Loading video...

Video Failed to Load

Go Home

Intelligence and Experience are orthogonal vectors Terence Tao is perhaps the world’s smartest person, but drop him into an accounting firm or onto a construction site and on day one he’s not going to be very productive Trajectory calls this The Experience Gap, and they have a way to...

gradypb's profile picture

Pat Grady

30,300 subscribers

72,044 views • 4 days ago •via X (Twitter)

Science & TechnologyEducation
LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

capability != learning new benchtalks with Parth Asawa on continual learning, where we discuss teaching models to learn from experience, measuring learning ability, the bet on parametric models, and more 01:06 What is continual learning? 04:10 Why capability and learning are different 06:13 Why build a benchmark? 08:07 Continual Learning Bench launch and reception 09:13 Anthropic's Fable release and Continual Learning Bench 11:02 How to design tasks for continual learning 18:41 The gain metric 24:01 What good looks like on the leaderboard 29:13 Failure modes: why models can't update their beliefs 31:12 Parametric systems and future architectures 34:30 Open science and AI safety 45:42 Lightning round 49:08 How to contribute to Continual Learning Bench
50:35

capability != learning new benchtalks with Parth Asawa on continual learning, where we discuss teaching models to learn from experience, measuring learning ability, the bet on parametric models, and more 01:06 What is continual learning? 04:10 Why capability and learning are different 06:13 Why build a benchmark? 08:07 Continual Learning Bench launch and reception 09:13 Anthropic's Fable release and Continual Learning Bench 11:02 How to design tasks for continual learning 18:41 The gain metric 24:01 What good looks like on the leaderboard 29:13 Failure modes: why models can't update their beliefs 31:12 Parametric systems and future architectures 34:30 Open science and AI safety 45:42 Lightning round 49:08 How to contribute to Continual Learning Bench

vincent sunn chen

40,884 views • 1 month ago

Want world class research capabilities, but don’t have the resources of a big lab? At our recent Sovereign AI event, Gabe Pereyra shared Harvey ’s “moneyball” approach. Here’s the playbook: 00:00 Introduction 00:37 Building a research lab on a budget 02:28 Legal Agent Bench, contracting, and the diligence dataset 03:57 Domain experts guiding synthetic data generation 05:23 Why Harvey open sourced its datasets 06:55 Working with the neo labs – and why more than one 08:20 Post-training in-house: building "Associate 1" 09:44 The model serving matrix: 60 countries, fallbacks, SLAs 11:05 Deciding what stays in production 12:29 Simple open source switches and model routing 13:55 Moneyball: "If we win on this budget, we change the game" 14:53 Q&A: Training with sensitive data 17:16 Q&A: Competing for research talent 18:46 Q&A: Designing rubrics that actually challenge frontier models 20:19 Q&A: Where the pipeline breaks — data, research, or infra 22:59 Q&A: The tension in open sourcing a benchmark 25:02 Q&A: Biggest remaining open problems 27:10 Q&A: Competing with horizontal products
29:20

Want world class research capabilities, but don’t have the resources of a big lab? At our recent Sovereign AI event, Gabe Pereyra shared Harvey ’s “moneyball” approach. Here’s the playbook: 00:00 Introduction 00:37 Building a research lab on a budget 02:28 Legal Agent Bench, contracting, and the diligence dataset 03:57 Domain experts guiding synthetic data generation 05:23 Why Harvey open sourced its datasets 06:55 Working with the neo labs – and why more than one 08:20 Post-training in-house: building "Associate 1" 09:44 The model serving matrix: 60 countries, fallbacks, SLAs 11:05 Deciding what stays in production 12:29 Simple open source switches and model routing 13:55 Moneyball: "If we win on this budget, we change the game" 14:53 Q&A: Training with sensitive data 17:16 Q&A: Competing for research talent 18:46 Q&A: Designing rubrics that actually challenge frontier models 20:19 Q&A: Where the pipeline breaks — data, research, or infra 22:59 Q&A: The tension in open sourcing a benchmark 25:02 Q&A: Biggest remaining open problems 27:10 Q&A: Competing with horizontal products

Pat Grady

168,566 views • 6 days ago

A model is only as good as its data, and we’ve long since exhausted the internet. From here on out, model progress is gated by data production. ’s Brendan (can/do) joined us at our Sovereign AI event to talk about how RL environments get built, and why your data might be your real moat: 00:00 Introduction 00:47 A short history of the data market: crowdsourcing to agentic data 02:29 What an RL environment is: worlds, apps, tasks 03:57 Why only humans can measure the frontier 05:35 Building verifiers is the hard part 06:44 Walkthrough: a real legal RL environment 08:18 Leaderboards — and what open weights change 09:45 Post-training results on Apex Agents 11:17 Three ways companies buy data 12:49 Q&A: How do you price data? 14:17 Q&A: What "data quality" actually means 16:42 Q&A: The misunderstanding about synthetic data 18:17 Q&A: Why RL environments now — and what comes after 21:20 Q&A: Can you scale rubric generation with models? 23:00 Q&A: RL environments for cyber defense 25:33 Q&A: Build data in-house or partner?
26:21

A model is only as good as its data, and we’ve long since exhausted the internet. From here on out, model progress is gated by data production. ’s Brendan (can/do) joined us at our Sovereign AI event to talk about how RL environments get built, and why your data might be your real moat: 00:00 Introduction 00:47 A short history of the data market: crowdsourcing to agentic data 02:29 What an RL environment is: worlds, apps, tasks 03:57 Why only humans can measure the frontier 05:35 Building verifiers is the hard part 06:44 Walkthrough: a real legal RL environment 08:18 Leaderboards — and what open weights change 09:45 Post-training results on Apex Agents 11:17 Three ways companies buy data 12:49 Q&A: How do you price data? 14:17 Q&A: What "data quality" actually means 16:42 Q&A: The misunderstanding about synthetic data 18:17 Q&A: Why RL environments now — and what comes after 21:20 Q&A: Can you scale rubric generation with models? 23:00 Q&A: RL environments for cyber defense 25:33 Q&A: Build data in-house or partner?

