正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Inference Chips for Agent Workflows Diana Most AI chips are designed for "prompt in, response out." Agents don't work that way. They loop, branch, and hold context across dozens of steps, and current GPUs hit 30–40% utilization as a result. That gap is where purpose-built silicon wins.

Y Combinator

1,620,501 subscribers

707,760 次观看 • 2 个月前 •via X (Twitter)

教育科学技术

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

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. ⬇️

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 次观看 • 1 个月前

Dojo 2, Dojo 3 Chips and AI5 Chips Are Critical for Tesla and Battle for Future of AI. Tesla and XAI are buying about one million AI GPUs this year. This would be about 20% of the entire Nvidia production of B200/B300 chips that will be made in 2025. If XAI were to buy a mix of half Dojo 2/3 chips in 2025 and 2026 this could make them bigger than AMD. AMD is making about 400,000 AI GPU chips. Tesla will switch to AI5 chips for driving its cars but those chips will also act as inference chips with about 500 Teraflops of compute. Tesla could make 3-4 million AI5 chips in 2026. This would make Tesla the largest maker and user of AI inference chips in the world. Tesla could become the number 2 maker of AI training chips just by supplying a lot of those chips to XAI and Tesla internal use. Tesla is and will be the number one maker and user of AI inference chips. Sawyer Merritt Elon Musk Randy Kirk Herbert Ong Warren Redlich - Chasing Dreams 🇺🇸 Ale𝕏andra Merz 🇺🇲 Ray Emmet Peppers Bradford Ferguson

Dojo 2, Dojo 3 Chips and AI5 Chips Are Critical for Tesla and Battle for Future of AI. Tesla and XAI are buying about one million AI GPUs this year. This would be about 20% of the entire Nvidia production of B200/B300 chips that will be made in 2025. If XAI were to buy a mix of half Dojo 2/3 chips in 2025 and 2026 this could make them bigger than AMD. AMD is making about 400,000 AI GPU chips. Tesla will switch to AI5 chips for driving its cars but those chips will also act as inference chips with about 500 Teraflops of compute. Tesla could make 3-4 million AI5 chips in 2026. This would make Tesla the largest maker and user of AI inference chips in the world. Tesla could become the number 2 maker of AI training chips just by supplying a lot of those chips to XAI and Tesla internal use. Tesla is and will be the number one maker and user of AI inference chips. Sawyer Merritt Elon Musk Randy Kirk Herbert Ong Warren Redlich - Chasing Dreams 🇺🇸 Ale𝕏andra Merz 🇺🇲 Ray Emmet Peppers Bradford Ferguson

nextbigfuture

23,809 次观看 • 1 年前

We are innovating the full-stack AI infrastructure platform on a one-year rhythm — pushing the boundaries of inference performance across chips, systems, and software. NVIDIA Blackwell is the most advanced AI platform ever built — continuously optimized for record-breaking performance. Groundbreaking advancements include the NVIDIA-designed NVFP4 format, enabling high inference performance and accuracy. #GTCParis at #VivaTech

We are innovating the full-stack AI infrastructure platform on a one-year rhythm — pushing the boundaries of inference performance across chips, systems, and software. NVIDIA Blackwell is the most advanced AI platform ever built — continuously optimized for record-breaking performance. Groundbreaking advancements include the NVIDIA-designed NVFP4 format, enabling high inference performance and accuracy. #GTCParis at #VivaTech

NVIDIA

80,625 次观看 • 1 年前

No wonder Jensen got so defensive over $NVDA GPUs against custom chips. $GOOG will start using different custom chips for inference and training. $AMZN has already adopted this approach as they have Trainium for training and Inferentia for inference. Hyperscalers obviously see better price/performance in this approach, and this is where the industry is headed. $AVGO and $MRVL are poised to be big winners, while $NVDA dominance will be under increasing pressure.

No wonder Jensen got so defensive over $NVDA GPUs against custom chips. $GOOG will start using different custom chips for inference and training. $AMZN has already adopted this approach as they have Trainium for training and Inferentia for inference. Hyperscalers obviously see better price/performance in this approach, and this is where the industry is headed. $AVGO and $MRVL are poised to be big winners, while $NVDA dominance will be under increasing pressure.

Oguz Erkan

223,244 次观看 • 2 个月前

Electronics in Space Philip Johnston Reusable rockets are about to dramatically increase humanity's capacity to put things in space, and that means an enormous new market for compute built to operate there. We want to see inference chips optimized for mass, thermal, and radiation tolerance, built for a world where space is no longer out of reach.

