Loading video...

Video Failed to Load

Go Home

Stuck waiting on training? Hit fast-forward. Amazon EC2 Trn3 UltraServers enable up to 4.4x higher performance, 3.9x higher memory bandwidth and 4x better performance/watt, meaning lower training costs. Built for agentic AI, MoE, and video models, with native PyTorch integration so you don’t change a single line of code.

awscloud's profile picture

Amazon Web Services

2,217,437 subscribers

1,281,551 views • 8 months ago •via X (Twitter)

EducationHealth & WellnessScience & Technology
LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

$HASHAI is Ahead of the Curve ✅ Today’s Amazon Web Services outage highlights the advantage of truly decentralised, reliable infrastructure. 🔹Hash AI’s Node network is purpose-built for real-world, high-performance computing, built on Ethereum. From training and deploying AI models to executing complex data processing and ETL workflows. 🔹Experience lower costs, higher reliability, and complete freedom from vendor lock-in with infrastructure engineered for scale, speed, and resilience. Rent now → Learn more →
0:41

$HASHAI is Ahead of the Curve ✅ Today’s Amazon Web Services outage highlights the advantage of truly decentralised, reliable infrastructure. 🔹Hash AI’s Node network is purpose-built for real-world, high-performance computing, built on Ethereum. From training and deploying AI models to executing complex data processing and ETL workflows. 🔹Experience lower costs, higher reliability, and complete freedom from vendor lock-in with infrastructure engineered for scale, speed, and resilience. Rent now → Learn more →

Hash AI

17,248 views • 10 months ago

Amazon Redshift RG, powered by AWS Graviton, deliver better performance - running data warehouse and data lake workloads up to 2.4x as fast as previous generation RA3 at 30% lower cost per vCPU. Run SQL analytics across your data warehouse and data lake from a single engine — including on open formats such as Apache Iceberg and Apache Parquet. Built to meet the demands of today’s analytics and agentic AI workloads.
0:28

Amazon Redshift RG, powered by AWS Graviton, deliver better performance - running data warehouse and data lake workloads up to 2.4x as fast as previous generation RA3 at 30% lower cost per vCPU. Run SQL analytics across your data warehouse and data lake from a single engine — including on open formats such as Apache Iceberg and Apache Parquet. Built to meet the demands of today’s analytics and agentic AI workloads.

Amazon Web Services

13,468 views • 3 months ago

Meet Amazon Nova 2 reasoning models Nova has introduced its next generation of foundation models that deliver frontier intelligence with industry-leading price performance 1/ Amazon Nova 2 Lite, is a fast, cost-effective model designed for everyday AI tasks. 2/ Amazon Nova 2 Pro (preview), Nova’s most intelligent model built for complex AI workloads. These models offer developers control to manage thinking budgets, native support for agentic workflows, and provide a one-million-token context window for richer interactions.
0:56

Meet Amazon Nova 2 reasoning models Nova has introduced its next generation of foundation models that deliver frontier intelligence with industry-leading price performance 1/ Amazon Nova 2 Lite, is a fast, cost-effective model designed for everyday AI tasks. 2/ Amazon Nova 2 Pro (preview), Nova’s most intelligent model built for complex AI workloads. These models offer developers control to manage thinking budgets, native support for agentic workflows, and provide a one-million-token context window for richer interactions.

Amazon Web Services

3,958,154 views • 8 months ago

Today, we’re launching Gemma 4, our most intelligent open models to date. Built with the same breakthrough technology as Gemini 3, Gemma 4 brings advanced reasoning to your personal hardware and devices. Here’s what Gemma 4 unlocks for developers: — Intelligence-per-parameter: Our 31B (Dense) and 26B (MoE) models deliver state-of-the-art performance for their size, outcompeting models 20x their size on Arena.ai — Commercial flexibility: Released under a permissive Apache 2.0 license for complete developer flexibility and digital sovereignty — Agentic workflows: Native support for function-calling and structured JSON output allows you to build reliable, autonomous agents — Multimodal edge AI: The E2B and E4B models bring native vision, audio, and low latency to mobile and IoT devices — Long-context reasoning: Up to 256K context windows allow you to process entire repositories or large documents in a single prompt Whether you're building global applications in 140+ languages or local-first AI code assistants, Gemma 4 is built to be your foundation. Explore in Google AI Studio or download the weights on Hugging Face, Kaggle, and ollama.
2:51

