel, Pialis noted. Cerebras doesn’t even cut the panel. What they do is they have software that works around the defects. Fundamentally, they’re using silicon as the foundation rather than packaging. They have multiple chips that are arrayed, but they’re not cut. They just connect on the wafer. It definitely is a different approach. However, Pialis added, there are advantages to cutting things up. Imagine a vendor like Intel is trying to build different units such as a CPU, a GPU, a DPU, or a networking device. By breaking things up into smaller pieces, they become like Lego building blocks. So you can have a processor core chiplet, a PCI Express connectivity chiplet, an Ethernet networking chiplet, a DDR memory I/O chiplet, and a memory I/O chiplet. These chiplets can be mixed and matched in a package to build out an entire product portfolio. Pialis thinks that from a design complexity and upfront investment perspective, that’s a winning formula. Pialis sees massive benefits in chiplets, reducing costs by over 60% and power consumption by 40%. This is a major incentive for hyperscale data centers. While he sees Apple, AMD, and Intel leading in chiplets today, Pialis believes chiplets will soon become essential for any company focused on leading-edge hardware. Join the Orchestrate All the Things Newsletter Stories about how Technology, Data, AI, and Media flow into each other shaping our lives.Analysis, Essays, Interviews, and News. Mid-to-long form, 1-3 times per month. Software and Chiplet Modularity, Composability, Programmability Drawing parallels from the world of software engineering, one can wonder how come the modularity that chiplets bring to the table did not prevail earlier. But in hardware, up to now, the winners were the ones that could integrate into a monolithic device the most. The reason is that this drove down cost, which Pialis noted has created “a maniacal focus on integration.” However, as we now approach the dimensions of a single atom, the manufacturing cost eclipses the cost of integration. But even in software, . excessive modularity can also incur excessive overhead Pialis expects that some hardware vendors may take the chiplet approach too far. If functionality is broken into too many tiny pieces, the cost of integrating all of these becomes limiting. So ultimately, it’ll be a hybrid approach that wins. There are two ways of splitting things up with chiplets. The first way of building chiplets is building a single Lego block, mirroring it over and over with the same functionality, and they just talk to each other. That seems more akin to the “traditional” approach. However, one may wonder how these identical chiplet building blocks can be synthesized. The answer is, again, software. Different packages can be composed for different needs, based on the same chiplet, using the software. For example, multiples of 1, 2, 4, or 8 chiplets. The same silicon, just packaged differently, priced differently, and with different software to leverage the incremental compute and memory bandwidth associated with these devices. The other way of building chiplets is by splitting and cutting, creating Lego blocks for different types of functionality. This creates chiplets like a compute Lego block, a training I/O Lego block, a network I/O Lego block, and so on. Pialis sees more momentum behind this approach, as reusable Lego blocks can accelerate other products too. Analog AI, Optical Computing, and AI-Assisted Hardware Design Chiplets may be the immediate way forward. But at a time when power-hungry digital accelerators like GPUs dominate, Pialis claims there’s more to the future of AI chips. And that involves a radical break with the current paradigm. AI needs massively parallelized arithmetic processing, and binary is the dominant approach for this. Using binary compute, numbers are represented as ones and zeros, and floating point arithmetic is done defining precision and range. In analog arithmetic processing, floating point can be represented via voltages or currents. In theory, that may have infinite precision. In the real world where noise exists, that may yield less precision. But that may well work for applications. Using very small currents keeps things operating at low power. Edge AI https://pod.co/orchestrate-all-the-things-podcast-connecting-the-dots-with-george-anadiotis/the-future-of-ai-hardware-leaders-dark-horses-and-rising-stars-featuring-tony-pialis-alphawave-ceo-co-founder?embedable=true There’s another form of compute that some companies are investing in: optical compute for arithmetic. Optical compute uses the properties of optics to implement what are called (multiply & accumulate functionality) which is the core of any arithmetic unit. This results in even lower power demands. MACs Pialis noted that analog and optical approaches are attracting billions in investment to meet specialized needs around computational scale, energy efficiency, and precision. However, it remains unclear if analog techniques can scale effectively to match digital compute for cutting-edge AI models. There is active debate on this question in the hardware community. Even though these techniques are often dismissed as impractical, Pialis believes the investment will eventually pay off. For example, recently, it was reported that capable of processing computer vision tasks with remarkable speed and energy efficiency. researchers at Tsinghua University have developed the world’s first super all-analog photoelectric chip Finally, there’s also the question of . Pialis said that the most effective hardware designers these days are the best software developers. But if AI models can be trained by leveraging the wealth of their experience, that may bring about a sea change. using AI to help design hardware to power AI While the future of AI chips is unknown, Pialis believes the fundamental principles of engineering are timeless. Hopefully, that translates to something that does not end up draining the world’s energy and resources on the promise of future gains to be made.

What Is the Future of AI Chips? Leaders, Dark Horses, and Rising Stars

Multi-Class Classification: Understanding Activation and Loss Functions in Neural Networks

মাল্টি-ক্লাস শ্রেণীবিভাগ: নিউরাল নেটওয়ার্কগুলিতে সক্রিয়করণ এবং ক্ষতি কার্যকারিতা বোঝা

Mehrklassenklassifizierung: Aktivierungs- und Verlustfunktionen in neuronalen Netzen verstehen

Clasificación de clases múltiples: comprensión de las funciones de activación y pérdida en redes neuronales

Classification multiclasse : comprendre les fonctions d'activation et de perte dans les réseaux de neurones

मल्टी-क्लास वर्गीकरण: तंत्रिका नेटवर्क में सक्रियण और हानि कार्यों को समझना

マルチクラス分類: ニューラル ネットワークの活性化関数と損失関数を理解する

다중 클래스 분류: 신경망의 활성화 및 손실 함수 이해

Classificação multiclasse: Compreendendo as funções de ativação e perda em redes neurais

Классификация нескольких классов: понимание функций активации и потерь в нейронных сетях

Çok Sınıflı Sınıflandırma: Sinir Ağlarında Aktivasyon ve Kayıp Fonksiyonlarını Anlamak

Phân loại nhiều lớp: Tìm hiểu các chức năng kích hoạt và mất mát trong mạng thần kinh

多类分类：了解神经网络中的激活函数和损失函数

Stories, News & Mentions about SCALER

Scaler’s Stories on HackerNoon

Scaler's News & Mentions