Episoder
-
Can GenAI allow us to connect our imagination to what we see on our screens? Decart’s Dean Leitersdorf believes it can.
In this episode, Dean Leitersdorf breaks down how Decart is pushing the boundaries of compute in order to create AI-generated consumer experiences, from fully playable video games to immersive worlds. From achieving real-time video inference on existing hardware to building a fully vertically integrated stack, Dean explains why solving fundamental limitations rather than specific problems could lead to the next trillion-dollar company.
Hosted by: Sonya Huang and Shaun Maguire, Sequoia Capital
00:00 Introduction
03:22 About Oasis
05:25 Solving a problem vs overcoming a limitation
08:42 The role of game engines
11:15 How video real-time inference works
14:10 World model vs pixel representation
17:17 Vertical integration
34:20 Building a moat
41:35 The future of consumer entertainment
43:17 Rapid fire questions -
Years before co-founding Glean, Arvind was an early Google employee who helped design the search algorithm. Today, Glean is building search and work assistants inside the enterprise, which is arguably an even harder problem. One of the reasons enterprise search is so difficult is that each individual at the company has different permissions and access to different documents and information, meaning that every search needs to be fully personalized. Solving this difficult ingestion and ranking problem also unlocks a key problem for AI: feeding the right context into LLMs to make them useful for your enterprise context. Arvind and his team are harnessing generative AI to synthesize, make connections, and turbo-change knowledge work. Hear Arvind’s vision for what kind of work we’ll do when work AI assistants reach their potential.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
00:00 - Introduction
08:35 - Search rankings
11:30 - Retrieval-Augmented Generation
15:52 - Where enterprise search meets RAG
19:13 - How is Glean changing work?
26:08 - Agentic reasoning
31:18 - Act 2: application platform
33:36 - Developers building on Glean
35:54 - 5 years into the future
38:48 - Advice for founders -
Manglende episoder?
-
In recent years there’s been an influx of theoretical physicists into the leading AI labs. Do they have unique capabilities suited to studying large models or is it just herd behavior? To find out, we talked to our former AI Fellow (and now OpenAI researcher) Dan Roberts.
Roberts, co-author of The Principles of Deep Learning Theory, is at the forefront of research that applies the tools of theoretical physics to another type of large complex system, deep neural networks. Dan believes that DLLs, and eventually LLMs, are interpretable in the same way a large collection of atoms is—at the system level. He also thinks that emphasis on scaling laws will balance with new ideas and architectures over time as scaling asymptotes economically.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
Mentioned in this episode:
The Principles of Deep Learning Theory: An Effective Theory Approach to Understanding Neural Networks, by Daniel A. Roberts, Sho Yaida, Boris Hanin
Black Holes and the Intelligence Explosion: Extreme scenarios of AI focus on what is logically possible rather than what is physically possible. What does physics have to say about AI risk?
Yang-Mills & The Mass Gap: An unsolved Millennium Prize problem
AI Math Olympiad: Dan is on the prize committee -
NotebookLM from Google Labs has become the breakout viral AI product of the year. The feature that catapulted it to viral fame is Audio Overview, which generates eerily realistic two-host podcast audio from any input you upload—written doc, audio or video file, or even a PDF. But to describe NotebookLM as a “podcast generator” is to vastly undersell it. The real magic of the product is in offering multi-modal dimensions to explore your own content in new ways—with context that’s surprisingly additive. 200-page training manuals become synthesized into digestible chapters, turned into a 10-minute podcast—or both—and shared with the sales team, just to cite one example. Raiza Martin and Jason Speilman join us to discuss how the magic happens, and what’s next for source-grounded AI.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital -
All of us as consumers have felt the magic of ChatGPT—but also the occasional errors and hallucinations that make off-the-shelf language models problematic for business use cases with no tolerance for errors. Case in point: A model deployed to help create a summary for this episode stated that Sridhar Ramaswamy previously led PyTorch at Meta. He did not. He spent years running Google’s ads business and now serves as CEO of Snowflake, which he describes as the data cloud for the AI era.
Ramaswamy discusses how smart systems design helped Snowflake create reliable "talk-to-your-data" applications with over 90% accuracy, compared to around 45% for out-of-the-box solutions using off the shelf LLMs. He describes Snowflake's commitment to making reliable AI simple for their customers, turning complex software engineering projects into straightforward tasks.