Today, we’re launching Gemma 4, our most intelligent open models to date. Built with the same breakthrough technology as Gemini 3, Gemma 4 brings advanced reasoning to your personal hardware and devices. Here’s what Gemma 4 unlocks for developers: — Intelligence-per-parameter: Our 31B (Dense) and 26B (MoE) models deliver state-of-the-art performance for their size, outcompeting models 20x their size on Arena.ai — Commercial flexibility: Released under a permissive Apache 2.0 license for complete developer flexibility and digital sovereignty — Agentic workflows: Native support for function-calling and structured JSON output allows you to build reliable, autonomous agents — Multimodal edge AI: The E2B and E4B models bring native vision, audio, and low latency to mobile and IoT devices — Long-context reasoning: Up to 256K context windows allow you to process entire repositories or large documents in a single prompt Whether you're building global applications in 140+ languages or local-first AI code assistants, Gemma 4 is built to be your foundation. Explore in Google AI Studio or download the weights on Hugging Face, Kaggle, and ollama.

Google AI

1,670,175 views • 4 months ago

GPU tradeoff series: A100 is not much more powerful than 4090 🫠 GPU Perf and Price: - 4090: 330 fp16 TFLOPs, $1,749 - A100 (80GB): 312 fp16 TFLOPs, $20,000 > A100 is 11.4X more pricy Training speed for GPT-2(124M) with llm.c: - 4090: 153K tokens/s - A100 (80GB): 195K tokens/s > A100 is only 1.3X faster (both trained using a single card, A100 llm.c training is shown in the video, 4090 video is in the quoted tweet) Conclusion: 4090 has a much better cost vs performance ratio Why: As in the H100 vs. 4090 comparison, the biggest difference between A100 and 4090 is their GPU memory size/bandwidth and cross-GPU communication bandwidth, which does not matter too much if your model can fit into a single 4090. Specs: 4090: - GPU memory size: 24GB - memory bandwidth: 1 TB/s - communication bandwidth: 64 GB/s A100: - GPU memory size: 80GB - memory bandwidth: 2 TB/s - communication bandwidth: 900 GB/s Nvidia killed off NVLink (a high-speed communication link that connects GPUs) on 4090. (Jensen Huang smiling face) If multiple 4090s could be interconnected via NVLink, their performance would be closer to datacenter-grade A100 GPUs, even for training larger models. Additionally, 4090 isn't allowed in datacenters, that's how Nvidia makes 💰💰💰
0:32

GPU tradeoff series: A100 is not much more powerful than 4090 🫠 GPU Perf and Price: - 4090: 330 fp16 TFLOPs, $1,749 - A100 (80GB): 312 fp16 TFLOPs, $20,000 > A100 is 11.4X more pricy Training speed for GPT-2(124M) with llm.c: - 4090: 153K tokens/s - A100 (80GB): 195K tokens/s > A100 is only 1.3X faster (both trained using a single card, A100 llm.c training is shown in the video, 4090 video is in the quoted tweet) Conclusion: 4090 has a much better cost vs performance ratio Why: As in the H100 vs. 4090 comparison, the biggest difference between A100 and 4090 is their GPU memory size/bandwidth and cross-GPU communication bandwidth, which does not matter too much if your model can fit into a single 4090. Specs: 4090: - GPU memory size: 24GB - memory bandwidth: 1 TB/s - communication bandwidth: 64 GB/s A100: - GPU memory size: 80GB - memory bandwidth: 2 TB/s - communication bandwidth: 900 GB/s Nvidia killed off NVLink (a high-speed communication link that connects GPUs) on 4090. (Jensen Huang smiling face) If multiple 4090s could be interconnected via NVLink, their performance would be closer to datacenter-grade A100 GPUs, even for training larger models. Additionally, 4090 isn't allowed in datacenters, that's how Nvidia makes 💰💰💰

