Nov 3, 2025 · 59m · 20vc
Cohere's Chief AI Officer, Joelle Pineau: Why Scaling Laws Will Continue & Future of Synthetic Data · 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 episode of 20VC, host Harry Stebbings interviews Joelle Pineau, Chief Scientist and Chief AI Officer at Cohere, on the realities of reinforcement learning, the economic shifts of enterprise AI adoption, the rise of agentic security vulnerabilities, and the necessity of pragmatic, open-source innovation over sensationalized existential fear.
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 3.9 Guest teaching 4.7 Guest disagreement 2.1 Harry pushing back 3.7
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 44:52 Dismissing catastrophic risk narratives
Joelle forcefully rejects doomer claims about AI overlords or existential risk, stating she has no patience for them as a scientist due to their total lack of scientific rigor.
Hardest push from Harry ▶ 16:02 Challenging 10X productivity claimHarry directly confronts Joelle's claim that AI will deliver 10X productivity, explicitly stating he finds her thesis far more unreal and intimidating than replacing 5% of workers.
Biggest teaching moment ▶ 36:56 The genetic island synthetic data analogyJoelle uses a vivid genetic island analogy to educate Harry on why synthetic data leads to model collapse in open domains like language and vision, but succeeds in structured domains like code and chess.
Harry holds his own ▶ 35:48 Deconstructing the AI data vendor marketplaceHarry showcases deep industry domain knowledge by naming key players (Surge, Turing) and analyzing the evolution of data vendors into three required operational pillars.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Establishing AI Hypotheses: Joelle's Tenure at Meta | 3 | 4 | 2 | 4 | Harry cites Andre Karpathy's comment that reinforcement learning is terrible to ask if AI has gotten over its skis. Joelle reframes RL's evolving utility, noting it is less terrible than 20 years ago while setting realistic limits on AGI expectations. | |
| Sequential Decision Making & Why RL is Inefficient | 1 | 6 | 1 | 1 | Harry admits ignorance and asks why RL is so inefficient. Joelle provides a masterclass on sequential decision making, compounding errors, and the difficulty of mathematically specifying reward functions for human behavior. | |
| Training vs. Inference Dynamics & Cohere’s Strategic Focus | 5 | 3 | 3 | 6 | Harry pushes on Cohere's enterprise model, questioning whether software vendors lack incentive to optimize inference if the customer pays for it. Joelle rejects the premise, explaining that customer value drives provider alignment. | |
| Linear Progress of Compute vs. Non-linear Algorithmic Jumps | 4 | 5 | 1 | 3 | Harry asks whether AI progress occurs linearly or via step functions like DeepSeek. Joelle breaks down how compute and data scale linearly, whereas algorithmic discoveries act as non-linear step functions. | |
| The search problem: Why Algorithmic Innovation is Challenging | 3 | 4 | 1 | 3 | Harry asks about the tension between pure academic research and product monetization. Joelle explains why enterprise feedback offers a superior signal compared to artificial academic benchmarks. | |
| Enterprise Productivity Barometers: Sequoia's 5% vs. Joelle's 10X | 5 | 5 | 4 | 7 | Harry cites Sequoia's David Cahn on replacing the bottom 5% of workforce as a productivity benchmark. Joelle rejects this framing in favor of 10X individual productivity, prompting Harry to forcefully challenge her 10X claim as unrealistic. | |
| Venture Budgets, Human Labor Transition, and Task Ambiguity | 5 | 4 | 1 | 4 | Harry re-evaluates VC investment assumptions about moving human labor budgets to AI software spend. Joelle explains that task ambiguity dictates whether automation or amplification succeeds. | |
| Legacy System Integration and Data Confidentiality | 3 | 3 | 1 | 2 | Harry quotes Sam Altman on generational differences in using AI. Joelle notes that enterprise adoption bottlenecks stem primarily from integrating with decades of legacy internal data systems. | |
| The Security Frontiers of AI Agents | 3 | 5 | 1 | 3 | Harry asks about overlooked security risks in AI. Joelle contrasts LLM hallucinations with agent impersonation risks, detailing security vectors in autonomous systems. | |
| AI Standards: The Balance Between Government and Enterprise | 5 | 6 | 4 | 6 | Harry questions whether governments are competent enough to set AI standards. Joelle rejects the pessimistic framing, pointing to historic regulatory successes like aviation safety. | |