Finally, he stresses that even as frontier models progress, there is significant value to be unlocked from current models by applying them more effectively across various domains.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
Mentioned in this episode:
Cortex Analyst: Snowflake’s talk-to-your-data API
Document AI: Snowflake feature that extracts in structured information from documents -
Combining LLMs with AlphaGo-style deep reinforcement learning has been a holy grail for many leading AI labs, and with o1 (aka Strawberry) we are seeing the most general merging of the two modes to date. o1 is admittedly better at math than essay writing, but it has already achieved SOTA on a number of math, coding and reasoning benchmarks.
Deep RL legend and now OpenAI researcher Noam Brown and teammates Ilge Akkaya and Hunter Lightman discuss the ah-ha moments on the way to the release of o1, how it uses chains of thought and backtracking to think through problems, the discovery of strong test-time compute scaling laws and what to expect as the model gets better.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
Mentioned in this episode:
Learning to Reason with LLMs: Technical report accompanying the launch of OpenAI o1.
Generator verifier gap: Concept Noam explains in terms of what kinds of problems benefit from more inference-time compute.
Agent57: Outperforming the human Atari benchmark, 2020 paper where DeepMind demonstrated “the first deep reinforcement learning agent to obtain a score that is above the human baseline on all 57 Atari 2600 games.”
Move 37: Pivotal move in AlphaGo’s second game against Lee Sedol where it made a move so surprising that Sedol thought it must be a mistake, and only later discovered he had lost the game to a superhuman move.
IOI competition: OpenAI entered o1 into the International Olympiad in Informatics and received a Silver Medal.
System 1, System 2: The thesis if Danial Khaneman’s pivotal book of behavioral economics, Thinking, Fast and Slow, that positied two distinct modes of thought, with System 1 being fast and instinctive and System 2 being slow and rational.
AlphaZero: The predecessor to AlphaGo which learned a variety of games completely from scratch through self-play. Interestingly, self-play doesn’t seem to have a role in o1.
Solving Rubik’s Cube with a robot hand: Early OpenAI robotics paper that Ilge Akkaya worked on.
The Last Question: Science fiction story by Isaac Asimov with interesting parallels to scaling inference-time compute.
Strawberry: Why?
O1-mini: A smaller, more efficient version of 1 for applications that require reasoning without broad world knowledge.
00:00 - Introduction
01:33 - Conviction in o1
04:24 - How o1 works
05:04 - What is reasoning?
07:02 - Lessons from gameplay
09:14 - Generation vs verification
10:31 - What is surprising about o1 so far
11:37 - The trough of disillusionment
14:03 - Applying deep RL
14:45 - o1’s AlphaGo moment?
17:38 - A-ha moments
21:10 - Why is o1 good at STEM?
24:10 - Capabilities vs usefulness
25:29 - Defining AGI
26:13 - The importance of reasoning
28:39 - Chain of thought
30:41 - Implication of inference-time scaling laws
35:10 - Bottlenecks to scaling test-time compute
38:46 - Biggest misunderstanding about o1?
41:13 - o1-mini
42:15 - How should founders think about o1? -
Adding code to LLM training data is a known method of improving a model’s reasoning skills. But wouldn’t math, the basis of all reasoning, be even better? Up until recently, there just wasn’t enough usable data that describes mathematics to make this feasible.
A few years ago, Vlad Tenev (also founder of Robinhood) and Tudor Achim noticed the rise of the community around an esoteric programming language called Lean that was gaining traction among mathematicians. The combination of that and the past decade’s rise of autoregressive models capable of fast, flexible learning made them think the time was now and they founded Harmonic. Their mission is both lofty—mathematical superintelligence—and imminently practical, verifying all safety-critical software.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
Mentioned in this episode:
IMO and the Millennium Prize: Two significant global competitions Harmonic hopes to win (soon)
Riemann hypothesis: One of the most difficult unsolved math conjectures (and a Millenium Prize problem) most recently in the sights of MIT mathematician Larry Guth
Terry Tao: perhaps the greatest living mathematician and Vlad’s professor at UCLA
Lean: an open source functional language for code verification launched by Leonardo de Moura when at Microsoft Research in 2013 that powers the Lean Theorem Prover
mathlib: the largest math textbook in the world, all written in Lean
Metaculus: online prediction platform that tracks and scores thousands of forecasters
Minecraft Beaten in 20 Seconds: The video Vlad references as an analogy to AI math
Navier-Stokes equations: another important Millenium Prize math problem. Vlad considers this more tractable that Riemann
John von Neumann: Hungarian mathematician and polymath that made foundational contributions to computing, the Manhattan Project and game theory
Gottfried Wilhelm Leibniz: co-inventor of calculus and (remarkably) creator of the “universal characteristic,” a system for reasoning through a language of symbols and calculations—anticipating Lean and Harmonic by 350 years!
