Apr 18, 2025 · 1h 3m · 20vc
20VC: Foundation Models: Who Wins & Who Loses | How Economies and Labour Markets Need to Change in a World of AI | China vs the US in an AI Race: What You Need to Know | Rich Socher, Founder @ You.com
0:00 / 0:00
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of TwentyVC, host Harry Stebbings interviews Rich Socher, founder and CEO of You.com, to discuss the economic realities of foundation models, enterprise AI adoption, and frontier applications across biology and robotics. Socher offers deep technical and strategic insights on model commoditization, workforce evolution, VC investment dynamics, and global technology policy.
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 26.8% of the talking time here. How this is scored →
Harry as informed peer 3.8 Guest teaching 4.1 Guest disagreement 1.1 Harry pushing back 2.5
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 53:52 Refusing host investment choices
Rich rejects Harry's forced selection between OpenAI at $300B, Anthropic at $60B, or Grok at $50B, arguing open-source models put too much pressure on pure infrastructure pricing.
Hardest push from Harry ▶ 8:59 Challenging telco analogy on retentionHarry directly interrupts Rich's comparison of LLMs to telcos by citing 10-year LTVs and high retention metrics in telecom, forcing Rich to admit his analogy breaks down.
Biggest teaching moment ▶ 21:26 Exposing true LLM training run economicsRich educates the host on AI compute economics, debunking DeepSeek's $6M PR figure by detailing the hidden costs of hyperparameter tuning and ablation runs.
Harry holds his own ▶ 8:59 Host demonstrates telco business metricsHarry uses specific unit economics knowledge regarding telecom customer retention and contract structures to correct the guest's financial comparison.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Welcome & Rich Socher's Background in AI | 1 | 2 | 0 | 0 | Harry introduces Rich and asks for his background. Rich provides an extensive overview of his academic and entrepreneurial career in NLP and search. | |
| Evaluating the Current State of LLMs and Intelligence Limits | 2 | 4 | 0 | 0 | Harry asks how to evaluate the current state of LLMs. Rich provides a framework dividing intelligence into ten dimensions to assess upper bounds. | |
| The Commoditization of LLMs and Value Capture | 6 | 3 | 2 | 6 | Harry interjects to challenge Rich's telco analogy by citing high customer retention and 10-year LTVs in telcos, which Rich concedes breaks his comparison. | |
| Why Chat Ads Fail and Google Search Moats | 5 | 4 | 1 | 3 | Harry cites conversion stats from Vercel's CEO while Rich explains why chat ads perform up to 100x worse than search ads due to default bias. | |
| Enterprise LLM Adoption and the Manager Mindset Shift | 3 | 5 | 0 | 1 | Rich explains why enterprise AI deployments fail, citing low active seat utilization due to individual contributors lacking management skills for agents. | |
| Horizontal vs. Vertical AI Agents and Action UI Limits | 5 | 4 | 2 | 4 | Harry questions data availability and defends Google's positioning, while Rich critiques natural language UI limitations for action agents. | |
| DeepSeek's Impact and the Real Cost of Training Models | 3 | 6 | 1 | 2 | Rich breaks down the difference between headline training run PR costs ($6M) and total ablation and GPU investment costs ($100M-$200M). | |
| VC Returns, Model Distillation, and Opportunities in Bio | 6 | 3 | 1 | 3 | Harry outlines a thesis on how VC dilution and stock-based compensation erode infrastructure LLM returns, which Rich validates before citing Jevons paradox. | |
| AI in Research, Medicine, and the AI Economist | 3 | 6 | 0 | 1 | Rich explains the AI Economist reinforcement learning model and highlights DeepMind's superior marketing in scientific research. | |
| The Bull Case for Humanoid Robots in Unstructured Home Environments | 5 | 4 | 2 | 5 | Harry challenges the timeline for home robotics by raising fine-grained hand dexterity requirements, forcing Rich to outline specific home use cases. | |
| AI's Potential in Deciphering Complex Biological Systems | 2 | 6 | 0 | 0 | Rich explains how AI is uniquely suited to model complex biological systems and emergent behaviors like the gut microbiome that traditional science misses. | |
| Social Impact, Job Disruption, and the Real Pace of AI Adoption | 6 | 4 | 3 | 6 | Harry pushes back against historical adoption comparisons, arguing software updates deploy instantly compared to past agricultural shifts. | |