| Sovereign AI Models and Local Multilingual Strategies | 4 | 5 | 1 | 3 | Harry asks about sovereign AI models and talent strategies. Joelle outlines why stacking AI superstars fails without execution focus and social glue within teams. | |
| The "Galacticos" Star Players Debate | 6 | 4 | 3 | 7 | Harry bluntly challenges Joelle, asking why labs buy 'Galacticos' like Daniel Gross or Alexandr Wang if superstar teams aren't required. Joelle clarifies that while a few core talents are needed, team composition and compensation alignment matter more. | |
| Specialized Curation and the Data Marketplace | 6 | 4 | 1 | 4 | Harry names major data platforms like Surge and Turing to probe market longevity. He demonstrates industry expertise by detailing how data vendors must now supply talent, curated data, and evaluation implementation. | |
| Model Collapse and the Genetic Island Analogy of Synthetic Data | 3 | 7 | 0 | 1 | Harry inquires about model collapse from synthetic data. Joelle provides an insightful 'genetic island' analogy to explain where synthetic training causes distribution collapse versus where it succeeds. | |
| Image Generation History as a Predictor for AI Code Quality | 3 | 6 | 2 | 3 | Harry voices concern over AI generating poor code. Joelle reframes the concern by drawing a historical parallel to primitive 2015 image generation, predicting vast improvements over a ten-year horizon. | |
| Team Curation over Creation and Fundamental Redesign | 5 | 5 | 3 | 5 | Harry argues that human roles limited to curation contradict true human-AI partnership. Joelle humorously counters that curation represents the promised 10X productivity leap while defending language as a dense symbolic interface. | |
| Scientific Rigor: Joelle's Past Skepticism of Neural Networks | 2 | 5 | 6 | 2 | Joelle admits her past scientific error regarding neural network viability over SVMs. She then aggressively dismisses existential risk and doomer narratives as lacking scientific rigor. | |
| Risk Variance and Tolling the AI Capital Bubble | 5 | 6 | 2 | 5 | Harry cites industry commentary calling evaluation benchmarks 'bullshit'. Joelle reframes evals as software unit tests rather than absolute metrics of enterprise ROI. | |
| The Academic-Corporate Disparity and Talent Flows | 5 | 4 | 1 | 4 | Harry asks if academic institutions are priced out of AI compute and questions massive founder valuations. Joelle defends university research relevance, noting NeurIPS paper awards consistently go to academic labs. | |
| Quickfire Round: Sandbox Agent Societies & Youth Social Dynamics | 2 | 2 | 1 | 1 | In a quickfire round, Joelle shares her desire to build sandbox agent societies and deadpans that her main parental restriction on kids is limiting sugar. | |
| The Impact of Social Media on Youth Mental Health | 3 | 5 | 3 | 3 | Harry asks about youth mental health and working with Mark Zuckerberg. Joelle cautions against blaming technology without rigorous data, praising Zuckerberg's intense technical deep-dives. | |
| Analyzing the Economics and Compensation of AI Talent | 4 | 6 | 4 | 5 | Harry asks about talent compensation and open source trends. Joelle drops a statistic showing 20 million monthly downloads for a 2019 open model (RoBERTa) to prove demand for efficient models, calling closed-source pivots a 'deep mistake'. |
Statements from this episode (45)
Pineau admits she was wrong to be skeptical of neural networks
“I used to be quite skeptical that neural networks were necessarily the ultimate
A solution to machine learning.
I seem to be quite wrong on this one.”
Pineau dismisses catastrophic AI existential risk predictions as unscientific
“I don't have a lot of patience as a scientist for people who are predicting the extremist scenarios, the catastrophic risks of AI, you know, AI becomes our overlord kind of scenario.”
Pineau: Reinforcement learning is fundamental to AI and will not disappear
“Oh, I'm still super bullish on RL in that, like, the concept itself is so fundamental. You know, this idea of training through a system of rewards, of indicating what's valuable and what's not valuable through numerical values, like, that is so fundamental. It…”
Pineau: Out-of-the-box reinforcement learning will not deliver AGI
“Now, you know, where we're maybe getting a little bit ahead is thinking that just RL out of the box is gonna give us AGI. That part, a lot less so. You know, if you look at the curve of progress, RL is terribly inefficient, and so the amount of signal you need…”
Pineau: RL costs are falling where clear reward functions exist
“It's coming down, especially in domains where we have good reward functions.”
