Jun 24, 2024 · 58m · news
David Luan: Why Nvidia Will Enter the Model Space & Models Will Enter the Chip Space | E1169 · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this comprehensive interview, Adept co-founder David Luan joins host Harry Stebbings to analyze the structural evolution of AI research, the geopolitical and economic battles between chipmakers and cloud giants, and the paradigm shift from basic chatbots to highly integrated, workflow-automating AI agents.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 20.9% of the talking time here. How this is scored →
Harry as informed peer 4.6 Guest teaching 3.5 Guest disagreement 1.7 Harry pushing back 3.6
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 53:44 Rejection of prompting as HCI
David forcefully pushes back when Harry reduces human-computer interaction to basic prompting, arguing that text prompting is a fundamentally flawed way to interact with intelligent systems.
Hardest push from Harry ▶ 35:53 Skepticism on agent novelty vs RPAHarry aggressively presses David on whether AI agents are genuinely novel or just a repackaging of legacy RPA concepts pioneered by companies like UiPath.
Biggest teaching moment ▶ 36:14 Yellow lines vs Full Self-Driving analogyDavid educates Harry on the structural difference between RPA and agents using a vivid comparison of factory floor yellow-line robots versus Full Self-Driving vehicles.
Harry holds his own ▶ 46:27 Defending the AI services thesisHarry demonstrates strong market expertise by defending his published thesis that AI services companies will out-earn foundation model providers, referencing feedback from industry figures.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| David Luan's Early Career and Takeaways from Google Brain | 2 | 2 | 0 | 1 | Harry sets an open tone asking about David's early career at Google Brain and follows up with a brief clarifying question regarding bottom-up research. David warmly praises Google Brain as the Bell Labs of its era and explains the curiosity-driven research culture. | |
| The Transformer Breakthrough and the Path to ChatGPT | 2 | 3 | 0 | 1 | Harry self-deprecatingly asks why there was a multi-year lag between the 2017 Transformer breakthrough and ChatGPT's consumer moment. David explains how model capabilities needed to hit a minimum threshold alongside consumer-friendly UX packaging. | |
| Key Takeaways from OpenAI and the Transition to Problem-Solving Culture | 6 | 4 | 2 | 5 | Harry challenges David by citing an interview with an unnamed top AI leader claiming OpenAI faced diminishing returns from compute. David politely counters by explaining logarithmic compute scaling and the pivot toward synthetic RL data loops. | |
| The S-Curve of Model Scaling and the Speciation of Agents vs. Chatbots | 5 | 3 | 1 | 2 | Harry synthesizes David's point to infer that agents lag because human work processes are not neatly codified in web data. David validates Harry's observation and articulates how agents and chatbots are becoming distinct technological species. | |
| Emergent AI Capabilities and the Concept of Minimum Viable Scale | 1 | 4 | 0 | 0 | Harry candidly admits he did not understand 'minimum viable capabilities' when David mentioned it off-air and invites him to explain. David breaks down how scaling yields emergent unpredictable abilities using arithmetic in GPT-2 as an example. | |
| Solving AI Reasoning and the Future Landscape of Model Providers | 5 | 3 | 2 | 4 | Harry questions whether reasoning breakthroughs prevent the widely predicted commoditization of foundation models. David explains that reasoning requires model-level breakthroughs and predicts only 5 to 7 major model providers will survive. | |
| The Challenge of AI Memory and LLMs as Software Components | 4 | 4 | 1 | 3 | Harry asks why memory remains difficult for AI given that traditional computers have memory. David distinguishes between working context windows and long-term user preference memory, emphasizing that LLMs are system components rather than standalone products. | |
| Chip Dominance, Google TPU, and the Apple Edge Advantage | 6 | 3 | 2 | 5 | Harry pushes back on the idea that cloud providers can easily commoditize Nvidia, pointing to Nvidia's extreme hardware sophistication. David agrees it is difficult but cites Google's TPU as evidence of successful in-house chip development. | |