00:00 - Introduction
01:42 - Math is reasoning
06:16 - Studying with the world's greatest living mathematician
10:18 - What does the math community think of AI math?
15:11 - Recursive self-improvement
18:31 - What is Lean?
21:05 - Why now?
22:46 - Synthetic data is the fuel for the model
27:29 - How fast will your model get better?
29:45 - Exploring the frontiers of human knowledge
34:11 - Lightning round -
AI researcher Jim Fan has had a charmed career. He was OpenAI’s first intern before he did his PhD at Stanford with “godmother of AI,” Fei-Fei Li. He graduated into a research scientist position at Nvidia and now leads its Embodied AI “GEAR” group. The lab’s current work spans foundation models for humanoid robots to agents for virtual worlds.
Jim describes a three-pronged data strategy for robotics, combining internet-scale data, simulation data and real world robot data. He believes that in the next few years it will be possible to create a “foundation agent” that can generalize across skills, embodiments and realities—both physical and virtual. He also supports Jensen Huang’s idea that “Everything that moves will eventually be autonomous.”
Hosted by: Stephanie Zhan and Sonya Huang, Sequoia Capital
Mentioned in this episode:
World of Bits: Early OpenAI project Jim worked on as an intern with Andrej Karpathy. Part of a bigger initiative called Universe
Fei-Fei Li: Jim’s PhD advisor at Stanford who founded the ImageNet project in 2010 that revolutionized the field of visual recognition, led the Stanford Vision Lab and just launched her own AI startup, World Labs
Project GR00T: Nvidia’s “moonshot effort” at a robotic foundation model, premiered at this year’s GTC
Thinking Fast and Slow: Influential book by Daniel Kahneman that popularized some of his teaching from behavioral economics
Jetson Orin chip: The dedicated series of edge computing chips Nvidia is developing to power Project GR00T
Eureka: Project by Jim’s team that trained a five finger robot hand to do pen spinning
MineDojo: A project Jim did when he first got to Nvidia that developed a platform for general purpose agents in the game of Minecraft. Won NeurIPS 2022 Outstanding Paper Award
ADI: artificial dog intelligence
Mamba: Selective State Space Models, an alternative architecture to Transformers that Jim is interested in (original paper here)
00:00 Introduction
01:35 Jim’s journey to embodied intelligence
04:53 The GEAR Group
07:32 Three kinds of data for robotics
10:32 A GPT-3 moment for robotics
16:05 Choosing the humanoid robot form factor
19:37 Specialized generalists
21:59 GR00T gets its own chip
23:35 Eureka and Issac Sim
25:23 Why now for robotics?
28:53 Exploring virtual worlds
36:28 Implications for games
39:13 Is the virtual world in service of the physical world?
42:10 Alternative architectures to Transformers
44:15 Lightning round -
There’s a new archetype in Silicon Valley, the AI researcher turned founder. Instead of tinkering in a garage they write papers that earn them the right to collaborate with cutting-edge labs until they break out and start their own.
This is the story of wunderkind Eric Steinberger, the founder and CEO of Magic.dev. Eric came to programming through his obsession with AI and caught the attention of DeepMind researchers as a high school student. In 2022 he realized that AGI was closer than he had previously thought and started Magic to automate the software engineering necessary to get there. Among his counterintuitive ideas are the need to train proprietary large models, that value will not accrue in the application layer and that the best agents will manage themselves. Eric also talks about Magic’s recent 100M token context window model and the HashHop eval they’re open sourcing.
Hosted by: Sonya Huang, Sequoia Capital
Mentioned in this episode:
David Silver: DeepMind researcher that led the AlphaGo team
Johannes Heinrich: a PhD student of Silver’s and DeepMind researcher who mentored Eric as a highschooler
Reinforcement Learning from Self-Play in Imperfect-Information Games: Johannes’s dissertation that inspired Eric
Noam Brown: DeepMind, Meta and now OpenAI reinforcement learning researcher who eventually collaborated with Eric and brought him to FAIR
ClimateScience: NGO that Eric co-founded in 2019 while a university student
Noam Shazeer: One of the original Transformers researchers at Google and founder of Charater.ai
DeepStack: Expert-Level Artificial Intelligence in Heads-Up No-Limit Poker: the first AI paper Eric ever tried to deeply understand
LTM-2-mini: Magic’s first 100M token context model, build using the HashHop eval (now available open source)
00:00 - Introduction
01:39 - Vienna-born wunderkind
04:56 - Working with Noam Brown
8:00 - “I can do two things. I cannot do three.”