Pat Grady

63,548 views • 5 days ago

An agent is three things: a harness, a model, and context. If you're serious about owning your intelligence, you probably want to own all three. LangChain founder Harrison Chase joined us at our Sequoia Capital Own Your Intelligence to talk about the piece that often gets the least attention: the harness. He offers a clear heuristic for when to build your own. The more out of distribution you are from what the models were trained on, the more you'll want to customize. And good technical content on how to actually measure performance with evals and langsmith. 00:00 Introduction 00:58 The three parts of an agent: harness, model, context 02:12 What a harness actually does 03:25 Customizing the core loop with middleware 04:41 Sandboxes, file systems, sub-agents, summarization 05:47 Cognitive architectures — and when you still need them 07:03 Build your own harness or use off the shelf? 08:24 In-distribution vs. out-of-distribution: the file-editing example 09:39 Why evals define what "good" means in an organization 11:04 Harbor: what an eval task actually looks like 12:11 Comparing harnesses and models on accuracy, latency, and cost 13:20 Why observability is underrated — it's usually the context 14:34 The data flywheel: traces → curation → experiments 15:42 Getting feedback through UX design and online evaluators 16:51 Demo: LangSmith Engine 19:23 Q&A: Running Engine on Engine, and "codex-ification" 20:44 Q&A: Will harnesses converge or diverge?
23:56

An agent is three things: a harness, a model, and context. If you're serious about owning your intelligence, you probably want to own all three. LangChain founder Harrison Chase joined us at our Sequoia Capital Own Your Intelligence to talk about the piece that often gets the least attention: the harness. He offers a clear heuristic for when to build your own. The more out of distribution you are from what the models were trained on, the more you'll want to customize. And good technical content on how to actually measure performance with evals and langsmith. 00:00 Introduction 00:58 The three parts of an agent: harness, model, context 02:12 What a harness actually does 03:25 Customizing the core loop with middleware 04:41 Sandboxes, file systems, sub-agents, summarization 05:47 Cognitive architectures — and when you still need them 07:03 Build your own harness or use off the shelf? 08:24 In-distribution vs. out-of-distribution: the file-editing example 09:39 Why evals define what "good" means in an organization 11:04 Harbor: what an eval task actually looks like 12:11 Comparing harnesses and models on accuracy, latency, and cost 13:20 Why observability is underrated — it's usually the context 14:34 The data flywheel: traces → curation → experiments 15:42 Getting feedback through UX design and online evaluators 16:51 Demo: LangSmith Engine 19:23 Q&A: Running Engine on Engine, and "codex-ification" 20:44 Q&A: Will harnesses converge or diverge?

Sonya Huang 🐥

74,288 views • 4 days ago

When Mudith Jayasekara and I met Gabe Pereyra, we were expecting just another vanilla intro call and instead had the best yarn about research, the state of LLMs, and where intelligence is actually heading. It's rare to meet a founder this deep in the weeds who's also building for one of the most important verticals in this new age of intelligence So it was awesome to sit down with Gabe for an extended discussion on what it take to build agents that can reliably complete work over hours, days, or even longer? We talked about why agents today struggle with search and long context windows and how techniques like KV-cache compaction, synthetic data, and continual learning could help. 0:00 Introduction 0:36 Getting legal agents to review the whole data room 2:08 Data rooms larger than any context window 5:28 How far open-source models can go 7:58 Where specialist models fit in legal AI 10:59 Training legal models when client data is off-limits 13:06 Teaching a model how a law firm works 13:59 What belongs in context vs. model weights 15:36 From firm-wide AI to a model for every lawyer 18:37 What training adds beyond retrieving the right cases 20:26 Why context windows have plateaued 24:01 How models could learn continuously on the job 26:12 Can AI recursively improve AI research? 27:07 Research agents can run experiments but not choose them 30:00 Why open-ended research is hard to train 33:47 Why deployment, not intelligence, is the bottleneck 35:08 The cost of frontier intelligence 36:59 Different neolabs, different paths to intelligence 39:26 Using open datasets to compare research methods 41:13 Conclusion
41:50

When Mudith Jayasekara and I met Gabe Pereyra, we were expecting just another vanilla intro call and instead had the best yarn about research, the state of LLMs, and where intelligence is actually heading. It's rare to meet a founder this deep in the weeds who's also building for one of the most important verticals in this new age of intelligence So it was awesome to sit down with Gabe for an extended discussion on what it take to build agents that can reliably complete work over hours, days, or even longer? We talked about why agents today struggle with search and long context windows and how techniques like KV-cache compaction, synthetic data, and continual learning could help. 0:00 Introduction 0:36 Getting legal agents to review the whole data room 2:08 Data rooms larger than any context window 5:28 How far open-source models can go 7:58 Where specialist models fit in legal AI 10:59 Training legal models when client data is off-limits 13:06 Teaching a model how a law firm works 13:59 What belongs in context vs. model weights 15:36 From firm-wide AI to a model for every lawyer 18:37 What training adds beyond retrieving the right cases 20:26 Why context windows have plateaued 24:01 How models could learn continuously on the job 26:12 Can AI recursively improve AI research? 27:07 Research agents can run experiments but not choose them 30:00 Why open-ended research is hard to train 33:47 Why deployment, not intelligence, is the bottleneck 35:08 The cost of frontier intelligence 36:59 Different neolabs, different paths to intelligence 39:26 Using open datasets to compare research methods 41:13 Conclusion