Yuchen Jin

234,951 views • 1 year ago

My startup, Positron AI just raised $51.6M Series A to rebuild the infrastructure powering AI inference and bring Superintelligence to everyone. Led by Valor Equity Partners (meek) , Atreides Management (Atreides Management, LP) , and DFJ Growth (DFJ Growth) —the same teams behind SpaceX, Tesla, xAI, and other companies pushing compute to its physical and economic limits. Why inference? Because as AI moves from research into production, the real bottleneck isn’t training—it’s deploying transformer models at scale. GPUs were built for flexibility, not predictable workloads. That mismatch means wasted bandwidth, poor density, and massive power costs. So at Positron, we built silicon specifically optimized for inference: •>90% memory bandwidth utilization (GPUs typically ~30%) •Multi-model hosting per card (higher density, lower power) •Zero code changes — drop in compatability with the HuggingFace Transformers ecosystem. U.S.-made silicon, stable supply chain, geopolitically resilient We’re already shipping and running in production environments today. Here’s what we’ve built, and link to our WSJ coverage and press release in the follow up reply ↓
6:20

My startup, Positron AI just raised $51.6M Series A to rebuild the infrastructure powering AI inference and bring Superintelligence to everyone. Led by Valor Equity Partners (meek) , Atreides Management (Atreides Management, LP) , and DFJ Growth (DFJ Growth) —the same teams behind SpaceX, Tesla, xAI, and other companies pushing compute to its physical and economic limits. Why inference? Because as AI moves from research into production, the real bottleneck isn’t training—it’s deploying transformer models at scale. GPUs were built for flexibility, not predictable workloads. That mismatch means wasted bandwidth, poor density, and massive power costs. So at Positron, we built silicon specifically optimized for inference: •>90% memory bandwidth utilization (GPUs typically ~30%) •Multi-model hosting per card (higher density, lower power) •Zero code changes — drop in compatability with the HuggingFace Transformers ecosystem. U.S.-made silicon, stable supply chain, geopolitically resilient We’re already shipping and running in production environments today. Here’s what we’ve built, and link to our WSJ coverage and press release in the follow up reply ↓

Thomas Sohmers

26,374 views • 1 year ago

Chips are the foundation of every AI experience. That's why understanding the hardware behind the software has never mattered more. When you ask a chatbot a question or generate an image, there's a physical chip somewhere doing the heavy lifting. For most of computing history, that was a CPU—the "brain" of a computer—great for general tasks needed to run software and operating systems. AI is more complex: for workloads called training and inference, AI needs to perform trillions of calculations in parallel. That's where AI accelerators come in. Purpose-built accelerators can deliver significantly better performance and efficiency than general-purpose chips. Amazon Web Services Trainium chips are an example—purpose-built for AI training and inference. New chips for a new era. ⬇️
0:43

Chips are the foundation of every AI experience. That's why understanding the hardware behind the software has never mattered more. When you ask a chatbot a question or generate an image, there's a physical chip somewhere doing the heavy lifting. For most of computing history, that was a CPU—the "brain" of a computer—great for general tasks needed to run software and operating systems. AI is more complex: for workloads called training and inference, AI needs to perform trillions of calculations in parallel. That's where AI accelerators come in. Purpose-built accelerators can deliver significantly better performance and efficiency than general-purpose chips. Amazon Web Services Trainium chips are an example—purpose-built for AI training and inference. New chips for a new era. ⬇️

Amazon

27,151 views • 2 months ago

Today, we're joined by Aakanksha Chowdhery, member of technical staff at Reflection, to explore the fundamental shifts required to build true agentic AI. While the industry has largely focused on post-training techniques to improve reasoning, Aakanksha draws on her experience leading pre-training efforts for Google’s PaLM and early Gemini models to argue that pre-training itself must be rethought to move beyond static benchmarks. We explore the limitations of next-token prediction for multi-step workflows and examine how attention mechanisms, loss objectives, and training data must evolve to support long-form reasoning and planning. Aakanksha shares insights on the difference between context retrieval and actual reasoning, the importance of "trajectory" training data, and why scaling remains essential for discovering emergent agentic capabilities like error recovery and dynamic tool learning. 🗒️ For the full list of resources for this episode, visit the show notes page: 📖 CHAPTERS =============================== 00:00 - Introduction 02:26 - Reflection 04:54 - Limitations of post-training for building agents 07:31 - Rethinking pre-training in agents 10:51 - Scaling 11:27 - Evolving attention mechanisms for agentic capabilities 12:39 - Memory as a tool 14:13 - Loss objectives and training data 15:50 - Fine-tuning loss in agent performance 19:37 - Training data 21:29 - Augmenting dominant training data source 24:11 - Overcoming challenges in training on synthetic data 25:47 - Benchmarks 30:44 - Scaling laws in large models versus small models 33:20 - Long-form versus short-form reasoning 37:57 - Agent’s ability to recover from failure 40:15 - Hallucinations and failure recovery 43:53 - Tool use in agents 46:38 - Coding agents 48:37 - How researchers can contribute to agentic AI
51:54