10:37 - AGI to-do list
13:27 - Advice for young researchers
20:35 - Reading every paper voraciously
23:06 - The army of Noams
26:46 - The leaps still needed in research
29:59 - What is Magic?
36:12 - Competing against the 800-pound gorillas
38:21 - Ideal team size for researchers
40:10 - AI that feels like a colleague
44:30 - Lightning round
47:50 - Bonus round: 200M token context announcement -
On Training Data, we learn from innovators pushing forward the frontier of AI’s capabilities. Today we’re bringing you something different. It’s the story of a company currently implementing AI at scale in the enterprise, and how it was built from a bootstrapped idea in the pre-AI era to a 150 billion dollar market cap giant.
It’s the Season 2 premiere of Sequoia’s other podcast, Crucible Moments, where we hear from the founders and leaders of some legendary companies about the crossroads and inflection points that shaped their journeys. In this episode, you’ll hear from Fred Luddy and Frank Slootman about building and scaling ServiceNow. Listen to Crucible Moments wherever you get your podcasts or go to:
Spotify: https://open.spotify.com/show/40bWCUSan0boCn0GZJNpPn
Apple: https://podcasts.apple.com/us/podcast/crucible-moments/id1705282398
Hosted by: Roelof Botha, Sequoia Capital
Transcript: https://www.sequoiacap.com/podcast/crucible-moments-servicenow/ -
Customer service is hands down the first killer app of generative AI for businesses. The reasons are simple: the costs of existing solutions are so high, the satisfaction so low and the margin for ROI so wide. But trusting your interactions with customers to hallucination-prone LLMs can be daunting.
Enter Sierra. Co-founder Clay Bavor walks us through the sophisticated engineering challenges his team solved along the way to delivering AI agents for all aspects of the customer experience that are delightful, safe and reliable—and being deployed widely by Sierra’s customers. The Company’s AgentOS enables businesses to create branded AI agents to interact with customers, follow nuanced policies and even handle customer retention and upsell. Clay describes how companies can capture their brand voice, values and internal processes to create AI agents that truly represent the business.
Hosted by: Ravi Gupta and Pat Grady, Sequoia Capital
Mentioned in this episode:
Bret Taylor: co-founder of Sierra
Towards a Human-like Open-Domain Chatbot: 2020 Google paper that introduced Meena, a predecessor of ChatGPT (followed by LaMDA in 2021)
PaLM: Scaling Language Modeling with Pathways: 2022 Google paper about their unreleased 540B parameter transformer model (GPT-3, at the time, had 175B)
Avocado chair: Images generated by OpenAI’s DALL·E model in 2022
Large Language Models Understand and Can be Enhanced by Emotional Stimuli: 2023 Microsoft paper on how models like GPT-4 can be manipulated into providing better results
𝛕-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains: 2024 paper authored by Sierra research team, led by Karthik Narasimhan (co-author of the 2022 ReACT paper and the 2023 Reflexion paper)
00:00:00 Introduction
00:01:21 Clay’s background
00:03:20 Google before the ChatGPT moment
00:07:31 What is Sierra?
00:12:03 What’s possible now that wasn’t possible 18 months ago?
00:17:11 AgentOS
00:23:45 The solution to many problems with AI is more AI
00:28:37 𝛕-bench
00:33:19 Engineering task vs research task
00:37:27 What tasks can you trust an agent with now?
00:43:21 What metrics will move?
00:46:22 The reality of deploying AI to customers today
00:53:33 The experience manager
01:03:54 Outcome-based pricing
01:05:55 Lightning Round -
After AlphaGo beat Lee Sedol, a young mechanical engineer at Google thought of another game reinforcement learning could win: energy optimization at data centers. Jim Gao convinced his bosses at the Google data center team to let him work with the DeepMind team to try. The initial pilot resulted in a 40% energy savings and led he and his co-founders to start Phaidra to turn this technology into a product.