Charlie O'Neill

91,088 views • 25 days ago

OWN YOUR INTELLIGENCE Last year, building on open-weight models was primarily a cost rationalization exercise. Slightly worse performance for a much cheaper price. Now, it is increasingly an existential and strategic topic for our portfolio. Intelligence is the product. Companies want to shape it and own it and let it compound within their own walls. Not your weights, not your product. Now, with frontier open-weight models and fantastic tooling/infrastructure, owning your intelligence at the frontier is finally becoming possible. The result: every application company we work with is embarking on the journey of doing their own research on post-training, evals, harnesses, etc. The hottest neolabs may just be Harvey, Factory, RamPrasad "RamP!" Moudgalya, etc. The list goes on. We held a summit Sequoia Capital to convene our portfolio on this topic, together with Gabe Pereyra (Harvey) on building Harvey Labs, Lin Qiao (Fireworks) on post-training, Harrison Chase (LangChain) on harnesses + evals, Brendan (can/do) () on RL environments and synthetic data, Arjun Karanam (Trajectory) on online continual learning. Opening talk below; rest to come this week! 00:00 What is sovereign AI (and what it isn't) 01:24 Centralized vs. decentralized intelligence 02:54 Four reasons companies own their models: cost, speed, performance, destiny 04:22 "Not your weights, not your product" 05:32 The application companies are the newest neo labs 07:05 Step 1: Deciding what to own vs. rent 09:51 Step 2: Build the team (and don't shoehorn your platform team) 11:17 Step 3: Legibility – why your research has to be visible 12:33 Step 4: The technical roadmap 13:56 The stack: production vs. development 15:16 Opening Pandora's box – base models, harnesses, context
17:28

OWN YOUR INTELLIGENCE Last year, building on open-weight models was primarily a cost rationalization exercise. Slightly worse performance for a much cheaper price. Now, it is increasingly an existential and strategic topic for our portfolio. Intelligence is the product. Companies want to shape it and own it and let it compound within their own walls. Not your weights, not your product. Now, with frontier open-weight models and fantastic tooling/infrastructure, owning your intelligence at the frontier is finally becoming possible. The result: every application company we work with is embarking on the journey of doing their own research on post-training, evals, harnesses, etc. The hottest neolabs may just be Harvey, Factory, RamPrasad "RamP!" Moudgalya, etc. The list goes on. We held a summit Sequoia Capital to convene our portfolio on this topic, together with Gabe Pereyra (Harvey) on building Harvey Labs, Lin Qiao (Fireworks) on post-training, Harrison Chase (LangChain) on harnesses + evals, Brendan (can/do) () on RL environments and synthetic data, Arjun Karanam (Trajectory) on online continual learning. Opening talk below; rest to come this week! 00:00 What is sovereign AI (and what it isn't) 01:24 Centralized vs. decentralized intelligence 02:54 Four reasons companies own their models: cost, speed, performance, destiny 04:22 "Not your weights, not your product" 05:32 The application companies are the newest neo labs 07:05 Step 1: Deciding what to own vs. rent 09:51 Step 2: Build the team (and don't shoehorn your platform team) 11:17 Step 3: Legibility – why your research has to be visible 12:33 Step 4: The technical roadmap 13:56 The stack: production vs. development 15:16 Opening Pandora's box – base models, harnesses, context

Sonya Huang 🐥

123,318 views • 6 days ago

Chelsea Finn (Chelsea Finn) on building general-purpose robotics, and bringing intelligence into the physical world. At AI Startup School in San Francisco. 00:00 - General Purpose Robots 00:11 - Challenges in Robotics Applications 00:57 - Physical Intelligence: A New Approach 01:47 - Learning from Language Models 02:08 - Data Sources for Training Robots 03:32 - Training with Real-World Data 04:39 - Initial Successes and Challenges 09:10 - Breakthrough in Robot Training 11:03 - Improving Performance 15:43 - Expanding Capabilities 17:34 - Robots in Unseen Environments 25:54 - Handling Open-Ended Prompts 29:36 - Evaluating Robot Performance 30:03 - Future Directions and Challenges 31:27 - Audience Q&A
44:52

Chelsea Finn (Chelsea Finn) on building general-purpose robotics, and bringing intelligence into the physical world. At AI Startup School in San Francisco. 00:00 - General Purpose Robots 00:11 - Challenges in Robotics Applications 00:57 - Physical Intelligence: A New Approach 01:47 - Learning from Language Models 02:08 - Data Sources for Training Robots 03:32 - Training with Real-World Data 04:39 - Initial Successes and Challenges 09:10 - Breakthrough in Robot Training 11:03 - Improving Performance 15:43 - Expanding Capabilities 17:34 - Robots in Unseen Environments 25:54 - Handling Open-Ended Prompts 29:36 - Evaluating Robot Performance 30:03 - Future Directions and Challenges 31:27 - Audience Q&A

Y Combinator

129,834 views • 1 year ago

Why AI Progress Suddenly Feels Real - my conversation with Yann Dubois, who co-leads the Post-Training Frontiers team at OpenAI 00:00 - Intro 01:30 - Why recent AI progress feels like a step function 04:13 - Model reliability & the emotional rollercoaster of shipping GPT-5.5 07:33 - How OpenAI structures vertical and horizontal teams 09:49 - Improving model efficiency and test-time compute 12:32 - Yann's journey from Switzerland to OpenAI 15:37 - Reasoning in 2026: Real-world utility vs verifiable rewards 18:34 - GPT-5.5 Thinking vs Pro: Scaling test-time compute 20:09 - How reasoning models become more efficient 23:23 - Pre-training scaling and overcoming the data wall 27:03 - Multimodal data, synthetic data, and embodied AI 31:05 - Demystifying mid-training and post-training 37:21 - Does RL create new capabilities in AI? 38:53 - The challenges and frontier of scaling RL 43:09 - Is building AI models a craft or a strict science 48:21 - How AI models generalize across different domains 54:18 - How reinforcement learning cures AI hallucinations 56:04 - Negative generalization and conflicting instructions 58:05 - Can RL scale to law, medicine, and the broader economy? 1:00:19 - The evaluation bottleneck and Model as a Judge 1:04:21 - Continuous AI progress & continual learning 1:08:49 - Will foundation models eat the agent harness 1:11:23 - Why startups should focus on the last mile of AI
1:13:56