| Apple's Partnership with OpenAI and Foundational Model Independence | 5 | 3 | 1 | 3 | Harry voices surprise at Apple's non-exclusive partnership with OpenAI, framing it as potential bad news for OpenAI's leverage. David reframes the deal as Apple smartly securing interface ownership while keeping foundation models hot-swappable. | |
| Adept's Vertical Strategy and the Shift to AI Workflows | 4 | 3 | 2 | 4 | Harry directly asks David whether Adept is actually a foundation model company. David clarifies that Adept is vertically integrated around workflow agents rather than selling raw model APIs to third-party developers. | |
| Traditional RPA vs. The New Era of AI Agents | 5 | 5 | 2 | 5 | Harry challenges David by asking if AI agents are simply re-badged Robotic Process Automation (RPA), citing UiPath founder Daniel Dines. David uses a vivid analogy comparing RPA to factory floor yellow lines and agents to Full Self-Driving. | |
| AI Business Models: Price per Seat vs. Price per Work | 6 | 3 | 3 | 5 | Harry challenges the popular idea that AI will eliminate seat-based pricing and quotes VC Miles Grimshaw's argument that 'co-pilots' are merely defensive incumbent strategies. David resists both premises, arguing knowledge work co-pilots augment human creativity. | |
| Organization Structures and Collapsing the Talent Stack | 6 | 4 | 2 | 4 | Harry cites Meta CMO Alex Schultz's warning that AI adoption could hit a prolonged plateau similar to autonomous driving. David counters by arguing that AI benefits from rapid scientific step-changes rather than endless edge-case whack-a-mole. | |
| AI Services Companies vs. Repeatable AI Products | 7 | 3 | 2 | 5 | Harry defends his viral prediction that AI implementation services will generate more revenue than model providers, despite peer criticism. David contends that repeatable software products will rapidly productize services and capture the majority of economic value. | |
| Human-Computer Interaction (HCI) and the Path to AGI | 4 | 5 | 4 | 5 | Harry bluntly asks if human-computer interaction (HCI) in AI is 'just basic prompting'. David strongly rejects this reductionist view, explaining that human collaboration relies on shared canvases and rich interactive workspaces rather than text prompts. | |
| Quick Fire Round: Speciation, Misconceptions, and Agents | 5 | 4 | 3 | 5 | Harry closes the quick-fire round by questioning the market size ceiling for workflow automation, pointing out UiPath's $6-7B valuation. David responds that RPA covers a tiny fraction of workflow tasks compared to the massive addressable market for intelligent agents. |
Statements from this episode (39)
Luan: OpenAI and DeepMind pivoted AI research strategy before competitors
“What OpenAI realized before basically everybody but DeepMind was that the next phase of AI after Transformer was not going to be about research paper writing. It was going to be about let's choose a major unsolved scientific problem and just try to solve it.”
Luan: AI scaling will not suffer diminishing returns on compute
“The second way of improving model performance is just starting to be tapped now, and that's also going to absorb a boatload of compute. So because of that, I actually am not worried about diminishing returns to the compute over time.”
Luan: AI success is existential for tier-one cloud providers
“I think every tier one cloud provider existentially needs to win here.”
Luan: Google Brain was the Bell Labs of the 2012–2018 AI era
“That, like, 2012 to 20 18 or so era, Google Brain was just, like, incredibly dominant. They did an amazing job picking talent. Like, the people who invented Transformer, the people who invented the diffusion model, people who did all of these new optimization …”
Luan: 2012-2018 AI breakthroughs came from bottom-up curiosity without corporate objectives
“During that twenty-twelve, twenty-eighteen era, the way people made progress what I mean by bottom-up basic research is you hire the most brilliant scientists, they come to work every day with, like, No near-term objective they're being held accountable to. An…”
Luan: ChatGPT was just GPT-3 with instruction tuning released a year later
“ChatGPT was really just GPT-III with instruction. It was basically like more chat tuning, but GPT-III API came out, I think, over a year before ChatGPT did, but only developers could play with it.”