Jim discusses the challenges of AI readiness in industrial settings and how we have to build on top of the control systems of the 70s and 80s to achieve the promise of the Fourth Industrial Revolution. He believes this new world of self-learning systems and self-improving infrastructure is a key factor in addressing global climate change.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
Mentioned in this episode:
Mustafa Suleyman: Co-founder of DeepMind and Inflection AI and currently CEO of Microsoft AI, known to his friends as “Moose”
Joe Kava: Google VP of data centers who Jim sent his initial email to pitching the idea that would eventually become Phaidra
Constrained optimization: the class of problem that reinforcement learning can be applied to in real world systems
Vedavyas Panneershelvam: co-founder and CTO of Phaidra; one of the original engineers on the AlphaGo project
Katie Hoffman: co-founder, President and COO of Phaidra
Demis Hassabis: CEO of DeepMind -
In the first wave of the generative AI revolution, startups and enterprises built on top of the best closed-source models available, mostly from OpenAI. The AI customer journey moves from training to inference, and as these first products find PMF, many are hitting a wall on latency and cost.
Fireworks Founder and CEO Lin Qiao led the PyTorch team at Meta that rebuilt the whole stack to meet the complex needs of the world’s largest B2C company. Meta moved PyTorch to its own non-profit foundation in 2022 and Lin started Fireworks with the mission to compress the timeframe of training and inference and democratize access to GenAI beyond the hyperscalers to let a diversity of AI applications thrive.
Lin predicts when open and closed source models will converge and reveals her goal to build simple API access to the totality of knowledge.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
Mentioned in this episode:
Pytorch: the leading framework for building deep learning models, originated at Meta and now part of the Linux Foundation umbrella
Caffe2 and ONNX: ML frameworks Meta used that PyTorch eventually replaced
Conservation of complexity: the idea that that every computer application has inherent complexity that cannot be reduced but merely moved between the backend and frontend, originated by Xerox PARC researcher Larry Tesler
Mixture of Experts: a class of transformer models that route requests between different subsets of a model based on use case
Fathom: a product the Fireworks team uses for video conference summarization
LMSYS Chatbot Arena: crowdsourced open platform for LLM evals hosted on Hugging Face
00:00 - Introduction
02:01 - What is Fireworks?
02:48 - Leading Pytorch
05:01 - What do researchers like about PyTorch?
07:50 - How Fireworks compares to open source
10:38 - Simplicity scales
12:51 - From training to inference
17:46 - Will open and closed source converge?
22:18 - Can you match OpenAI on the Fireworks stack?
26:53 - What is your vision for the Fireworks platform?
31:17 - Competition for Nvidia?
32:47 - Are returns to scale starting to slow down?
34:28 - Competition
36:32 - Lightning round -
GithHub invented collaborative coding and in the process changed how open source projects, startups and eventually enterprises write code. GitHub Copilot is the first blockbuster product built on top of OpenAI’s GPT models. It now accounts for more than 40 percent of GitHub revenue growth for an annual revenue run rate of $2 billion. Copilot itself is already a larger business than all of GitHub was when Microsoft acquired it in 2018.
We talk to CEO Thomas Dohmke about how a small team at GitHub built on top of GPT-3 and quickly created a product that developers love—and can’t live without. Thomas describes how the product has grown from simple autocomplete to a fully featured workspace for enterprise teams. He also believes that tools like Copilot will bring the power of coding to a billion developers by 2030.
Hosted by: Stephanie Zhan and Sonya Huang, Sequoia Capital
Mentioned in this episode:
Nat Friedman: Former Microsoft VP (and now investor) who came up with the idea that Microsoft should buy GitHub
Oege de Moor: Github developer (and now founder of XBOW) who came up with the idea of using GPT-3 for code and went on to create Copilot
Alex Graveley: principal engineer and Chief Architect for Copilot (now CEO of Minion.ai) who came up with the name Copilot (because his boss, Nat Firedman, is an amateur pilot)
Productivity Assessment of Neural Code Completion: Original GitHub research paper on the impact of Copilot on Developer productivity
Escaping a room in Minecraft with an AI-powered NPC: Recent Minecraft AI assistant demo from Microsoft
With AI, anyone can be a coder now: TED2024 talk by Thomas Dohmke
JFrog: The software supply chain platform that GitHub just partnered with
00:00:00 - Introduction
00:01:18 - Getting started with code
00:03:43 - Microsoft’s acquisition of GitHub
00:11:40 - Evolving Copilot beyond autocomplete
00:14:18 - In hindsight, you can always move faster
00:15:56 - Building on top of OpenAI
00:20:21 - The latest metrics
00:22:11 - The surprise of Copilot’s impact
00:25:11 - Teaching kids to code in the age of Copilot
00:26:38 - The momentum mindset
00:29:46 - Agents vs Copilots
00:32:06 - The Roadmap
00:37:31 - Making maintaining software easier
00:38:48 - The creative new world
00:42:38 - The AI 10x software engineer
00:45:12 - Creativity and systems engineering in AI
00:48:55 - What about COBOL?