| Model Orchestration and the Future of AI Prompting UI | 5 | 3 | 1 | 3 | Harry criticizes manual model selection in product UIs as archaic, leading into a broader discussion on UBI and meaning in work. | |
| Career Advice for Young Graduates: Combining CS with Domain Focus | 4 | 5 | 2 | 4 | Harry asks why young graduates should learn computer science if coding is automated, and Rich frames CS as a fundamental way of structured thinking. | |
| The Transformation of Software Engineering Teams and Vibe Coding | 3 | 5 | 1 | 2 | Rich explains why non-coders encounter limits when vibe coding due to a lack of understanding regarding complexity bounds and algorithmic efficiency. | |
| Developer AI Tools and Switching Costs | 4 | 4 | 1 | 3 | Harry brings up enterprise fears around data privacy and retention, while Rich details security prerequisites for enterprise LLM vendors. | |
| AI Startup Valuations and the Return to Fundamental Moats | 6 | 3 | 2 | 4 | Harry points to poor retention data for DeepSeek and examples of OpenAI wiping out thin startups to question startup defensibility. | |
| Common Public Misconceptions Surrounding AI Capabilities | 3 | 5 | 0 | 1 | Rich critiques public AI misconceptions and details how quantum computing could enable exact molecular and cellular simulations. | |
| AGI Bets and Audacious Goal-Setting | 5 | 4 | 2 | 4 | Harry asks Rich to pick between investing in OpenAI, Anthropic, or Grok, but Rich rejects the premise and declines all three due to high valuations. | |
| Defining Personal Success, Impact, and Legacy | 1 | 2 | 0 | 0 | Harry asks personal reflection questions about legacy, money, introversion, and health habits for longevity. | |
| Policy Recommendations to Fix EU AI Regulation and Boost Innovation | 2 | 5 | 1 | 1 | Rich delivers a comprehensive policy and regulatory blueprint for European tech, advocating for mandatory CS and sovereign wealth funds. |
Statements from this episode (56)
Socher: Peer reviewers publicly rejected the original prompt engineering paper
“We invented prompt engineering which was majorly rejected publicly on open review an idea that made no sense to the reviewers.”
Socher: Enterprise Is the Killer Application for Large Language Models
“We realized eventually the killer app for large language models and complex answers is an enterprise.”
Socher: LLMs are already good enough for most business tasks
“And so LMs are already good enough. They just need to be brought into companies to be actually made useful.”
Socher: Object detection in computer vision is practically solved
“For example, object detection and computer vision. It's actually kind of solved. We can classify most objects on the planet.”
Socher: Thin-Layer LLM Providers Will Function Like High-Capex Telcos
“I think LLM companies, especially just the pure thin infrastructure layer of LLMs are going to look, I think, more and more like telcos in the sense that it's high capex, huge, you know, expenditure to build it, especially if you want to build it from scratch.…”
Socher: OpenAI Is a Consumer App Company, Not Pure Infrastructure
“That's why I said, if you're in that thin infrastructure layer, and that was an important qualification because OpenAI is a consumer app company.”
Socher: OpenAI generates vast majority of revenue from ChatGPT
“They make their revenue, the vast majority of their revenue from a consumer app called ChatGPT.”
Socher: Anthropic under pressure due to Claude's low consumer market share
“Anthropic has a lot more pressure to keep building the best models because Claude is so much smaller in terms of market share for the consumer app.”
Socher: Other consumer LLM apps are rounding errors compared to ChatGPT
“Consumers, once you're really famous and you cross that threshold of just like being well known, being the default for a lot of people, all the other LM apps companies are almost rounding errors to ChatGPT.”
Socher: Chat ads perform up to 100x worse than search ads
“We actually evaluated that they work about 10 to a hundred X worse than search ads and you have. Twice the cost, about.”
Socher: 80% of iPhone users never change a single device setting
“80% of all iPhone users never change a single setting of any kind.”
Socher: LLMs will capture search whenever queries are complex
“And so LMs as part of that unbundling wave of Google, LMs will capture whenever you have a more complex question.”