Why AI Progress Suddenly Feels Real - my conversation with Yann Dubois, who co-leads the Post-Training Frontiers team at OpenAI 00:00 - Intro 01:30 - Why recent AI progress feels like a step function 04:13 - Model reliability & the emotional rollercoaster of shipping GPT-5.5 07:33 - How OpenAI structures vertical and horizontal teams 09:49 - Improving model efficiency and test-time compute 12:32 - Yann's journey from Switzerland to OpenAI 15:37 - Reasoning in 2026: Real-world utility vs verifiable rewards 18:34 - GPT-5.5 Thinking vs Pro: Scaling test-time compute 20:09 - How reasoning models become more efficient 23:23 - Pre-training scaling and overcoming the data wall 27:03 - Multimodal data, synthetic data, and embodied AI 31:05 - Demystifying mid-training and post-training 37:21 - Does RL create new capabilities in AI? 38:53 - The challenges and frontier of scaling RL 43:09 - Is building AI models a craft or a strict science 48:21 - How AI models generalize across different domains 54:18 - How reinforcement learning cures AI hallucinations 56:04 - Negative generalization and conflicting instructions 58:05 - Can RL scale to law, medicine, and the broader economy? 1:00:19 - The evaluation bottleneck and Model as a Judge 1:04:21 - Continuous AI progress & continual learning 1:08:49 - Will foundation models eat the agent harness 1:11:23 - Why startups should focus on the last mile of AI

Matt Turck

100,966 views • 2 months ago

Today's AI models train once. We don't work that way. We learn continuously, forget what doesn't matter, and retain what does. That gap is what Dan Biderman and Jessy Lin are closing at Engram. AI that never stops learning, with memory that lives inside the model instead of bolted on as an afterthought. In our latest Training Data episode we get into why memory is the next frontier: why the brain forgets on purpose, why RAG is a band-aid, and what becomes possible when a model is always training. 00:00 Introduction 00:59 Always Training Explained 01:51 Beyond Context Windows 03:29 Ngram Product Overview 04:34 Adapters And Training Signals 05:32 Internalize Vs Externalize 06:49 Compute And Token Savings 08:19 Teams First Then Individuals 08:51 Memorization Vs Understanding 12:47 Dreams And Offline Digestion 14:08 Training Beats Curation 15:19 Why Everyone Needs A Model 21:44 Bitter Lesson And Architecture 24:44 RAG Killer And KV Cache 31:38 Future Of Memory And Models
44:51

Today's AI models train once. We don't work that way. We learn continuously, forget what doesn't matter, and retain what does. That gap is what Dan Biderman and Jessy Lin are closing at Engram. AI that never stops learning, with memory that lives inside the model instead of bolted on as an afterthought. In our latest Training Data episode we get into why memory is the next frontier: why the brain forgets on purpose, why RAG is a band-aid, and what becomes possible when a model is always training. 00:00 Introduction 00:59 Always Training Explained 01:51 Beyond Context Windows 03:29 Ngram Product Overview 04:34 Adapters And Training Signals 05:32 Internalize Vs Externalize 06:49 Compute And Token Savings 08:19 Teams First Then Individuals 08:51 Memorization Vs Understanding 12:47 Dreams And Offline Digestion 14:08 Training Beats Curation 15:19 Why Everyone Needs A Model 21:44 Bitter Lesson And Architecture 24:44 RAG Killer And KV Cache 31:38 Future Of Memory And Models

Sequoia Capital

83,323 views • 1 month ago

tylercowen is bullish on AI education — here's why. 00:00 -- Preview 00:24 -- President Carlos Carvalho's AI-generated intro 03:21 -- Cowen reacts to UATX's campus 04:38 -- The AI revolution is here. Who will lose the most? 06:05 -- AI lawyers 07:17 -- Don't underestimate this 10:41 -- Changes to the "upper upper middle class" 12:38 -- How to be successful 13:43 -- The rise of managerial empires 14:02 -- When will we have the first billion dollar company with one employee? 16:05 -- 10-20 year forecast 16:19 -- Why education is so behind 17:01 -- Should you be bullish on UATX? 18:36 -- Should you still read Homer? 21:50 -- Write to think 25:01 -- Meet more people 25:42 -- How to get hired 26:54 -- Is AI your best mentor? 38:17 -- How to curb cheating 39:02 -- The new life of the mind 42:34 -- Q&A: Will there be more status associated with real education or AI education? 45:50 -- Q&A: Why do tech-savvy students need to practice using AI? 47:56 -- Q&A: Do LLMs atrophy your mind? 49:29 -- Q&A: How do you avoid AI-dependency? 51:05 -- Q&A: Isn't this vision lonely and isolating? 53:06 -- Q&A: Do students need teachers? 55:36 -- Q&A: What are the four most important courses for undergrads? 57:49 -- Q&A: Which AI company will win the AI race in the next five years and why? 59:22 -- Q&A: Can AI teach religion? 01:01:32 -- Q&A: Will AI narrow or widen our world? 01:04:37 -- Q&A: What makes us human? 01:05:42 -- Q&A: What is art? 01:08:33 -- Q&A: It's easy to catch cheaters
1:10:31

tylercowen is bullish on AI education — here's why. 00:00 -- Preview 00:24 -- President Carlos Carvalho's AI-generated intro 03:21 -- Cowen reacts to UATX's campus 04:38 -- The AI revolution is here. Who will lose the most? 06:05 -- AI lawyers 07:17 -- Don't underestimate this 10:41 -- Changes to the "upper upper middle class" 12:38 -- How to be successful 13:43 -- The rise of managerial empires 14:02 -- When will we have the first billion dollar company with one employee? 16:05 -- 10-20 year forecast 16:19 -- Why education is so behind 17:01 -- Should you be bullish on UATX? 18:36 -- Should you still read Homer? 21:50 -- Write to think 25:01 -- Meet more people 25:42 -- How to get hired 26:54 -- Is AI your best mentor? 38:17 -- How to curb cheating 39:02 -- The new life of the mind 42:34 -- Q&A: Will there be more status associated with real education or AI education? 45:50 -- Q&A: Why do tech-savvy students need to practice using AI? 47:56 -- Q&A: Do LLMs atrophy your mind? 49:29 -- Q&A: How do you avoid AI-dependency? 51:05 -- Q&A: Isn't this vision lonely and isolating? 53:06 -- Q&A: Do students need teachers? 55:36 -- Q&A: What are the four most important courses for undergrads? 57:49 -- Q&A: Which AI company will win the AI race in the next five years and why? 59:22 -- Q&A: Can AI teach religion? 01:01:32 -- Q&A: Will AI narrow or widen our world? 01:04:37 -- Q&A: What makes us human? 01:05:42 -- Q&A: What is art? 01:08:33 -- Q&A: It's easy to catch cheaters