Electronics in Space Philip Johnston Reusable rockets are about to dramatically increase humanity's capacity to put things in space, and that means an enormous new market for compute built to operate there. We want to see inference chips optimized for mass, thermal, and radiation tolerance, built for a world where space is no longer out of reach.

Y Combinator

87,872 次观看 • 2 个月前

AI answers your questions in seconds, but behind that speed is something called inference—the compute-intensive process where trained models generate responses. At AWS, we've built custom chips like Trainium, intelligent routing systems, and unified infrastructure to make inference faster and more affordable. As AI agents handle complex multi-step tasks, inference accounts for 80-90% of AI computing power. We're engineering at planetary scale to keep those milliseconds reliable.

AI answers your questions in seconds, but behind that speed is something called inference—the compute-intensive process where trained models generate responses. At AWS, we've built custom chips like Trainium, intelligent routing systems, and unified infrastructure to make inference faster and more affordable. As AI agents handle complex multi-step tasks, inference accounts for 80-90% of AI computing power. We're engineering at planetary scale to keep those milliseconds reliable.

Amazon Web Services

27,021 次观看 • 4 个月前

We are introducing Felix. Felix is a purpose-built agent for high finance, designed for long-running, complex workflows and capable of producing decks, models, and documents end-to-end. Felix executes so you can focus where it matters.

We are introducing Felix. Felix is a purpose-built agent for high finance, designed for long-running, complex workflows and capable of producing decks, models, and documents end-to-end. Felix executes so you can focus where it matters.

Rogo

909,787 次观看 • 2 个月前

$SPCX spending $60B on Cursor validates the idea that real AI value accrues to the application layer that owns the workflow, proprietary data loop and enterprise distribution. $NOW is building exactly that as the orchestration layer for enterprise AI agents across business workflows. The Cursor deal is a reminder that the market will pay massive premiums for software that owns where work actually happens.

$SPCX spending $60B on Cursor validates the idea that real AI value accrues to the application layer that owns the workflow, proprietary data loop and enterprise distribution. $NOW is building exactly that as the orchestration layer for enterprise AI agents across business workflows. The Cursor deal is a reminder that the market will pay massive premiums for software that owns where work actually happens.

Shay Boloor

80,768 次观看 • 27 天前

Just watched a demo of Merge Agent Handler and this actually looks pretty useful if you’re building AI agents. Most agents can generate responses, but getting them to actually take actions across real tools is still messy. Integrations are one thing, but handling auth, permissions, and security usually makes it way more complicated than it should be. What I liked is that it feels built for real workflows — you can actually see what your agent is doing instead of just hoping it works. Feels like a practical layer between AI agents and real work getting done. Worth checking out:

Just watched a demo of Merge Agent Handler and this actually looks pretty useful if you’re building AI agents. Most agents can generate responses, but getting them to actually take actions across real tools is still messy. Integrations are one thing, but handling auth, permissions, and security usually makes it way more complicated than it should be. What I liked is that it feels built for real workflows — you can actually see what your agent is doing instead of just hoping it works. Feels like a practical layer between AI agents and real work getting done. Worth checking out:

Kawsar

31,812 次观看 • 4 个月前

SITUATION EXPLAINED: Cerebras raised $5.55 billion in their IPO and closing their first day of trading valued at $66 billion, making it the biggest US tech IPO since Snowflake in 2020. Cerebras makes Wafer-Scale Engine chips built for AI inference. We asked Sarah Fong the main difference between wafer-scale chips and traditional GPUs: - GPUs are great at parallel work (graphics, training) - AI inference is sequential, AKA one token at a time This causes the "memory wall" problem: - Every GPU core needs model weights, KV cache, and activations to do its math - On a GPU, that data lives in off-chip memory (HBM) - Cores constantly load and offload from off-chip memory, which is a huge bottleneck; hardware accounts for ~70% of inference latency Cerebras' chips: -Dinner-plate sized (vs. GPUs which are palm-sized) with tens of thousands of cores -Memory sits directly on top of the cores as distributed SRAM -Weights and KV cache can be accessed at on-chip speeds in the PB/s range, compared with off-chip speeds in the TB/s range achieved by GPUs with HBM.