00:50:23 - Will GitHub build its own models?
00:57:19 - Rapid incubation at GitHub Next
00:59:21 - The future of AI?
01:03:18 - Advice for founders
01:05:08 - Lightning round -
As head of Product Management for Generative AI at Meta, Joe Spisak leads the team behind Llama, which just released the new 3.1 405B model. We spoke with Joe just two days after the model’s release to ask what’s new, what it enables, and how Meta sees the role of open source in the AI ecosystem.
Joe shares that where Llama 3.1 405B really focused is on pushing scale (it was trained on 15 trillion tokens using 16,000 GPUs) and he’s excited about the zero-shot tool use it will enable, as well as its role in distillation and generating synthetic data to teach smaller models. He tells us why he thinks even frontier models will ultimately commoditize—and why that’s a good thing for the startup ecosystem.
Hosted by: Stephanie Zhan and Sonya Huang, Sequoia Capital
Mentioned in this episode:
Llama 3.1 405B paper
Open Source AI Is the Way Forward: Mark Zuckerberg essay released with Llama 3.1.
Mistral Large 2
The Bitter Lesson by Rich Sutton
00:00 Introduction
01:28 The Llama 3.1 405B launch
05:02 The open source license
07:01 What's in it for Meta?
10:19 Why not open source?
11:16 Will frontier models commoditize?
12:41 What about startups?
16:29 The Mistral team
19:36 Are all frontier strategies comparable?
22:38 Is model development becoming more like software development?
26:34 Agentic reasoning
29:09 What future levers will unlock reasoning?
31:20 Will coding and math lead to unlocks?
33:09 Small models
34:08 7X more data
37:36 Are we going to hit a wall?
39:49 Lightning round -
In February, Sebastian Siemiatkowski boldly announced that Klarna’s new OpenAI-powered assistant handled two thirds of the Swedish fintech’s customer service chats in its first month. Not only were customer satisfaction metrics better, but by replacing 700 full-time contractors the bottom line impact is projected to be $40M. Since then, every company we talk to wants to know, “How do we get the Klarna customer support thing?”
Co-founder and CEO Sebastian Siemiatkowski tells us how the Klarna team shipped this new product in record time—and how embracing AI internally with an experimental mindset is transforming the company. He discusses how AI development is proliferating inside the company, from customer support to marketing to internal knowledge to customer-facing experiences.
Sebastian also reflects on the impacts of AI on employment, society, and the arts while encouraging lawmakers to be open minded about the benefits.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
Mentioned in this episode:
DeepL: Language translation app that Sebastian says makes 10,000 translators in Brussels redundant
The Klarna brand: The offbeat optimism that the company is now augmenting with AI
Neo4j: The graph database management system that Klarna is using to build Kiki, their internal knowledge base
00:00 Introduction
01:57 Klarna’s business
03:00 Pitching OpenAI
08:51 How we built this
10:46 Will Klara ever completely replace its CS team with AI?
14:22 The benefits
17:25 If you had a policy magic wand…
21:12 What jobs will be most affected by AI?
23:58 How about marketing?
27:55 How creative are LLMs?
30:11 Klarna’s knowledge graph, Kiki
33:10 Reducing the number of enterprise systems
35:24 Build vs buy?
39:59 What’s next for Klarna with AI?
48:48 Lightning round -
LLMs are democratizing digital intelligence, but we’re all waiting for AI agents to take this to the next level by planning tasks and executing actions to actually transform the way we work and live our lives.
Yet despite incredible hype around AI agents, we’re still far from that “tipping point” with best in class models today. As one measure: coding agents are now scoring in the high-teens % on the SWE-bench benchmark for resolving GitHub issues, which far exceeds the previous unassisted baseline of 2% and the assisted baseline of 5%, but we’ve still got a long way to go.