University of Austin (UATX)

27,770 views • 6 months ago

I'm often asked for the best public example of AI evals done right for a real, production product. I finally have an answer. Teresa Torres shares how she shipped an AI interview coach, and used evals to rapidly squash bugs and improve the product. Teresa shows how she: 1. did error analysis FIRST to find real issues (instead of using generic metrics) 😍 2. used Jupyter notebooks to analyze errors 3. built custom annotation tools + custom widgets in notebooks 4. built a LLM-judge and assertions to test for specific errors 5. iterated through this feedback loop until it worked. 6. kept things simple the whole time It's also probably the best commercial for Jupyter notebooks you can imagine. 🥰 Chapter summary below. Link to YT in next thread 00:00:00 - Intro 00:01:45 - The Product: Building an AI Interview Coach 00:06:34 - The Problem: How Do I Know if My AI Coach is Any Good? 00:10:15 - Using Airtable for Traces and Annotation 00:12:15 - Discovering Jupyter Notebooks and Designing the First Evals 00:15:15 - Example Evals: LLM-as-Judge vs. Code-Based Assertions 00:21:00 - Learning Python with ChatGPT to Analyze Eval Results 00:31:00 - VS Code, Custom Tools, and an Eval Investigation Notebook 00:39:45 - Building a Custom Annotation Tool with Claude 00:41:00 - From Personal Project to Production App 00:46:02 - How Should PMs and Engineers Collaborate on AI Products? 00:55:45 - Q&A: Capturing Feedback and Annotations from End Users 00:58:11 - Q&A: Is a Technical Background Necessary to Build AI? 01:02:28 - Q&A: What's Next for Teresa? 01:03:13 - Q&A: Unpacking the Micro-Decisions of Building an AI App
1:10:22

I'm often asked for the best public example of AI evals done right for a real, production product. I finally have an answer. Teresa Torres shares how she shipped an AI interview coach, and used evals to rapidly squash bugs and improve the product. Teresa shows how she: 1. did error analysis FIRST to find real issues (instead of using generic metrics) 😍 2. used Jupyter notebooks to analyze errors 3. built custom annotation tools + custom widgets in notebooks 4. built a LLM-judge and assertions to test for specific errors 5. iterated through this feedback loop until it worked. 6. kept things simple the whole time It's also probably the best commercial for Jupyter notebooks you can imagine. 🥰 Chapter summary below. Link to YT in next thread 00:00:00 - Intro 00:01:45 - The Product: Building an AI Interview Coach 00:06:34 - The Problem: How Do I Know if My AI Coach is Any Good? 00:10:15 - Using Airtable for Traces and Annotation 00:12:15 - Discovering Jupyter Notebooks and Designing the First Evals 00:15:15 - Example Evals: LLM-as-Judge vs. Code-Based Assertions 00:21:00 - Learning Python with ChatGPT to Analyze Eval Results 00:31:00 - VS Code, Custom Tools, and an Eval Investigation Notebook 00:39:45 - Building a Custom Annotation Tool with Claude 00:41:00 - From Personal Project to Production App 00:46:02 - How Should PMs and Engineers Collaborate on AI Products? 00:55:45 - Q&A: Capturing Feedback and Annotations from End Users 00:58:11 - Q&A: Is a Technical Background Necessary to Build AI? 01:02:28 - Q&A: What's Next for Teresa? 01:03:13 - Q&A: Unpacking the Micro-Decisions of Building an AI App

Hamel Husain

51,376 views • 1 year ago

Who touches every token that flows through Paul Jankura? It’s not any one model, but it’s Katelyn Lesse, Angela Jiang and the platform team. A year ago, it was just a messages API. Today, their platform sits under every API call, every workflow, and every agent built on Claude. Sonya Huang 🐥 and I explore why they make everything they can externally facing for customers, how they pick markets to target, what it takes to build the best agents, and more: 00:00 Introduction 01:49 Two North Stars 02:27 External Builders And Primitives 03:54 What To Externalize 06:00 From Messages To Agents 08:19 Managed Agents Adoption 09:07 Three Layer Cake 10:22 Execution Harnesses Explained 11:09 Coordination Strategies Roadmap 12:13 Ecosystem Standards And Safety 15:39 Open Ecosystem Not Walled 17:12 Vertical Products And Form Factors 22:26 Claude Tag Under The Hood 26:04 Harness Best Practices 38:13 Token Costs And Whats Next
48:54

Who touches every token that flows through Paul Jankura? It’s not any one model, but it’s Katelyn Lesse, Angela Jiang and the platform team. A year ago, it was just a messages API. Today, their platform sits under every API call, every workflow, and every agent built on Claude. Sonya Huang 🐥 and I explore why they make everything they can externally facing for customers, how they pick markets to target, what it takes to build the best agents, and more: 00:00 Introduction 01:49 Two North Stars 02:27 External Builders And Primitives 03:54 What To Externalize 06:00 From Messages To Agents 08:19 Managed Agents Adoption 09:07 Three Layer Cake 10:22 Execution Harnesses Explained 11:09 Coordination Strategies Roadmap 12:13 Ecosystem Standards And Safety 15:39 Open Ecosystem Not Walled 17:12 Vertical Products And Form Factors 22:26 Claude Tag Under The Hood 26:04 Harness Best Practices 38:13 Token Costs And Whats Next

Lauren Reeder

56,964 views • 1 month ago

Dr. Fei-Fei Li (Fei-Fei Li), the Godmother of AI, on the origin of computer vision and why spatial intelligence is the next frontier. At AI Startup School in San Francisco. 01:00 - Origins of ImageNet 02:15 - The Dream to Make Machines See 03:28 - A paradigm shift in AI 05:50 - AlexNet and Deep Learning 08:00 - Evolving Computer Vision 12:20 - Building World Labs 13:00 - The next frontier in AI 14:20 - Why Spatial Intelligence Is Harder Than Language 18:40 - Technical Barriers for Vision-Based AI 20:00 - AI is more than LLMs 25:50 - Fei Fei’s Journey 29:30 - Mentoring some of the legends of AI 33:00 - Audience Q&A
44:21