Today, we're joined by Aakanksha Chowdhery, member of technical staff at Reflection, to explore the fundamental shifts required to build true agentic AI. While the industry has largely focused on post-training techniques to improve reasoning, Aakanksha draws on her experience leading pre-training efforts for Google’s PaLM and early Gemini models to argue that pre-training itself must be rethought to move beyond static benchmarks. We explore the limitations of next-token prediction for multi-step workflows and examine how attention mechanisms, loss objectives, and training data must evolve to support long-form reasoning and planning. Aakanksha shares insights on the difference between context retrieval and actual reasoning, the importance of "trajectory" training data, and why scaling remains essential for discovering emergent agentic capabilities like error recovery and dynamic tool learning. 🗒️ For the full list of resources for this episode, visit the show notes page: 📖 CHAPTERS =============================== 00:00 - Introduction 02:26 - Reflection 04:54 - Limitations of post-training for building agents 07:31 - Rethinking pre-training in agents 10:51 - Scaling 11:27 - Evolving attention mechanisms for agentic capabilities 12:39 - Memory as a tool 14:13 - Loss objectives and training data 15:50 - Fine-tuning loss in agent performance 19:37 - Training data 21:29 - Augmenting dominant training data source 24:11 - Overcoming challenges in training on synthetic data 25:47 - Benchmarks 30:44 - Scaling laws in large models versus small models 33:20 - Long-form versus short-form reasoning 37:57 - Agent’s ability to recover from failure 40:15 - Hallucinations and failure recovery 43:53 - Tool use in agents 46:38 - Coding agents 48:37 - How researchers can contribute to agentic AI

The TWIML AI Podcast

44,888 views • 8 months ago

Sakana AI CEO David Ha hardmaru: "If you assemble a team of Avengers, that team will be better than the individual heroes. The same with AI." Sakana Fugu, their new AI “model” goes toe-to-toe with Mythos and Fable 5 on a few benchmarks — without training a single frontier model of its own. Fugu orchestrates between different frontier AI models to achieve higher results. On the real risk, David adds: "I'm more concerned about the structure of dependency… you don't want to be tied to a single particular AI model or a single particular vendor." An exclusive 1:1 with CROSS DIG’s 中川雅博 / Masahiro Nakagawa @ TBS CROSS DIG #SakanaAI #SakanaFugu
24:51

Sakana AI CEO David Ha hardmaru: "If you assemble a team of Avengers, that team will be better than the individual heroes. The same with AI." Sakana Fugu, their new AI “model” goes toe-to-toe with Mythos and Fable 5 on a few benchmarks — without training a single frontier model of its own. Fugu orchestrates between different frontier AI models to achieve higher results. On the real risk, David adds: "I'm more concerned about the structure of dependency… you don't want to be tied to a single particular AI model or a single particular vendor." An exclusive 1:1 with CROSS DIG’s 中川雅博 / Masahiro Nakagawa @ TBS CROSS DIG #SakanaAI #SakanaFugu