Socher: Enterprise OpenAI deployments see weekly active usage drop to 6%
“They had to pay a thousand seat licenses for OpenAI, and then six months later, they realize only six percent are actually using them every week.”
Socher: AI will turn every individual contributor into a manager
“With AI, every person will become a manager, but most people are not used to managing other people or processes.”
Socher: Natural language is not the best interface for all AI tasks
“As much as I love natural language is not the single best interface for a lot of different types of answers. Sometimes you want to see a map. Like sometimes you want to see a table. Sometimes you want to see a map with a bunch of specific overlays.”
Socher: Web action AI agents are in a valley of disillusionment
“We're sort of in this valley of disillusionment on a lot of these, what I call action agents that go on the web and actually do something for you and take actions that you can't undo and say you buy a ticket that's not refundable or something. There's this val…”
Socher: Google's AI struggles stem from ad-model innovator dilemma, not tech
“I think it's never been a question of technical strength for Google. It's just a question of classic innovator's dilemma. You make money by showing ads and lists or blue links. So it's hard to give people just a straightforward, useful answer.”
Socher: Apple has failed to execute effectively on its consumer AI position
“In theory, Apple would be so well positioned, but in practice, they've been not doing much.”
Socher: DeepSeek bypassed top LLMs rapidly due to low switching costs
“I think one is that personalization, and that's why it's so easy for people to switch around LLMs too, you know, like DeepSeek overtook almost every other thing other than ChatGPT within the weeks.”
Socher: DeepSeek model training likely cost $100M-$200M, not single-digit millions
“It's also clear that it probably cost them a 102 hundred million dollars, but it's still incredibly cheaper than billions of dollars that we're told it would take to train these kinds of models.”
Socher: AI model development requires hundreds of smaller ablation runs
“Generally in model development, you train one finally, like really good final good model. On the path to that, you had to run many, what we call ablations or hyperparameter runs where you tune a little bit, like, Should this joint be, you know, moving this muc…”
Stebbings: Venture investors will not make money from LLMs
“I don't think any venture investors are going to make money from LLMs.”
Socher: Avoids personal investments in pure language model infrastructure
“I have personally stayed away from investing in any pure LM infrastructure companies.”
Socher: Lower marginal cost of AI will expand usage everywhere
“And so I think the analogy here is that yes, AI, like the cost, the marginal cost of intelligence keeps going down, but that just means we're going to use it more and more places.”
Socher: Economics lacks an AI breakthrough because it moves too slowly
“And the field of economics hasn't had their ChatGPT moment yet, because it's such an old, slow moving field. They don't have archive papers. They don't have conferences where you publish every couple of months.”
Socher: AI simulation of economic policy is an underrated application
“And instead we could ask an AI to give us some advice on like, what would be the actual best setup if you have a certain set of goals? You wouldn't have to like try out like a massive tariff change in the world. You could just like ask an AI model first to sim…”
Socher: DeepMind Has the World's Best Science Marketing
“Their science marketing is, is probably the best in the world, and so that paper never quite had its moment in the sun yet.”
Socher: Humanoids Are Suboptimal for Highly Repeatable Industrial Tasks
“And so, for almost every process that is highly repeatable, there's a better, more quickly evolved new hardware form than five fingers on two arms.”
Socher: Household robotics face same long-tail complexity as radiology AI
“It's very easy to build one radiology classifier for one thing and make that better than a human, but then there's this very long tail of things, and so you need a lot of money and a lot of resources and a lot of data to see the long tail of all things in radi…”
Socher: AI will solve major medical problems within ten years
“Yes. That is one of, like, several, like, several chapters in my book are about that.”
Socher: 60% of U.S. adults have never used an AI chat model
“I don't think it'll happen as quickly as people think. There's still, I think, 60% of all adults in the U.S. Have never talked to a chat model, like, at all. And that's the U.S. Like, just go to Europe, I'm sure that is even higher.”