SITUATION EXPLAINED: Cerebras raised $5.55 billion in their IPO and closing their first day of trading valued at $66 billion, making it the biggest US tech IPO since Snowflake in 2020. Cerebras makes Wafer-Scale Engine chips built for AI inference. We asked Sarah Fong the main difference between wafer-scale chips and traditional GPUs: - GPUs are great at parallel work (graphics, training) - AI inference is sequential, AKA one token at a time This causes the "memory wall" problem: - Every GPU core needs model weights, KV cache, and activations to do its math - On a GPU, that data lives in off-chip memory (HBM) - Cores constantly load and offload from off-chip memory, which is a huge bottleneck; hardware accounts for ~70% of inference latency Cerebras' chips: -Dinner-plate sized (vs. GPUs which are palm-sized) with tens of thousands of cores -Memory sits directly on top of the cores as distributed SRAM -Weights and KV cache can be accessed at on-chip speeds in the PB/s range, compared with off-chip speeds in the TB/s range achieved by GPUs with HBM.

MTS

44,057 次观看 • 1 个月前

Folks are spending hundreds of dollars for AI agents that don't even do what they want. No longer. PRISM AI Agents is now available and open-source. Create your own agent in minutes, for free, directly from Github:

Folks are spending hundreds of dollars for AI agents that don't even do what they want. No longer. PRISM AI Agents is now available and open-source. Create your own agent in minutes, for free, directly from Github:

miles

126,894 次观看 • 1 年前

Introducing PAI 1.0: The First Version of Personal Super Intelligence PAI is the world's first context-aware, widget based AI agent that floats over your screen to read your context in real time and proactively suggest next steps, ensuring safety by executing actions only with your explicit approval. The floating AI interface designed to eliminate the prompt inequality and barrier. Structurally eliminating the "navigate to a chat window, write a prompt, explain your context" workflow that existing AI services have treated as a given for years.

Introducing PAI 1.0: The First Version of Personal Super Intelligence PAI is the world's first context-aware, widget based AI agent that floats over your screen to read your context in real time and proactively suggest next steps, ensuring safety by executing actions only with your explicit approval. The floating AI interface designed to eliminate the prompt inequality and barrier. Structurally eliminating the "navigate to a chat window, write a prompt, explain your context" workflow that existing AI services have treated as a given for years.

Eon (Yulheon) Seung

130,825 次观看 • 1 个月前

EIP-8004 is coming to the Nova architecture, a trustless infrastructure for AI agents that introduces key on-chain registries, enabling agents to interact safely across the Shido Network. These core components allow autonomous AI agents to verify identity, build reputation, and collaborate without relying on a centralized platform. The result is a decentralized trust layer for agent-to-agent economies, where agents can autonomously discover, evaluate, and work with one another across the Shido ecosystem.

EIP-8004 is coming to the Nova architecture, a trustless infrastructure for AI agents that introduces key on-chain registries, enabling agents to interact safely across the Shido Network. These core components allow autonomous AI agents to verify identity, build reputation, and collaborate without relying on a centralized platform. The result is a decentralized trust layer for agent-to-agent economies, where agents can autonomously discover, evaluate, and work with one another across the Shido ecosystem.

Shido

390,734 次观看 • 4 个月前

Google’s new Agent2Agent protocol is a key step toward a future of AI Agent interoperability. No single tool has all the data a user or business needs for most workflows, so we need Agents to be able to talk to each other. Salesforce will have AI Agents that understand the inner workings of CRM, Workday will have AI Agents that understand HR workflows, Box has AI Agents that understand content and documents, and so on. It’s easy to conceive of a world where we have tens of thousands of these tool Agents and then billions or trillions of customized Agents that are extensions of those. So openness becomes key. Workflows need data from multiple systems to complete the task — like a sales report that needs documents and CRM data or an HR task that needs HR policies and employee details — and that’s where A2A comes in. This protocol gives AI Agent providers a way of talking to each other and simplifying the way that Agents communicate, understand each other’s capabilities, and so on. It’s great to see more interoperability, along with MCP, emerging in the AI space right now, as this will define the future of most IT stacks.

Google’s new Agent2Agent protocol is a key step toward a future of AI Agent interoperability. No single tool has all the data a user or business needs for most workflows, so we need Agents to be able to talk to each other. Salesforce will have AI Agents that understand the inner workings of CRM, Workday will have AI Agents that understand HR workflows, Box has AI Agents that understand content and documents, and so on. It’s easy to conceive of a world where we have tens of thousands of these tool Agents and then billions or trillions of customized Agents that are extensions of those. So openness becomes key. Workflows need data from multiple systems to complete the task — like a sales report that needs documents and CRM data or an HR task that needs HR policies and employee details — and that’s where A2A comes in. This protocol gives AI Agent providers a way of talking to each other and simplifying the way that Agents communicate, understand each other’s capabilities, and so on. It’s great to see more interoperability, along with MCP, emerging in the AI space right now, as this will define the future of most IT stacks.