TBS CROSS DIG with Bloomberg

78,986 views • 1 month ago

💥AI is no longer confined to the cloud. At Computex, Jack Huynh showcased the AMD Ryzen AI Halo and Ryzen AI Max PRO 400 Series processors. The next wave of innovation will be built, trained, and run locally. As AI evolves from chatbots to intelligent agents, developers and enterprises need the performance, memory capacity, and flexibility to bring these experiences to life directly on the device. ⚡ Ryzen AI Halo enables local AI development with support for models up to 200B parameters, up to 128GB of unified memory, ROCm optimization, and Windows + Linux support. ⚡ Ryzen AI Max PRO 400 Series combines AI, graphics, and compute in a single architecture, delivering workstation-class performance for next-gen Agent Computers and AI PCs. Some of the most impactful AI experiences will happen right at your desk, where latency, privacy, and control matter most. We're just getting started.
1:24

💥AI is no longer confined to the cloud. At Computex, Jack Huynh showcased the AMD Ryzen AI Halo and Ryzen AI Max PRO 400 Series processors. The next wave of innovation will be built, trained, and run locally. As AI evolves from chatbots to intelligent agents, developers and enterprises need the performance, memory capacity, and flexibility to bring these experiences to life directly on the device. ⚡ Ryzen AI Halo enables local AI development with support for models up to 200B parameters, up to 128GB of unified memory, ROCm optimization, and Windows + Linux support. ⚡ Ryzen AI Max PRO 400 Series combines AI, graphics, and compute in a single architecture, delivering workstation-class performance for next-gen Agent Computers and AI PCs. Some of the most impactful AI experiences will happen right at your desk, where latency, privacy, and control matter most. We're just getting started.

AMD

24,315 views • 2 months ago

Jupiter Lend rewrote the rules. With our tick-based liquidity system, we’re able to clear risky positions so efficiently that we can empower borrowers with better terms. Higher LTVs, higher LTs, and liquidations that only restore you back to health, and no further. The system is so efficient that it has incurred 0 bad debt in its entire history. Fast forward to 6 months after release, and Lend 1.5 is pushing the boundaries even further. - Native Stake as collateral to unlock the billions of dollars of Solana natively staked - Custodied collateral to bring more assets onchain - The $JUICED Loop: deposit $JupUSD, receive $JUICED, and loop against other stables for amplified yield Better for borrowers. Simpler for lenders. Built for everyone. Just use Jupiter (Lend).
0:40

Jupiter Lend rewrote the rules. With our tick-based liquidity system, we’re able to clear risky positions so efficiently that we can empower borrowers with better terms. Higher LTVs, higher LTs, and liquidations that only restore you back to health, and no further. The system is so efficient that it has incurred 0 bad debt in its entire history. Fast forward to 6 months after release, and Lend 1.5 is pushing the boundaries even further. - Native Stake as collateral to unlock the billions of dollars of Solana natively staked - Custodied collateral to bring more assets onchain - The $JUICED Loop: deposit $JupUSD, receive $JUICED, and loop against other stables for amplified yield Better for borrowers. Simpler for lenders. Built for everyone. Just use Jupiter (Lend).

Jupiter

49,539 views • 6 months ago

🔥Really excited to see the release of PAN world model, a project I had been working over the past years. PAN is a general world model capable of simulating physical, agentic, and nested worlds, synthesizing infinite interactive experiences for training AI agents. Building on top of pretrained LLMs and video diffusion models, PAN connects language, perception, action, and latent thoughts, for long-horizon simulation and reasoning. PAN shows overwhelming performance gains over JEPA-2, Cosmos-2, and other prior models. More in the thread👇 ... 1/
0:54

🔥Really excited to see the release of PAN world model, a project I had been working over the past years. PAN is a general world model capable of simulating physical, agentic, and nested worlds, synthesizing infinite interactive experiences for training AI agents. Building on top of pretrained LLMs and video diffusion models, PAN connects language, perception, action, and latent thoughts, for long-horizon simulation and reasoning. PAN shows overwhelming performance gains over JEPA-2, Cosmos-2, and other prior models. More in the thread👇 ... 1/