Socher: Humanity's long-term purpose is spreading consciousness across the universe
“In fact, I think one of the biggest civilizationary unlocks would be to all agree that entropy or darkness in the universe is the enemy and that we should, like, spread consciousness into the universe. If we can all agree that's the goal, then we can always be…”
Socher: UBI will deprive people of meaning and worsen societal crisis
“I think the problem is that while most people complain about their jobs. It does give them meaning. It does give them meaning to be a valuable part of society and to have earned something that they can then give to their family, to their kids, and so on. And s…”
Socher: Humans will have more significant AI friends in the future
“I think we will have more significant AI friends, but I hope it's not more significant than our friendships with people.”
Socher: Universal AI real-time audio translation will arrive within five years
“And maybe that technology is like fully there in five years or so, like so, so, so good, so cheap, so prevalent that everyone will just travel with it and everything.”
Socher: Learning programming teaches a mental framework, not just code writing
“Programming isn't just about programming itself and being like an IT or being a programmer or a developer or so on. It's also about a different way of thinking. And it's a different way of understanding the world that you're in. And so if you have an understan…”
Socher: Students in medicine and law must combine studies with computer science
“Even if you want to study law, medicine, chemistry, or whatever, you should combine it with computer science, because all of these fields are going to change over the next couple of decades.”
Socher: Entry-Level Tech Jobs Are Increasingly Automatable
“I think the biggest problem and biggest worry I have is that the entry level jobs right now are more and more automatable.”
Socher: Programming Knowledge Is Required for Effective Vibe Coding
“But if you don't know how to program at all, you're also not going to be as good of a vibe coder.”
Socher: AI tools lack switching costs without proprietary enterprise data integration
“There is very little switching cost. Same is true for LLMs in the consumer world, right? None of them are that amazing yet. None of them do enough personalization yet. Now, where there is a lot of switching costs is if you have company internal data or you hav…”
Socher: AI startups cannot be great consumer and enterprise companies simultaneously
“And that's where you really can't be like a great consumer company and a great enterprise like company at the same time.”
Socher: AI startups trading at 80x-100x ARR without moats face correction
“We're going to see a correction when companies are trading hundred ADX their ARR and they don't have a real moat and like there's the switching cost is close to zero to go to DeepSeek or something else.”
Socher: Point-solution AI startups will not all die overnight
“I actually don't think they're all gonna just die overnight.”
Socher: Speech recognition startups reach $100M+ revenue despite commoditization
“There's still several companies that make millions and, like, tens or even hundreds of millions of revenue doing speech recognition just perfectly.”
Socher: Current computing can accurately simulate only a few hundred atoms
“We can right now like really, really accurately only simulate like a few hundred atoms at best and how they really, truly interact with one another.”
Socher: AI can solve every problem in any domain that can be simulated
“In AI, anything you can simulate, AI can solve every problem in that domain.”
Socher: Quantum biological simulation will enable AI to cure cancer and MS
“Once you can simulate a cell and then multiple cells and organs and organisms, like all of a sudden AI can try billions of different things on how to cure that cancer, how to cure MS, how to cure like all the bacteria and viruses and like all these different t…”
Socher: Expects to win $1,000 bet that AGI won't arrive by 2027
“We did this bet and in the bet he has to win. I think it ends in 20, 27. So three things have to be true. We have to have a personal robot that cleans the whole house the way my cleaning team does. And it needs to be purchasable, like for reasonable amounts of…”
Socher: Sonnet 3.7 is the best model for coding and engineering
“Sonnet 3.7 is actually the best model for a lot of things, especially in coding and engineering and so on.”
Socher: OpenAI, Anthropic, and Grok won't generate 1,000x returns
“I personally would just, I also just love investing in early stage where you can have thousand X's and so on. I just don't see a thousand X's for those companies.”
Socher: Money is one of the best indicators of impact
“I realized at some point you have to care about money because money is one of the best indicators of impact and allows you to do epic, cool things, but I don't intrinsically care about it.”
Socher: Longevity Is Massively Underrated and Will Improve via AI
“Improving longevity as like one of those AI plus bio corollaries or follows that is just really, really exciting. And I think longevity is massively underrated.”
Socher: EU AI regulation is destroying Europe's fledgling tech ecosystem
“Unfortunately, the EU has shot itself in the foot. A lot of different kinds of regulation, and in particular AI regulation, destroying a fledgling ecosystem that could never get off the ground.”