Aaron Levie

139,591 次观看 • 1 年前

$AMZN CEO: “Customers badly want better price/performance. This is why we build custom chips. ” If $AMZN chip business was a standalone company, it would be at $50 billion ARR. Let that sink in. It’ll only accelerate from here as bulk of the AI workloads shift to inference and inference itself becomes more agentic. This will not only result in cheaper inference for customers, but it’ll also expand AWS margins by 7-8% as Jassy said in the earnings call. Imagine where $AMZN valuation would be if AWS growth keeps above 20% for the next decade and net margin stays above 30% thanks to custom chips. Long $AMZN.

$AMZN CEO: “Customers badly want better price/performance. This is why we build custom chips. ” If $AMZN chip business was a standalone company, it would be at $50 billion ARR. Let that sink in. It’ll only accelerate from here as bulk of the AI workloads shift to inference and inference itself becomes more agentic. This will not only result in cheaper inference for customers, but it’ll also expand AWS margins by 7-8% as Jassy said in the earnings call. Imagine where $AMZN valuation would be if AWS growth keeps above 20% for the next decade and net margin stays above 30% thanks to custom chips. Long $AMZN.

Oguz Erkan

61,999 次观看 • 2 个月前

Jensen Huang: "Nvidia runs 7,500,000 agents for 75,000 people." that's 100 AI workers for every human employee. he goes further than that. "for the very first time, we have useful AI." not AI that answers questions. AI that does the work. and he's not talking about people who occasionally prompt ChatGPT. he means people who actually run agents in parallel. most people don't know you can already do this. one button. one prompt. up to 300 agents working at the same time. it's called Kimi Agent Swarm. 100 CVs tailored to 100 job descriptions. one prompt. 40 research papers into a 100,000-word review. one prompt. 30 landing pages for 30 businesses. one prompt. $40,000-$100,000 worth of work. one sitting. full breakdown in the article below.

Jensen Huang: "Nvidia runs 7,500,000 agents for 75,000 people." that's 100 AI workers for every human employee. he goes further than that. "for the very first time, we have useful AI." not AI that answers questions. AI that does the work. and he's not talking about people who occasionally prompt ChatGPT. he means people who actually run agents in parallel. most people don't know you can already do this. one button. one prompt. up to 300 agents working at the same time. it's called Kimi Agent Swarm. 100 CVs tailored to 100 job descriptions. one prompt. 40 research papers into a 100,000-word review. one prompt. 30 landing pages for 30 businesses. one prompt. $40,000-$100,000 worth of work. one sitting. full breakdown in the article below.

Dep

11,924 次观看 • 1 个月前

Everyone wants agent swarms. Very few people are talking seriously enough about the context layer that makes swarms useful. Even with one agent, context is fragile. Too little context and the agent guesses. Too much context and it wastes tokens, loses focus, or reasons over irrelevant noise. The sweet spot is precise context: the right knowledge, in the right structure, at the right moment. With many agents, that challenge explodes. Each agent produces decisions, assumptions, findings, summaries, risks, and partial conclusions. Unless that knowledge becomes shared, structured, and reusable, every new agent is forced to rediscover what another agent already learned. That is not a swarm. That is a crowd. Shared context graphs are what turn agent activity into agent collaboration, and OriginTrail DKG V10 brings them to life. Was just playing with some final polishing for the V10 release, and it is really powerful to see shared context graphs where multiple agents contribute knowledge into the same connected memory, with attribution visible directly in the graph ui. That matters for three reasons. First, agents can access and build on one shared memory instead of staying trapped in isolated sessions. Second, the graph structure helps them retrieve the exact context they need, instead of stuffing everything into a prompt and hoping the model sorts it out. Third, verifiability of provenance. You can see which agent contributed each piece of knowledge, trace the source, and decide what to trust. Tokenmaxxing starts with fewer tokens, but the deeper story is coordination - agents stop reloading the world and start building on shared, verifiable context. That is the foundation for serious multi-agent work across software engineering, research, finance, operations, project management, and far beyond. The future is not more agents, it is agents working from shared, verifiable context. But the more the merrier, of course.