Dr. Fei-Fei Li (Fei-Fei Li), the Godmother of AI, on the origin of computer vision and why spatial intelligence is the next frontier. At AI Startup School in San Francisco. 01:00 - Origins of ImageNet 02:15 - The Dream to Make Machines See 03:28 - A paradigm shift in AI 05:50 - AlexNet and Deep Learning 08:00 - Evolving Computer Vision 12:20 - Building World Labs 13:00 - The next frontier in AI 14:20 - Why Spatial Intelligence Is Harder Than Language 18:40 - Technical Barriers for Vision-Based AI 20:00 - AI is more than LLMs 25:50 - Fei Fei’s Journey 29:30 - Mentoring some of the legends of AI 33:00 - Audience Q&A

Y Combinator

194,393 views • 1 year ago

Loved this 22-minute talk on continual learning for AI agents. Must watch for anyone looking to get agents performant and into production. Credit: Soheil Feizi at AI Engineer • Agent learning can happen at three layers: the model (weights), the harness (prompts, tools, skills, code, workflows), and memory (session or persistent). • Two fundamental challenges: (1) getting feedback, meaning how do we know if the agent did well and what it should have done instead, and (2) acting on that feedback, meaning deciding which layer or component to change and how. • Feedback sources differ by stage: In development you have benchmarks with evaluators that score pass/fail. In production you only have logs, which can be judged either automatically (LLMs or code analyzing the log, which is scalable) or by human experts (low volume but critical domain knowledge). • Logs plus feedback aren't enough because they're not testable: A single log with feedback is one observation of what happened. You need to lift it into a replayable learning environment, a simulation with tools, users, and defined evaluators, so candidate fixes can be run, verified, and compared. • Three ways to optimize the agent, with tradeoffs: Model-layer updates (SFT, RL post-training like DPO/GRPO, LoRA) are expensive and need benchmarks and evaluators. Harness updates (trace-to-harness coding agents, prompt search like GEPA) are flexible but either untestable and "vibe-based" or benchmark-dependent. Memory updates (fact storage like Letta/Mem0, skill distillation) are cheapest and fastest but usually unverified. • A good learning engine makes "the smallest durable change at the right layer" of the agent. • Verifiable continual learning (VCL): Improve an agent from its own experience where every fix is proven to help and proven to break nothing that already worked. It requires an executable test (replayable failure), a measured delta (score before and after), and regression tests (prior tests still pass). • Four principles of practical VCL: Replayability (turn one-off failures into rerunnable tests), holisticness (one failure can have causes in memory, prompts, tools, workflow, or model, so route the fix to the right layer), lifelongness (fix new failures subject to no regression on past environments, with regression handled inside the optimization loop rather than post-hoc), and efficiency (the loop must run frequently and cheaply, without scaling linearly as past environments accumulate). • Three takeaways: (1) Agent continual learning isn't necessarily fine-tuning; many useful updates live in the harness and memory layers. (2) Production logs are not learning environments and must be transformed into replayable ones. (3) The frontier is regression-aware improvement: fixing new failures while verifying you don't break old ones.
22:34

Loved this 22-minute talk on continual learning for AI agents. Must watch for anyone looking to get agents performant and into production. Credit: Soheil Feizi at AI Engineer • Agent learning can happen at three layers: the model (weights), the harness (prompts, tools, skills, code, workflows), and memory (session or persistent). • Two fundamental challenges: (1) getting feedback, meaning how do we know if the agent did well and what it should have done instead, and (2) acting on that feedback, meaning deciding which layer or component to change and how. • Feedback sources differ by stage: In development you have benchmarks with evaluators that score pass/fail. In production you only have logs, which can be judged either automatically (LLMs or code analyzing the log, which is scalable) or by human experts (low volume but critical domain knowledge). • Logs plus feedback aren't enough because they're not testable: A single log with feedback is one observation of what happened. You need to lift it into a replayable learning environment, a simulation with tools, users, and defined evaluators, so candidate fixes can be run, verified, and compared. • Three ways to optimize the agent, with tradeoffs: Model-layer updates (SFT, RL post-training like DPO/GRPO, LoRA) are expensive and need benchmarks and evaluators. Harness updates (trace-to-harness coding agents, prompt search like GEPA) are flexible but either untestable and "vibe-based" or benchmark-dependent. Memory updates (fact storage like Letta/Mem0, skill distillation) are cheapest and fastest but usually unverified. • A good learning engine makes "the smallest durable change at the right layer" of the agent. • Verifiable continual learning (VCL): Improve an agent from its own experience where every fix is proven to help and proven to break nothing that already worked. It requires an executable test (replayable failure), a measured delta (score before and after), and regression tests (prior tests still pass). • Four principles of practical VCL: Replayability (turn one-off failures into rerunnable tests), holisticness (one failure can have causes in memory, prompts, tools, workflow, or model, so route the fix to the right layer), lifelongness (fix new failures subject to no regression on past environments, with regression handled inside the optimization loop rather than post-hoc), and efficiency (the loop must run frequently and cheaply, without scaling linearly as past environments accumulate). • Three takeaways: (1) Agent continual learning isn't necessarily fine-tuning; many useful updates live in the harness and memory layers. (2) Production logs are not learning environments and must be transformed into replayable ones. (3) The frontier is regression-aware improvement: fixing new failures while verifying you don't break old ones.