Luan: Base AI model performance requires doubling compute for consistent gains
“So put another way for just scaling up a base language model you need to double the amount of compute for that language model for it to be predictably consistently smarter.”
Luan: AI model progress is pivoting to synthetic data and reinforcement learning
“Now the critical path for model improvement is, is is shifting over to this to this sort of broader sort of simulation slash synthetic data slash like RL loop sort of path. I think it's just a natural consequence of the fact that it's so expensive to just keep…”
Luan: Pre-Trained LLMs Cannot Discover New Knowledge Beyond Human Data
“A model trained that way is only as good as the smartest data in the training set. Like, it cannot discover new knowledge, because its job, the way the models are trained, is to do what a human would do in that situation.”
Luan: Chatbots and AI Agents Are Becoming Distinct Technology Species
“I kind of think that chat bots like chat GPT and stuff and agents are kind of becoming different species of technology a little bit, right? Like I think they'll be useful in very different ways and what they need to be used for super different, right?”
Luan: Hallucinations help chatbots but destroy AI agent reliability
“Having hallucinations in chatbots and in, like, image generators is, like, a really good thing, right? Because it gives you like, a starter tool for, like, getting to, like, solve the blank page problem, right? Like, and, like, gives you, like, little bits of …”
Luan: AI development feels more like gardening than traditional software engineering
“And I know exactly the behavior of the system that I've built will be. But the cool thing about AI is that every day you come to work and you make some tweaks to the model. And what you get on the other end is actually somewhat unpredictable. Like you kind of …”
Luan: Scaling GPT-2 without architectural changes unlocked three-digit arithmetic
“When we were training GPT-II, we trained GPT-II in various different sizes. And at the smallest size, the model was just, like, unable to do three-digit arithmetic. But as the models got bigger and bigger and bigger, we didn't change anything else. We just had…”
David Luan: Pure model scaling will not solve AI reasoning
“But I think pure model scaling does not deliver solutions to reasoning.”
David Luan: AI reasoning must be solved at the model provider level
“I think the general capability of reasoning will need to be solved at the model provider level. And that's because What you're actually doing is you're not just using the model to reason. You are trying to improve the model's ability to reason, which means the…”
Luan: There will be only five to seven max-scale LLM providers long-term
“I do think there will not be that many LLM players. I think that there will probably be, my guess is somewhere between five to seven long-term steady state LLM providers at maximum scale, just because of the costs involved.”
David Luan: LLMs alone are not products, software systems are
“The underlying thing that everyone's realizing now is that LLMs themselves are not a product. Like, an actual product is this entire software system that uses LLMs in it.”
Luan: Controlling the AI model layer means controlling all underlying compute
“In the future, when more and more software is just like the logic of software is actually just handled by a, by an LLM, nobody cares anymore about what the base computing primitive is. All you need to do is access these models and compose these models to go so…”
Luan: Every major cloud and LLM provider is developing custom chips
“Every one of the major clouds is working on And every major LLM provider is working on a strategy to have their in-house chips because that way they have better margins.”
Luan: Vertical integration pressure will merge model builders and chip makers
“That's my expectation. To me, like what's interesting about AI from a business side, right, is like it forces the question of like, what Companies or offerings are going to be bundled or integrated and which ones are going to get unbundled. And I actually thin…”
Luan: Google's TPU team had under 500 people on a shoestring budget
“I think the TPU team when I was at Google was, like, sub-five hundred people and their budget was a shoestring budget, and yet somehow every generation they taped out quite good chips that were then used to train Gemini and Palm and are used by third parties n…”
Luan: Apple will dominate private on-device AI running at the edge
“And so I think as a result, I think Apple is just going to completely crush at everything that looks like something that's really private, something that's fine-tuned on your own particular data, but doesn't require massive reasoning capability. And that will …”
David Luan: GPT-4o is significantly underhyped relative to its scientific improvements
“I think the degree to which GPT four O was like, I think relatively under hyped relative to what I think the true scientific improvements have been in that model is, is, is this pretty big gap?”