Zhiting Hu

31,213 views • 9 months ago

✨ #GOOGL Key Milestone ✨ shared on Power Lunch Market Navigator Back in July, Alphabet hit a $191 high, but bearish signals kept it in check. Now, those signals (like key moving averages) have flipped to support, paving the way for a potential bullish breakout! 🚀 A weekly close above $191 could push the stock toward $220, based on Fibonacci levels. 📈 Keep an eye on this trend! 👀 ✨A key theme for 2025 is AI Agents✨ Smarter AI Agents Alphabet’s Trillium TPU is a cloud differentiator. TPUs are used to train large language models, and fast and more energy-efficient training capabilities position Alphabet to compete with AWS, and Azure. The trillium TPU increases performance by 4x and is 67% more energy efficient. Salesforce states that increase training speed – which means this will be key for AI agents. ✅It requires increased training speed, preferably an energy-efficient one. ✅Key cloud differentiator with Alphabet’s Trillium TPU ✅Performance is increased by 4 times in comparison to its predecessor and is 67% more energy efficient ✅On Dec 11 Google announced Gemini 2.0 which was trained on these TPUs ✅Combine this training with cloud customers and Google is now positioned to compete with AWS and Azure.
2:20

✨ #GOOGL Key Milestone ✨ shared on Power Lunch Market Navigator Back in July, Alphabet hit a $191 high, but bearish signals kept it in check. Now, those signals (like key moving averages) have flipped to support, paving the way for a potential bullish breakout! 🚀 A weekly close above $191 could push the stock toward $220, based on Fibonacci levels. 📈 Keep an eye on this trend! 👀 ✨A key theme for 2025 is AI Agents✨ Smarter AI Agents Alphabet’s Trillium TPU is a cloud differentiator. TPUs are used to train large language models, and fast and more energy-efficient training capabilities position Alphabet to compete with AWS, and Azure. The trillium TPU increases performance by 4x and is 67% more energy efficient. Salesforce states that increase training speed – which means this will be key for AI agents. ✅It requires increased training speed, preferably an energy-efficient one. ✅Key cloud differentiator with Alphabet’s Trillium TPU ✅Performance is increased by 4 times in comparison to its predecessor and is 67% more energy efficient ✅On Dec 11 Google announced Gemini 2.0 which was trained on these TPUs ✅Combine this training with cloud customers and Google is now positioned to compete with AWS and Azure.

Jessica Inskip

28,234 views • 1 year ago

Labour's reckless anti-business jobs tax has arrived, and it's doing untold damage to businesses up and down the country. The Chancellor wants you to believe that you will be better off under Labour. The reality is, her choices mean: 👉🏻 fewer jobs 👉🏻 lower wages 👉🏻 higher prices. From today, businesses will be hit with higher National Insurance contributions, meaning an additional £900 for the average worker. As someone who has built businesses from scratch, I know that this is a burden that is simply too much for many employers to absorb. We can't afford Labour 👇🏻
1:12

Labour's reckless anti-business jobs tax has arrived, and it's doing untold damage to businesses up and down the country. The Chancellor wants you to believe that you will be better off under Labour. The reality is, her choices mean: 👉🏻 fewer jobs 👉🏻 lower wages 👉🏻 higher prices. From today, businesses will be hit with higher National Insurance contributions, meaning an additional £900 for the average worker. As someone who has built businesses from scratch, I know that this is a burden that is simply too much for many employers to absorb. We can't afford Labour 👇🏻

Mel Stride

41,966 views • 1 year ago

BOOM! Because of X Creator monetization and YOU I have the funds to get to 2200 simultaneous streams of live 24/7 radio recordings from around the world for AI training. You know, you get higher quality human interactions on radio than on Reddit. I just can’t understand why? Absolutely no AI company is doing what I am doing. This is just one way I never run out of AI training data. I have many, many more. And guess what? I don’t need “ai alignment for ‘safety’” when I train in high protein data. Come take a trip around the world with me, and hear 1000s of different perspectives…
6:21

BOOM! Because of X Creator monetization and YOU I have the funds to get to 2200 simultaneous streams of live 24/7 radio recordings from around the world for AI training. You know, you get higher quality human interactions on radio than on Reddit. I just can’t understand why? Absolutely no AI company is doing what I am doing. This is just one way I never run out of AI training data. I have many, many more. And guess what? I don’t need “ai alignment for ‘safety’” when I train in high protein data. Come take a trip around the world with me, and hear 1000s of different perspectives…

