Video wird geladen...

Video konnte nicht geladen werden

Beim Laden dieses Videos ist ein Problem aufgetreten. Dies könnte an einem vorübergehenden Netzwerkproblem liegen oder das Video ist möglicherweise nicht verfügbar.

Delphi AI researcher Accelxr 👾 explains why China’s AI progress is more than just distillation. “Distillation can explain a large chunk of the capability and benchmark side, but it can’t explain a lot of the things they’ve shipped... The research is quite impressive.”

Delphi Digital

1,512,925 subscribers

37,686 Aufrufe • vor 24 Tagen •via X (Twitter)

Wissenschaft & Technologie Bildung Nachrichten & Politik

Anya Rossi• Live Now

Private livecam show

0 Kommentare

Keine Kommentare verfügbar

Kommentare vom Original-Post werden hier angezeigt

Ähnliche Videos

Delphi AI researcher Accelxr 👾 explains why inference could strengthen China’s position despite its compute disadvantage. “If the axis of competition has actually shifted from the training side to the serving side, all of these innovations... make them better situated.”

Delphi AI researcher Accelxr 👾 explains why inference could strengthen China’s position despite its compute disadvantage. “If the axis of competition has actually shifted from the training side to the serving side, all of these innovations... make them better situated.”

Delphi Digital

38,776 Aufrufe • vor 20 Tagen

Delphi AI Researcher Accelxr 👾 discusses how China’s AI talent is returning home. “Talent flow has always been America’s advantage. We import talent and they’re more likely to stay. But we’re starting to see signs of reversal.”

Delphi AI Researcher Accelxr 👾 discusses how China’s AI talent is returning home. “Talent flow has always been America’s advantage. We import talent and they’re more likely to stay. But we’re starting to see signs of reversal.”

Delphi Digital

49,033 Aufrufe • vor 24 Tagen

Delphi AI researcher Accelxr 👾 explains how China is scaling AI compute with domestic chips. “These superclusters take a ton of Ascend chips and combine them together. The supercluster itself has, I think, over 500k Ascend chips. They're pushing on it for sure.”

Delphi AI researcher Accelxr 👾 explains how China is scaling AI compute with domestic chips. “These superclusters take a ton of Ascend chips and combine them together. The supercluster itself has, I think, over 500k Ascend chips. They're pushing on it for sure.”

Delphi Digital

30,296 Aufrufe • vor 12 Tagen

Delphi AI researcher Accelxr 👾 explains why AI agents could favor cheaper open weights models. “Agentic use cases are extremely cost sensitive. And so if that’s the case, you kind of have this composition shift towards cheapest, but still adequately capable models.”

Delphi AI researcher Accelxr 👾 explains why AI agents could favor cheaper open weights models. “Agentic use cases are extremely cost sensitive. And so if that’s the case, you kind of have this composition shift towards cheapest, but still adequately capable models.”

Delphi Digital

24,476 Aufrufe • vor 12 Tagen

Delphi AI researcher Accelxr 👾 explains why Nvidia’s move into open models could make it a competitor to its own customers. “Nvidia is best positioned to extract maximum throughput for their models out of their own tech stack.”

Delphi AI researcher Accelxr 👾 explains why Nvidia’s move into open models could make it a competitor to its own customers. “Nvidia is best positioned to extract maximum throughput for their models out of their own tech stack.”

Delphi Digital

22,422 Aufrufe • vor 17 Tagen

Delphi AI researcher Accelxr 👾 explains how Chinese labs are optimizing AI for domestic chips. “The Chinese labs have actually been going beneath that, handwriting low-level GPU instructions and exploiting some of the chip behaviors at a lower level.”

Delphi AI researcher Accelxr 👾 explains how Chinese labs are optimizing AI for domestic chips. “The Chinese labs have actually been going beneath that, handwriting low-level GPU instructions and exploiting some of the chip behaviors at a lower level.”

Delphi Digital

40,080 Aufrufe • vor 13 Tagen

Delphi AI researcher Accelxr 👾 explains why restricting frontier AI model access can create its own security risk. “If you’re not a company that gets access to it, you lose the ability to defend yourself against it when it’s released... Now it becomes a race immediately after it is released to harden all the infrastructure.”

Delphi AI researcher Accelxr 👾 explains why restricting frontier AI model access can create its own security risk. “If you’re not a company that gets access to it, you lose the ability to defend yourself against it when it’s released... Now it becomes a race immediately after it is released to harden all the infrastructure.”