Why is that? What do we need to truly unlock agentic capability for LLMs? What can we learn from researchers who have built both the most powerful agents in the world, like AlphaGo, and the most powerful LLMs in the world?
To find out, we’re talking to Misha Laskin, former research scientist at DeepMind. Misha is embarking on his vision to build the best agent models by bringing the search capabilities of RL together with LLMs at his new company, Reflection AI. He and his cofounder Ioannis Antonoglou, co-creator of AlphaGo and AlphaZero and RLHF lead for Gemini, are leveraging their unique insights to train the most reliable models for developers building agentic workflows.
Hosted by: Stephanie Zhan and Sonya Huang, Sequoia Capital
00:00 Introduction
01:11 Leaving Russia, discovering science
10:01 Getting into AI with Ioannis Antonoglou
15:54 Reflection AI and agents
25:41 The current state of Ai agents
29:17 AlphaGo, AlphaZero and Gemini
32:58 LLMs don’t have a ground truth reward
37:53 The importance of post-training
44:12 Task categories for agents
45:54 Attracting talent
50:52 How far away are capable agents?
56:01 Lightning round
Mentioned:
The Feynman Lectures on Physics: The classic text that got Misha interested in science.
Mastering the game of Go with deep neural networks and tree search: The original 2016 AlphaGo paper.
Mastering the game of Go without human knowledge: 2017 AlphaGo Zero paper
Scaling Laws for Reward Model Overoptimization: OpenAI paper on how reward models can be gamed at all scales for all algorithms.
Mapping the Mind of a Large Language Model: Article about Anthropic mechanistic interpretability paper that identifies how millions of concepts are represented inside Claude Sonnet
Pieter Abeel: Berkeley professor and founder of Covariant who Misha studied with
A2C and A3C: Advantage Actor Critic and Asynchronous Advantage Actor Critic, the two algorithms developed by Misha’s manager at DeepMind, Volodymyr Mnih, that defined reinforcement learning and deep reinforcement learning -
The current LLM era is the result of scaling the size of models in successive waves (and the compute to train them). It is also the result of better-than-Moore’s-Law price vs performance ratios in each new generation of Nvidia GPUs. The largest platform companies are continuing to invest in scaling as the prime driver of AI innovation.
Are they right, or will marginal returns level off soon, leaving hyperscalers with too much hardware and too few customer use cases? To find out, we talk to Microsoft CTO Kevin Scott who has led their AI strategy for the past seven years. Scott describes himself as a “short-term pessimist, long-term optimist” and he sees the scaling trend as durable for the industry and critical for the establishment of Microsoft’s AI platform.
Scott believes there will be a shift across the compute ecosystem from training to inference as the frontier models continue to improve, serving wider and more reliable use cases. He also discusses the coming business models for training data, and even what ad units might look like for autonomous agents.
Hosted by: Pat Grady and Bill Coughran, Sequoia Capital
Mentioned:
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, the 2018 Google paper that convinced Kevin that Microsoft wasn’t moving fast enough on AI.
Dennard scaling: The scaling law that describes the proportional relationship between transistor size and power use; has not held since 2012 and is often confused with Moore’s Law.
Textbooks Are All You Need: Microsoft paper that introduces a new large language model for code, phi-1, that achieves smaller size by using higher quality “textbook” data.
GPQA and MMLU: Benchmarks for reasoning
Copilot: Microsoft product line of GPT consumer assistants from general productivity to design, vacation planning, cooking and fitness.
Devin: Autonomous AI code agent from Cognition Labs that Microsoft recently announced a partnership with.
Ray Solomonoff: Participant in the 1956 Dartmouth Summer Research Project on Artificial Intelligence that named the field; Kevin admires his prescience about the importance of probabilistic methods decades before anyone else.
00:00 - Introduction
01:20 - Kevin’s backstory
06:56 - The role of PhDs in AI engineering
09:56 - Microsoft’s AI strategy
12:40 - Highlights and lowlights
16:28 - Accelerating investments
18:38 - The OpenAI partnership
22:46 - Soon inference will dwarf training
27:56 - Will the demand/supply balance change?
30:51 - Business models for data
36:54 - The value function
39:58 - Copilots
44:47 - The 98/2 rule
49:34 - Solving zero-sum games
57:13 - Lightning round -
As impressive as LLMs are, the growing consensus is that language, scale and compute won’t get us to AGI. Although many AI benchmarks have quickly achieved human-level performance, there is one eval that has barely budged since it was created in 2019.