Brian Roemmele

30,843 views • 7 months ago

$MU $SKHY $DRAM Holy Shit $AMD's Lisa Su just spent two full minutes at the 2026 AI keynote raving about HBM4 capacity as the thing that separates Helios from the competition. Not compute. Memory. "When you compare Helios to the competition, we're delivering 15% more compute, 50% more HBM4 memory capacity and memory bandwidth, and 50% more scale-out bandwidth." Notice which number is bigger. "What that means is that every Helios can deliver more performance for the largest models, more capacity for longer context, and the bandwidth to scale across thousands of racks.... More than 18,000 CDNA five GPU compute units, over 4,600 Zen 6 CPU cores, and 31 terabytes of HBM4 memory, all in a single rack." "That's what it takes to run agentic AI at scale. Today, I'm excited to announce that Helios is in full production. We have shipments on track to start at the end of the third quarter and ramping into the fourth quarter in the second half of the year. I can tell you, customer demand for Helios is extremely strong, we're extremely proud of the work that we have done across the leading AI labs to adopt Helios."
1:53

$MU $SKHY $DRAM Holy Shit $AMD's Lisa Su just spent two full minutes at the 2026 AI keynote raving about HBM4 capacity as the thing that separates Helios from the competition. Not compute. Memory. "When you compare Helios to the competition, we're delivering 15% more compute, 50% more HBM4 memory capacity and memory bandwidth, and 50% more scale-out bandwidth." Notice which number is bigger. "What that means is that every Helios can deliver more performance for the largest models, more capacity for longer context, and the bandwidth to scale across thousands of racks.... More than 18,000 CDNA five GPU compute units, over 4,600 Zen 6 CPU cores, and 31 terabytes of HBM4 memory, all in a single rack." "That's what it takes to run agentic AI at scale. Today, I'm excited to announce that Helios is in full production. We have shipments on track to start at the end of the third quarter and ramping into the fourth quarter in the second half of the year. I can tell you, customer demand for Helios is extremely strong, we're extremely proud of the work that we have done across the leading AI labs to adopt Helios."

Trade Whisperer

179,812 views • 24 days ago

"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"
2:21

"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"

Dwarkesh Patel

502,249 views • 8 months ago

I have built a Windows x64 sampling profiler for native code+C# to be used by coding agents. Why? Because it is irresponsible to let computers write code but give them no tools to evaluate the performance of their code. Stop guessing, start measuring. -> How does it work? Just point your agent to the profiler and let them measure and optimize. Or record a trace yourself and let the agent interpret the results for you. Works with Claude Code, Codex, pi, etc. The profiler also has a couple of neat features by itself like giving you estimates for how much time your code is spending on memory stalls and annotating disassembly with inlining information. Take a look and try it for free!
12:40

I have built a Windows x64 sampling profiler for native code+C# to be used by coding agents. Why? Because it is irresponsible to let computers write code but give them no tools to evaluate the performance of their code. Stop guessing, start measuring. -> How does it work? Just point your agent to the profiler and let them measure and optimize. Or record a trace yourself and let the agent interpret the results for you. Works with Claude Code, Codex, pi, etc. The profiler also has a couple of neat features by itself like giving you estimates for how much time your code is spending on memory stalls and annotating disassembly with inlining information. Take a look and try it for free!

Sebastian Schöner

13,057 views • 3 months ago

📢 BREAKING: Nvidia, Microsoft, and Anthropic announce new strategic AI partnerships of roughly $45 Billion. They struck a 3-way deal where Anthropic commits $30B for Azure compute with capacity up to 1GW, and NVIDIA and Microsoft plan to invest up to $10B and $5B in Anthropic. Claude will scale on Azure using NVIDIA systems, which expands access for enterprises and makes Claude available across the major clouds. Anthropic and NVIDIA will co-design to tune both model stacks and future chips for performance, efficiency, and total cost of ownership, so the models run faster per dollar. The first wave targets NVIDIA Grace Blackwell and Vera Rubin systems, which pair high-bandwidth memory with dense interconnect to push training and inference throughput. Microsoft will expose Anthropic models in Azure AI Foundry, including Claude Sonnet 4.5, Claude Opus 4.1, and Claude Haiku 4.5, so teams can pick latency and cost profiles. Claude will also remain available across the Copilot family, which means coding, document automation, and workflow agents can call the same model line. The 1GW commitment signals many thousands of GPUs worth of capacity, which reduces queue times and stabilizes unit economics for long-running training runs. The joint optimization loop, models shaping chips and chips shaping models, typically yields double-digit efficiency gains over off-the-shelf deployments.
1:10