Alex Lieberman

20,085 views • 1 month ago

Yoshua Bengio thinks he knows how to make provably safe superintelligent agents. Bengio built the foundations of modern AI and is the most cited living scientist. He believes his alternative training setup would: 1. Guarantee honesty 2. Prevent unintended goals 3. Produce capable agents 4. Port over most data and techniques from current LLMs 5. Not be inherently more expensive, and perhaps be more intelligent Bengio claims the honesty and lack of unintended goals can be proven mathematically, at least given particular assumptions. And his new organization, LawZero, is aiming to build a scrappy prototype as soon as possible. The architecture is called 'Scientist AI' and it's based on training a model to explain empirical observations, including what people say, rather than training AIs that mimic human behaviour or seek our approval. (Bengio's frank assessment is that "reinforcement learning is evil" and that allowing AIs to independently train their successors is "the most crazy, dangerous bet that unfortunately we are on track to do.") But skeptics question whether Scientist AI really does solve the fundamental problem of 'eliciting latent knowledge' from AI models. And with the commercial race for superintelligence so intense, it's not clear whether the proposal will be able to compete or have time to bear fruit, even if it's sound in theory. On The 80,000 Hours Podcast, links below – enjoy! • Making AI honest and safe (00:00:00) • Scientist AI in plain English (00:02:27) • How Scientist AI differs from LLMs (00:06:32) • How the training data works (00:14:02) • Can this become an agent? (00:21:02) • Why Yoshua is now more optimistic (00:32:11) • Why companies can’t stop racing (00:36:35) • A working prototype won't take long (00:49:15) • Scientist models might be more capable (00:53:34) • “Reinforcement learning is evil” (01:01:27) • Scientist AI from guardrail to agent (01:08:37) • Can safe AI still be competent? (01:12:38) • How much will this cost? (01:19:29) • Can it generalise beyond maths and science? (01:23:26) • A multi-national push for superintelligence (01:39:19) • Want to work with or fund Yoshua? (01:51:16) • Why smart people ignore AI risk (01:54:45) • Don’t let AI build the next AI (02:01:33) • Why politicians miss the real risks (02:12:28) • Why Yoshua changed his mind about AI risk (02:21:27)
2:35:26

Yoshua Bengio thinks he knows how to make provably safe superintelligent agents. Bengio built the foundations of modern AI and is the most cited living scientist. He believes his alternative training setup would: 1. Guarantee honesty 2. Prevent unintended goals 3. Produce capable agents 4. Port over most data and techniques from current LLMs 5. Not be inherently more expensive, and perhaps be more intelligent Bengio claims the honesty and lack of unintended goals can be proven mathematically, at least given particular assumptions. And his new organization, LawZero, is aiming to build a scrappy prototype as soon as possible. The architecture is called 'Scientist AI' and it's based on training a model to explain empirical observations, including what people say, rather than training AIs that mimic human behaviour or seek our approval. (Bengio's frank assessment is that "reinforcement learning is evil" and that allowing AIs to independently train their successors is "the most crazy, dangerous bet that unfortunately we are on track to do.") But skeptics question whether Scientist AI really does solve the fundamental problem of 'eliciting latent knowledge' from AI models. And with the commercial race for superintelligence so intense, it's not clear whether the proposal will be able to compete or have time to bear fruit, even if it's sound in theory. On The 80,000 Hours Podcast, links below – enjoy! • Making AI honest and safe (00:00:00) • Scientist AI in plain English (00:02:27) • How Scientist AI differs from LLMs (00:06:32) • How the training data works (00:14:02) • Can this become an agent? (00:21:02) • Why Yoshua is now more optimistic (00:32:11) • Why companies can’t stop racing (00:36:35) • A working prototype won't take long (00:49:15) • Scientist models might be more capable (00:53:34) • “Reinforcement learning is evil” (01:01:27) • Scientist AI from guardrail to agent (01:08:37) • Can safe AI still be competent? (01:12:38) • How much will this cost? (01:19:29) • Can it generalise beyond maths and science? (01:23:26) • A multi-national push for superintelligence (01:39:19) • Want to work with or fund Yoshua? (01:51:16) • Why smart people ignore AI risk (01:54:45) • Don’t let AI build the next AI (02:01:33) • Why politicians miss the real risks (02:12:28) • Why Yoshua changed his mind about AI risk (02:21:27)

Rob Wiblin

65,088 views • 3 months ago

When should you start post-training your own models? Fireworks CEO Lin Qiao’s answer: after product-market fit. Not because it's hard... but because only after PMF is the data coming off your product surface worth training on. Lin joined us for our Sequoia Capital "Own Your Intelligence" event to host a workshop on all things post-training; what works, what breaks, and how not to let the model outsmart you. Must listen!! 00:00 Introduction 00:37 What Fireworks sees across thousands of AI applications 02:47 Off-the-shelf APIs and the problem of keeping your taste 03:58 What "owning your intelligence" actually means 05:43 The progression: prompting → RAG → SFT → preferences → RL 07:20 Why this mirrors how humans learn 09:03 Matching the technique to the problem you actually have 10:46 Where teams get stuck: data quality and vibe evals 12:28 Reward hacking: the model that wrote zero lines of code 13:59 Training-to-serving alignment (and why quality drops) 15:55 Post-training in healthcare and security 17:31 From coding to every co-work domain 19:35 Incumbents, cost burden, and not scaling into bankruptcy 21:26 How much control do you want? 23:24 Q&A: What makes a good reward signal 25:00 Q&A: When to start thinking about post-training
28:28

When should you start post-training your own models? Fireworks CEO Lin Qiao’s answer: after product-market fit. Not because it's hard... but because only after PMF is the data coming off your product surface worth training on. Lin joined us for our Sequoia Capital "Own Your Intelligence" event to host a workshop on all things post-training; what works, what breaks, and how not to let the model outsmart you. Must listen!! 00:00 Introduction 00:37 What Fireworks sees across thousands of AI applications 02:47 Off-the-shelf APIs and the problem of keeping your taste 03:58 What "owning your intelligence" actually means 05:43 The progression: prompting → RAG → SFT → preferences → RL 07:20 Why this mirrors how humans learn 09:03 Matching the technique to the problem you actually have 10:46 Where teams get stuck: data quality and vibe evals 12:28 Reward hacking: the model that wrote zero lines of code 13:59 Training-to-serving alignment (and why quality drops) 15:55 Post-training in healthcare and security 17:31 From coding to every co-work domain 19:35 Incumbents, cost burden, and not scaling into bankruptcy 21:26 How much control do you want? 23:24 Q&A: What makes a good reward signal 25:00 Q&A: When to start thinking about post-training

Sonya Huang 🐥

23,604 views • 5 days ago

.River AI has a new vision for personal AI. We co-led River’s $1.1B funding round because we believe in Igor Babuschkin's ambition to build open-weight models that individuals can train and improve continuously. Watch Hemant Taneja in conversation with Igor. Chapters 02:22 — Co-Founding xAI 04:36 — An AI Experience You Own 05:59 — Betting Early on Open Weights 07:47 — How Do You Make Money on Free Models? 09:25 — Why the US Needs Its Own Open Models 11:27 — Enterprises Want to Control Their AI 13:36 — What Is Continual Learning? 16:56 — An API That Updates the Weights 19:17 — A Personal AI for Everyone 20:46 — The 20-Minute Training Run
22:16