Luan: Developer-focused model sellers must align with clouds or face commoditization
“Companies that sell models to developers will either need to effectively be the like first party effort of one of these big clouds, or they have a short window between now and commoditization to build such a big economic flywheel that they can afford to stay i…”
Luan: Reliable AI agents require full integration from UI to model
“I do think that in the agent space, it's extremely important that you own the entire stack from what is the end user interface. I think like we were talking about the Apple example earlier, owning the interface gives you tremendous leverage in this era of AI t…”
Luan: Every enterprise workflow is an edge case
“I was talking to Parag, who used to be the CEO of Twitter, we were just hanging out the other day, and he's like, dude, every enterprise workflow is an edge case, and he's absolutely right, and that's why you need to control the same thing.”
Luan: In 5–10 Years, Computer Interaction Will Be Goal-Driven
“I just think in five to 10 years, people are gonna use their computers by giving them high-level goals, right?”
Luan: Co-pilots are a great strategy for incumbents adapting to AI
“Like, I think co-pilots are a great incumbent strategy because it lets them morph their existing software business model to something that kind of looks the same while getting in on the AI thing.”
Luan: AI will not take all jobs, humans will drive agentic systems
“Like, that's not, I don't think this is how this is going to play out. I think the way this is going to play out is, Is that what we're going to have as humans fundamentally be the drivers of these agentic systems that like that, like basically give everybody …”
Luan: AI will make workers generalists supervising specialized AI co-pilots
“So I think what it's gonna do is it's gonna make humans at work much more like generalists, and it's gonna have, like, causes to create larger and larger, oh, sorry, like or giving people sort of, like, larger and larger scope over various different, like, are…”
Luan: Enterprise AI adoption will play out over a very long timeline
“We're going to be on this adoption curve for enterprise AI for a very, very long time.”
Adept avoids signing enterprise AI deals funded by experimental budgets
“One of the things we do, for example, is we really try to not sign deals that are coming out of experiment budget because we want quality revenue basically.”
Luan: AI will not plateau like autonomous vehicles due to ongoing model breakthroughs
“What's going to prevent this for what I think has a hope of preventing this from just being a hype cycle that falls flat like AB is that those things are, those shoes are yet to drop. And as they do, the capabilities of these models are going to continue to im…”
Luan: Leading AI companies are already lobbying for regulatory capture
“I think I think the move to go pull up the ladder behind them is already beginning. And I think that lawmakers don't really understand this technology at all. And so their default instinct is sort of listened to the most credible source. And usually those cred…”
Luan: Open-source AI will lag closed models over the next five years
“And I think in the next five years, open will always lag closed. And because open will always lag closed, because open just has fewer resources behind them and fewer incentives for people to go to go make things to be open as these things become more and more …”
David Luan: Chat interfaces are not the right paradigm for AI
“And so that's why the HCI problem, like people just aren't spending enough time thinking about like chat is obviously not it.”
David Luan: Iterative prompting is a flawed interface for AI
“You know, when you go work with a coworker, right, you don't just you don't just talk back and forth. You, like, actually, like, share the same canvas. You'll go use a whiteboard. You'll go, like, maybe look at the same thing on the computer together, and then…”
Luan: AI agents in five years will function like brain-computer interfaces
“Agents in five years time, I mean it's kind of gonna be, like, a non-invasive, like, brain-computer interface, basically. I think that's what an agent will be. Like, all of us are gonna be up-leveled. We're going to, you know, it's gonna feel like the same tra…”
Shorts cut from this episode
▶ How to solve AI reasoning 🧠 · 20VC with Harry Stebbings (@21:01)
▶ Can anyone beat Nvidia? 🚀 · 20VC with Harry Stebbings (@24:37)
▶ Is this the future of machine learning? 🤖 · 20VC with Harry (@22:24)
▶ How OpenAI changed the game 🚀 · 20VC with Harry Stebbings (@0:00)
▶ Will AI Agents supplant RPA? 🤖 · 20VC with Harry Stebbings (@36:28)