📢 BREAKING: Nvidia, Microsoft, and Anthropic announce new strategic AI partnerships of roughly $45 Billion. They struck a 3-way deal where Anthropic commits $30B for Azure compute with capacity up to 1GW, and NVIDIA and Microsoft plan to invest up to $10B and $5B in Anthropic. Claude will scale on Azure using NVIDIA systems, which expands access for enterprises and makes Claude available across the major clouds. Anthropic and NVIDIA will co-design to tune both model stacks and future chips for performance, efficiency, and total cost of ownership, so the models run faster per dollar. The first wave targets NVIDIA Grace Blackwell and Vera Rubin systems, which pair high-bandwidth memory with dense interconnect to push training and inference throughput. Microsoft will expose Anthropic models in Azure AI Foundry, including Claude Sonnet 4.5, Claude Opus 4.1, and Claude Haiku 4.5, so teams can pick latency and cost profiles. Claude will also remain available across the Copilot family, which means coding, document automation, and workflow agents can call the same model line. The 1GW commitment signals many thousands of GPUs worth of capacity, which reduces queue times and stabilizes unit economics for long-running training runs. The joint optimization loop, models shaping chips and chips shaping models, typically yields double-digit efficiency gains over off-the-shelf deployments.

Rohan Paul

39,781 views • 9 months ago

Here's my conversation with Lucas Atkins and the team at Arcee.ai on their path to training and releasing Trinity Large today. From going all in on open models built end to end in the US 6 months ago to having the model in hand is no easy feet. I loved this conversation on how to design a startup around open models and take a bold step to scale it up. I'm openly an Arcee fan, watching them take risk and pull it off. We discuss: - The state (and future) of open vs. closed models, - The business of selling open models for on-prem deployments, - The story of Arcee AI & going “all-in” on this training run, - The ATOM project, - Building frontier model training teams in 6 months, - and other great topics. I really loved this one, and think you well too. Chapters: 00:00:00 Intro: Arcee AI, Trinity Models & Trinity Large 00:08:26 Transitioning a Company to Pre-training 00:13:00 Technical Decisions: Muon and MoE 00:18:41 Scaling and MoE Training Pain 00:23:14 Post-training and RL Strategies 00:28:09 Team Structure and Data Scaling 00:31:31 The Trinity Manifesto: US Open Weights 00:42:31 Specialized Models and Distillation 00:47:12 Infrastructure and Hosting 400B 00:50:53 Open Source as a Business Moat 00:56:31 Predictions: Best Model in 2026 01:02:29 Lightning Round & Conclusions More great open model builder podcasts coming soon!
1:12:15

Here's my conversation with Lucas Atkins and the team at Arcee.ai on their path to training and releasing Trinity Large today. From going all in on open models built end to end in the US 6 months ago to having the model in hand is no easy feet. I loved this conversation on how to design a startup around open models and take a bold step to scale it up. I'm openly an Arcee fan, watching them take risk and pull it off. We discuss: - The state (and future) of open vs. closed models, - The business of selling open models for on-prem deployments, - The story of Arcee AI & going “all-in” on this training run, - The ATOM project, - Building frontier model training teams in 6 months, - and other great topics. I really loved this one, and think you well too. Chapters: 00:00:00 Intro: Arcee AI, Trinity Models & Trinity Large 00:08:26 Transitioning a Company to Pre-training 00:13:00 Technical Decisions: Muon and MoE 00:18:41 Scaling and MoE Training Pain 00:23:14 Post-training and RL Strategies 00:28:09 Team Structure and Data Scaling 00:31:31 The Trinity Manifesto: US Open Weights 00:42:31 Specialized Models and Distillation 00:47:12 Infrastructure and Hosting 400B 00:50:53 Open Source as a Business Moat 00:56:31 Predictions: Best Model in 2026 01:02:29 Lightning Round & Conclusions More great open model builder podcasts coming soon!

Nathan Lambert

26,872 views • 6 months ago