.River AI has a new vision for personal AI. We co-led River’s $1.1B funding round because we believe in Igor Babuschkin's ambition to build open-weight models that individuals can train and improve continuously. Watch Hemant Taneja in conversation with Igor. Chapters 02:22 — Co-Founding xAI 04:36 — An AI Experience You Own 05:59 — Betting Early on Open Weights 07:47 — How Do You Make Money on Free Models? 09:25 — Why the US Needs Its Own Open Models 11:27 — Enterprises Want to Control Their AI 13:36 — What Is Continual Learning? 16:56 — An API That Updates the Weights 19:17 — A Personal AI for Everyone 20:46 — The 20-Minute Training Run

General Catalyst

106,552 views • 6 days ago

Anthropic co-founder Jared Kaplan on scaling and the road to human-level AI. At AI Startup School in San Francisco. 00:17 - From Physics to AI 01:41 - Initial Skepticism and Shift to AI 02:12 - AI Training Phases 02:32 - Pre-Training 03:16 - Reinforcement Learning 04:19 - Scaling Laws in Training 08:19 - Unlocking AI Capabilities 11:27 - Organizational Knowledge and Memory 12:19 - Oversight and Nuanced Tasks 13:38 - Preparing for the Future 15:48 - Claude 4 and Beyond 21:18 - Human-AI Collaboration 29:50 - Scaling Laws and Compute Efficiency 35:26 - Audience Q&A
40:47

Anthropic co-founder Jared Kaplan on scaling and the road to human-level AI. At AI Startup School in San Francisco. 00:17 - From Physics to AI 01:41 - Initial Skepticism and Shift to AI 02:12 - AI Training Phases 02:32 - Pre-Training 03:16 - Reinforcement Learning 04:19 - Scaling Laws in Training 08:19 - Unlocking AI Capabilities 11:27 - Organizational Knowledge and Memory 12:19 - Oversight and Nuanced Tasks 13:38 - Preparing for the Future 15:48 - Claude 4 and Beyond 21:18 - Human-AI Collaboration 29:50 - Scaling Laws and Compute Efficiency 35:26 - Audience Q&A

Y Combinator

116,444 views • 1 year ago

In the future, you’ll be able to accomplish a goal by just giving Claude an outcome and a budget. That’s the direction Anthropic is building in with its new Managed Agents features, announced at this week’s Code with Claude developer event. The basic idea: Claude, wrapped in a computer in the cloud, that you can spin up, scale, and manage as needed. Anthropic is taking on the infrastructure that kills most agent products, and making sure that it scales to meet the needs of agents running 24/7. On this week’s AI & I from Every 📧, I talk with Angela Jiang (Angela Jiang), head of product for the Claude platform, and Katelyn Lesse (Katelyn Lesse), head of engineering for the Claude platform, about what Anthropic is building and what it takes to make agents reliable in production. We get into: - Why the "build a generic harness, hot-swap any model behind it" playbook is already outdated. Angela points to eval data on Memory where the same task across different harnesses performed drastically differently. - The infrastructure wall every team hits in production—and why Katelyn thinks “my sandbox died and took the agent with it” is the real reason internal agents don't ship. - Why Anthropic is so bullish on using file systems and skills within Claude, including Angela's argument that those early design choices can compound for years. This is a must-watch for anyone trying to take an agent past the demo and into production. Watch below! Timestamps: How the Claude platform evolved from API to agents: 00:01:48 The primitives that make up Claude Managed Agents: 00:04:09 Why the harness and the model are becoming a single unit: 00:10:37 The infrastructure wall that kills most agent projects in production: 00:18:49 Why team agents need a different shape than individual productivity tools: 00:24:49 How Anthropic's legal team uses an agent to review marketing copy: 00:26:36 Using multi-agent orchestration for advisor strategies, adversarial pairs, and swarms: 00:34:24 How to measure agent success with outcome and budget as the end state: 00:35:50 What the platform looks like a year from now, when Claude writes its own harness: 00:39:11
43:20

In the future, you’ll be able to accomplish a goal by just giving Claude an outcome and a budget. That’s the direction Anthropic is building in with its new Managed Agents features, announced at this week’s Code with Claude developer event. The basic idea: Claude, wrapped in a computer in the cloud, that you can spin up, scale, and manage as needed. Anthropic is taking on the infrastructure that kills most agent products, and making sure that it scales to meet the needs of agents running 24/7. On this week’s AI & I from Every 📧, I talk with Angela Jiang (Angela Jiang), head of product for the Claude platform, and Katelyn Lesse (Katelyn Lesse), head of engineering for the Claude platform, about what Anthropic is building and what it takes to make agents reliable in production. We get into: - Why the "build a generic harness, hot-swap any model behind it" playbook is already outdated. Angela points to eval data on Memory where the same task across different harnesses performed drastically differently. - The infrastructure wall every team hits in production—and why Katelyn thinks “my sandbox died and took the agent with it” is the real reason internal agents don't ship. - Why Anthropic is so bullish on using file systems and skills within Claude, including Angela's argument that those early design choices can compound for years. This is a must-watch for anyone trying to take an agent past the demo and into production. Watch below! Timestamps: How the Claude platform evolved from API to agents: 00:01:48 The primitives that make up Claude Managed Agents: 00:04:09 Why the harness and the model are becoming a single unit: 00:10:37 The infrastructure wall that kills most agent projects in production: 00:18:49 Why team agents need a different shape than individual productivity tools: 00:24:49 How Anthropic's legal team uses an agent to review marketing copy: 00:26:36 Using multi-agent orchestration for advisor strategies, adversarial pairs, and swarms: 00:34:24 How to measure agent success with outcome and budget as the end state: 00:35:50 What the platform looks like a year from now, when Claude writes its own harness: 00:39:11

Dan Shipper 📧

66,339 views • 3 months ago