Google researcher François Chollet wrote a paper that year defining intelligence as skill-acquisition efficiency—the ability to learn new skills as humans do, from a small number of examples. To make it testable he proposed a new benchmark, the Abstraction and Reasoning Corpus (ARC), designed to be easy for humans, but hard for AI. Notably, it doesn’t rely on language.
Zapier co-founder Mike Knoop read Chollet’s paper as the LLM wave was rising. He worked quickly to integrate generative AI into Zapier’s product, but kept coming back to the lack of progress on the ARC benchmark. In June, Knoop and Chollet launched the ARC Prize, a public competition offering more than $1M to beat and open-source a solution to the ARC-AGI eval.
In this episode Mike talks about the new ideas required to solve ARC, shares updates from the first two weeks of the competition, and shares why he’s excited for AGI systems that can innovate alongside humans.
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
Mentioned:
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models: The 2019 paper that first caught Mike’s attention about the capabilities of LLMs
On the Measure of Intelligence: 2019 paper by Google researcher François Chollet that introduced the ARC benchmark, which remains unbeaten
ARC Prize 2024: The $1M+ competition Mike and François have launched to drive interest in solving the ARC-AGI eval
Sequence to Sequence Learning with Neural Networks: Ilya Sutskever paper from 2014 that influenced the direction of machine translation with deep neural networks.
Etched: Luke Miles on LessWrong wrote about the first ASIC chip that accelerates transformers on silicon
Kaggle: The leading data science competition platform and online community, acquired by Google in 2017
Lab42: Swiss AU lab that hosted ARCathon precursor to ARC Prize
Jack Cole: Researcher on team that was #1 on the leaderboard for ARCathon
Ryan Greenblatt: Researcher with current high score (50%) on ARC public leaderboard
(00:00) Introduction
(01:51) AI at Zapier
(08:31) What is ARC AGI?
(13:25) What does it mean to efficiently acquire a new skill?
(19:03) What approaches will succeed?
(21:11) A little bit of a different shape
(25:59) The role of code generation and program synthesis
(29:11) What types of people are working on this?
(31:45) Trying to prove you wrong
(34:50) Where are the big labs?
(38:21) The world post-AGI
(42:51) When will we cross 85% on ARC AGI?
(46:12) Will LLMs be part of the solution?
(50:13) Lightning round -
Archimedes said that with a large enough lever, you can move the world. For decades, software engineering has been that lever. And now, AI is compounding that lever. How will we use AI to apply 100 or 1000x leverage to the greatest lever to move the world?
Matan Grinberg and Eno Reyes, co-founders of Factory, have chosen to do things differently than many of their peers in this white-hot space. They sell a fleet of “Droids,” purpose-built dev agents which accomplish different tasks in the software development lifecycle (like code review, testing, pull requests or writing code). Rather than training their own foundation model, their approach is to build something useful for engineering orgs today on top of the rapidly improving models, aligning with the developer and evolving with them.
Matan and Eno are optimistic about the effects of autonomy in software development and on building a company in the application layer. Their advice to founders, “The only way you can win is by executing faster and being more obsessed.”
Hosted by: Sonya Huang and Pat Grady, Sequoia Capital
Mentioned:
Juan Maldacena, Institute for Advanced Study, string theorist that Matan cold called as an undergrad
SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering, small-model open-source software engineering agent
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?, an evaluation framework for GitHub issues
Monte Carlo tree search, a 2006 algorithm for solving decision making in games (and used in AlphaGo)
Language agent tree search, a framework for LLM planning, acting and reasoning
The Bitter Lesson, Rich Sutton’s essay on scaling in search and learning
Code churn, time to merge, cycle time, metrics Factory thinks are important to eng orgs
Transcript: https://www.sequoiacap.com/podcast/training-data-factory/
00:00 Introduction
01:36 Personal backgrounds
10:54 The compound lever
12:41 What is Factory?
16:29 Cognitive architectures
21:13 800 engineers at OpenAI are working on my margins
24:00 Jeff Dean doesn't understand your code base
25:40 Individual dev productivity vs system-wide optimization
30:04 Results: Factory in action
32:54 Learnings along the way
35:36 Fully autonomous Jeff Deans
37:56 Beacons of the upcoming age
40:04 How far are we?
43:02 Competition
45:32 Lightning round
49:34 Bonus round: Factory's SWE-bench results - Vis mere