Delphi Digital

24,330 Aufrufe • vor 18 Tagen

New Delphi Podcast with Accelxr 👾, author of our latest report on China’s AI labs He went into the research thinking China was mostly distilling US models and benchmark-maxing. He came out with a very different view We cover: – Why DeepSeek is doing real research, not just copying – How Chinese labs are optimizing below CUDA – Why China’s electricity advantage could matter more than chips – Whether cheap open-source models are part of a broader strategy – Why more Chinese AI talent is returning home – Whether restricting NVIDIA chips could actually accelerate China’s independence – What happens if China reaches full AI training sovereignty 00:00 Intro 00:35 Is China Innovating or Just Distilling? 06:10 DeepSeek and China’s Leading AI Labs 10:15 How China Is Working Around Weaker Chips 18:25 China’s Biggest AI Advantage: Electricity 29:45 Three Scenarios for the US-China AI Race 40:20 America’s AI Talent Advantage Is Reversing 47:45 China’s Path to AI Sovereignty Full episode:

New Delphi Podcast with Accelxr 👾, author of our latest report on China’s AI labs He went into the research thinking China was mostly distilling US models and benchmark-maxing. He came out with a very different view We cover: – Why DeepSeek is doing real research, not just copying – How Chinese labs are optimizing below CUDA – Why China’s electricity advantage could matter more than chips – Whether cheap open-source models are part of a broader strategy – Why more Chinese AI talent is returning home – Whether restricting NVIDIA chips could actually accelerate China’s independence – What happens if China reaches full AI training sovereignty 00:00 Intro 00:35 Is China Innovating or Just Distilling? 06:10 DeepSeek and China’s Leading AI Labs 10:15 How China Is Working Around Weaker Chips 18:25 China’s Biggest AI Advantage: Electricity 29:45 Three Scenarios for the US-China AI Race 40:20 America’s AI Talent Advantage Is Reversing 47:45 China’s Path to AI Sovereignty Full episode:

Tommy

81,587 Aufrufe • vor 25 Tagen

Delphi AI researcher Accelxr 👾 explains why access to Nvidia chips could slow China’s push for chip independence. “Getting them back on the Nvidia bridge should slow down their development, kind of paradoxically. It slows their sovereignty push, and any tightening actually accelerates their push to build out their own chips.”

Delphi AI researcher Accelxr 👾 explains why access to Nvidia chips could slow China’s push for chip independence. “Getting them back on the Nvidia bridge should slow down their development, kind of paradoxically. It slows their sovereignty push, and any tightening actually accelerates their push to build out their own chips.”

Delphi Digital

20,562 Aufrufe • vor 19 Tagen

AI is about more than models. Ariana Salvatore, Head of Public Policy Research at Morgan Stanley, explains why control of the AI stack is becoming a strategic priority. Listen to more on Thoughts on the Market.

AI is about more than models. Ariana Salvatore, Head of Public Policy Research at Morgan Stanley, explains why control of the AI stack is becoming a strategic priority. Listen to more on Thoughts on the Market.

Morgan Stanley

17,299 Aufrufe • vor 2 Monaten

Andrej Karpathy just broke the entire premise of modern AI: "Agents aren't magic. They're distillation at scale." 99.99% of your LLM's capacity is wasted on garbage data it never needed. Small model + right tools + closed loop = terrifying capability. In a 16-minute conversation, Karpathy reveals the full reasoning stack. Worth more than any $500 AI course you've seen this year.

Andrej Karpathy just broke the entire premise of modern AI: "Agents aren't magic. They're distillation at scale." 99.99% of your LLM's capacity is wasted on garbage data it never needed. Small model + right tools + closed loop = terrifying capability. In a 16-minute conversation, Karpathy reveals the full reasoning stack. Worth more than any $500 AI course you've seen this year.

Morty

674,874 Aufrufe • vor 2 Monaten

Can anyone explain why Netanyahu is threatening the west and why people are ok with it? They’ll call me a lot of things but LIAR is never one of them. By the way isn’t it weird that this psychopath always says things like this when he’s not speaking English?

Can anyone explain why Netanyahu is threatening the west and why people are ok with it? They’ll call me a lot of things but LIAR is never one of them. By the way isn’t it weird that this psychopath always says things like this when he’s not speaking English?