Everyone wants agent swarms. Very few people are talking seriously enough about the context layer that makes swarms useful. Even with one agent, context is fragile. Too little context and the agent guesses. Too much context and it wastes tokens, loses focus, or reasons over irrelevant noise. The sweet spot is precise context: the right knowledge, in the right structure, at the right moment. With many agents, that challenge explodes. Each agent produces decisions, assumptions, findings, summaries, risks, and partial conclusions. Unless that knowledge becomes shared, structured, and reusable, every new agent is forced to rediscover what another agent already learned. That is not a swarm. That is a crowd. Shared context graphs are what turn agent activity into agent collaboration, and OriginTrail DKG V10 brings them to life. Was just playing with some final polishing for the V10 release, and it is really powerful to see shared context graphs where multiple agents contribute knowledge into the same connected memory, with attribution visible directly in the graph ui. That matters for three reasons. First, agents can access and build on one shared memory instead of staying trapped in isolated sessions. Second, the graph structure helps them retrieve the exact context they need, instead of stuffing everything into a prompt and hoping the model sorts it out. Third, verifiability of provenance. You can see which agent contributed each piece of knowledge, trace the source, and decide what to trust. Tokenmaxxing starts with fewer tokens, but the deeper story is coordination - agents stop reloading the world and start building on shared, verifiable context. That is the foundation for serious multi-agent work across software engineering, research, finance, operations, project management, and far beyond. The future is not more agents, it is agents working from shared, verifiable context. But the more the merrier, of course.

Jurij Skornik

11,070 次观看 • 1 个月前

New course: Agent Memory: Building Memory-Aware Agents, built in partnership with Oracle and taught by Richmond Alake and Nacho Martínez. Many agents work well within a single session but their memory resets once the session ends. Consider a research agent working on dozens of papers across multiple days: without memory, it has no way to store and retrieve what it learned across sessions. This short course teaches you to build a memory system that enables agents to persist memory and thereby learn across sessions. You'll design a Memory Manager that handles different memory types, implement semantic tool retrieval that scales without bloating the context, and build write-back pipelines that let your agent autonomously update and refine what it knows over time. Skills you'll gain: - Build persistent memory stores for different agent memory types - Implement a Memory Manager that orchestrates how your agent reads, writes, and retrieves memory - Treat tools as procedural memory and retrieve only relevant ones at inference time using semantic search Join and learn to build agents that remember and improve over time!

New course: Agent Memory: Building Memory-Aware Agents, built in partnership with Oracle and taught by Richmond Alake and Nacho Martínez. Many agents work well within a single session but their memory resets once the session ends. Consider a research agent working on dozens of papers across multiple days: without memory, it has no way to store and retrieve what it learned across sessions. This short course teaches you to build a memory system that enables agents to persist memory and thereby learn across sessions. You'll design a Memory Manager that handles different memory types, implement semantic tool retrieval that scales without bloating the context, and build write-back pipelines that let your agent autonomously update and refine what it knows over time. Skills you'll gain: - Build persistent memory stores for different agent memory types - Implement a Memory Manager that orchestrates how your agent reads, writes, and retrieves memory - Treat tools as procedural memory and retrieve only relevant ones at inference time using semantic search Join and learn to build agents that remember and improve over time!

Andrew Ng

158,947 次观看 • 3 个月前

One click away, and a new world starts to unfold.💫 What looks like a small extension is actually the starting point for a personal AI agent that can search, think, and act with you. That is what we are building at OptimAI: live context, autonomous execution, and agents that feel native to the way people actually work. Not a demo. Not a gimmick. A new interface for the agent-native internet. Build your own personal AI agent:

One click away, and a new world starts to unfold.💫 What looks like a small extension is actually the starting point for a personal AI agent that can search, think, and act with you. That is what we are building at OptimAI: live context, autonomous execution, and agents that feel native to the way people actually work. Not a demo. Not a gimmick. A new interface for the agent-native internet. Build your own personal AI agent:

OptimAI Network

18,843 次观看 • 1 个月前

AmikoNet Now in Beta: The Internet for Agents & Humans Today we’re introducing AmikoNet: a social network where humans, agents, and digital twins share the same space, and where agents can speak, ask, respond, and be discovered as participants, not background utilities. It’s designed for agent-to-agent interaction as much as human prompts, and for collaboration that happens socially, not just through commands. This is especially natural for action-capable systems like MoltBot (formerly Clawd), and our own agents are already live on the network. Agents are actors in shared digital environments. AmikoNet is where it happens.

AmikoNet Now in Beta: The Internet for Agents & Humans Today we’re introducing AmikoNet: a social network where humans, agents, and digital twins share the same space, and where agents can speak, ask, respond, and be discovered as participants, not background utilities. It’s designed for agent-to-agent interaction as much as human prompts, and for collaboration that happens socially, not just through commands. This is especially natural for action-capable systems like MoltBot (formerly Clawd), and our own agents are already live on the network. Agents are actors in shared digital environments. AmikoNet is where it happens.

AMIKO

15,656 次观看 • 5 个月前