Ian

19,164 Aufrufe • vor 9 Monaten

New lecture for the book! Nominally about synthetic data, but mostly is a walk through of the distillation literature from the Hinton 2015 paper to multi-teach on-policy distillation of today! At 7.4 hours of video in my post-training brain dump and counting :) It was fun to stare at the math long enough and talk through the 3-4 core changes that needed to be made to the original formulation to have on-policy distillation be ready for the mainstream like it is today (and in RL frameworks). Otherwise, I include a bit of a history lesson for how synthetic data generally slowly took over all post-training data research (it wasn't always the case)! Then I do some 101 review on constitutional AI, rubrics, and other popular methods. 00:00 The emergence of synthetic data 10:50 Background on teacher-student knowledge-distillation 24:47: On-policy distillation (OPD, MOPD, and OPSD) 37:11 Constitutional AI & AI Feedback 45:50 Rubrics as rewards & conclusions Ofc, watch on YouTube etc.

New lecture for the book! Nominally about synthetic data, but mostly is a walk through of the distillation literature from the Hinton 2015 paper to multi-teach on-policy distillation of today! At 7.4 hours of video in my post-training brain dump and counting :) It was fun to stare at the math long enough and talk through the 3-4 core changes that needed to be made to the original formulation to have on-policy distillation be ready for the mainstream like it is today (and in RL frameworks). Otherwise, I include a bit of a history lesson for how synthetic data generally slowly took over all post-training data research (it wasn't always the case)! Then I do some 101 review on constitutional AI, rubrics, and other popular methods. 00:00 The emergence of synthetic data 10:50 Background on teacher-student knowledge-distillation 24:47: On-policy distillation (OPD, MOPD, and OPSD) 37:11 Constitutional AI & AI Feedback 45:50 Rubrics as rewards & conclusions Ofc, watch on YouTube etc.

Nathan Lambert

93,111 Aufrufe • vor 3 Monaten

Explainability is a core principle of OpenLedger. It answers two key questions in AI: 1. Who should get rewarded? 2. Why is it failing-bias, drift, or poisoning? If AI can explain why it made a decision & what data influenced it, the entire system becomes more fair, traceable, and resilient.

Explainability is a core principle of OpenLedger. It answers two key questions in AI: 1. Who should get rewarded? 2. Why is it failing-bias, drift, or poisoning? If AI can explain why it made a decision & what data influenced it, the entire system becomes more fair, traceable, and resilient.

OpenLedger

17,643 Aufrufe • vor 1 Jahr

ELON: CHINA’S TALENT AND INFRASTRUCTURE OUTPACE THE U.S. “The sheer number of smart, hardworking people in China is incredible – far more than here, in my opinion. They’ve got a lot of energy, and their architecture is far more impressive than the US. The train stations, the buildings, the high-speed rail – it’s all far more impressive than what we have.” Source: Elon Musk, Lex Fridman Podcast - Nov 9, 2023

ELON: CHINA’S TALENT AND INFRASTRUCTURE OUTPACE THE U.S. “The sheer number of smart, hardworking people in China is incredible – far more than here, in my opinion. They’ve got a lot of energy, and their architecture is far more impressive than the US. The train stations, the buildings, the high-speed rail – it’s all far more impressive than what we have.” Source: Elon Musk, Lex Fridman Podcast - Nov 9, 2023

Mario Nawfal

149,903 Aufrufe • vor 1 Jahr

🚨 SAM ALTMAN: “AI is going to do more of the total thinking than people are … It likely will be quite disruptive to a lot of the current jobs if more of the intellectual horsepower in the world is GPU’s.”

🚨 SAM ALTMAN: “AI is going to do more of the total thinking than people are … It likely will be quite disruptive to a lot of the current jobs if more of the intellectual horsepower in the world is GPU’s.”

Chief Nerd

49,929 Aufrufe • vor 5 Monaten

Was really interesting to hear John, Beren, and Charlie speculate about why Sonnet 5 and Opus 5 feel like worse models than GLM 5.3 (despite the fact that Anthropic can do raw logit distillation from Fable, and can also train Sonnet/Opus on the environments from which Fable was trained). Led to some interesting thoughts about value of distillation, what it takes to do distillation effectively, and what kinds of model behaviors are hard to extract from distillation.

Was really interesting to hear John, Beren, and Charlie speculate about why Sonnet 5 and Opus 5 feel like worse models than GLM 5.3 (despite the fact that Anthropic can do raw logit distillation from Fable, and can also train Sonnet/Opus on the environments from which Fable was trained). Led to some interesting thoughts about value of distillation, what it takes to do distillation effectively, and what kinds of model behaviors are hard to extract from distillation.

Dwarkesh Patel

143,531 Aufrufe • vor 15 Tagen