Pineau: Using reinforcement learning to teach AI social behavior remains unsolved
“RL, to shape the behavior of models, to get them to be social creatures, that we have no idea how to do. I mean, I don't know if you have children, but like shaping their behaviors, you know, the number of times you can repeat the same thing, and still they do…”
Pineau: Cohere develops on-premise AI models to offload inference costs
“One of the things that Cohere is doing is actually to develop AI models that run on premise. So that means enterprise bring it in, they run it locally, so the company has to worry about the training Of the models. Obviously we want world-class models for the n…”
Pineau: Unpredictable GPU needs and returns are AI's biggest economic challenge
“I think one of the biggest challenges, the fact that it's very hard to have predictability, right? Everyone wants to know when are we going to hit the breakthrough? Everyone wants to know how many GPUs do I actually need? Everyone wants to know, like, what's t…”
Pineau: Compute and Data Scale AI Linearly, While Algorithms Cause Step-Functions
“Compute and data have a more linear effect on progress. You build more compute. You run bigger models. You can typically get better performance. You feed in more data. It's not just quantity. You need to worry about quality and diversity as well, but roughly i…”
Pineau warns against betting that AI compute scaling will hit a wall
“They've been, the scaling laws have been remarkably robust. They don't play exactly as we expect, but still, they've been remarkably robust. Lots of people have bet against scaling laws in the past. And I would say overall, you know, we've seen a pretty, prett…”
Pineau: Algorithmic innovation is the hardest AI area for investors to evaluate
“So, in that sense, I think it's the most interesting one, it's the most frustrating one, and it's the most difficult one, certainly from investors' point of view, to know where to put your chips.”
Pineau: AI is starting to be useful, but trails public perception
“We're at a stage where AI is really starting to be useful, maybe not as useful as people think it is, but we are there.”
Pineau: Selling enterprise AI offers better feedback than academic benchmarks
“When you need to sell AI to a business, you get a real signal of what works, what doesn't work. And you know, we've been using these academic benchmarks for many years. You get some signal, but it's not the same as getting this to do productive work.”
Pineau: Replacing portions of enterprise workforces with AI is unrealistic
“So to just like flat out replace a portion of your workforce is actually Pretty unrealistic.”
Pineau predicts AI will enable 10x employee productivity within two years
“Oh, I don't think that's unrealistic at all.”
Joelle Pineau: AI automation succeeds when task results are precisely specified
“So in any case where we can be very precise about what a great result looks like, we'll be able to make that task automatic much more easily than tasks that are much more nuanced and have a lot of complexity.”
Pineau: Users treat AI as a work tool rather than a companion
“I see a lot of people using it as a tool, more than as a companion. You know, people have this, you know, Swiss knife in their work life all of a sudden that can be super helpful, but that's really most of what I see.”
Pineau: Practical AI applications will be discovered by field practitioners
“We don't have all the answers of how it should be used. That's gonna come from people on the field.”
Stebbings: VCs routinely glaze over security in application-layer AI deals
“Security is a topic that we quite often glaze over, especially when investing in kind of application layer AI tools.”
Pineau: AI agents are opening an unmapped new frontier in security
“With respect to AI security, I think there's a new front that's opening up with the development of agents, and frankly, there's a lot we don't know yet in terms of the vulnerability of these systems.”
Pineau: Impersonation is the AI agent parallel to LLM hallucinations
“In terms of agents, you know, we worry a lot about hallucinations in LLMs. The parallel in agents is impersonation.”
Pineau: Building AI models outside the US and China creates a healthier industry
“I do think it's healthy that there are models that are getting built in different places around the world, not just in the US and China right now. I think this is healthy in terms of diversity of thoughts. I think it's healthy in terms of having a greater amou…”
Pineau: Assembling AI superstars without execution and glue fails
“I don't think it becomes that productive to put a bunch of AI superstars all together in a room without the execution machine, without the social glue.”
Pineau: AI teams need a few elite talents, not a full roster of superstars
“You do need a few of these, like, uber talents in the team. There's a relatively, you know, short number of people who just Understand this technology very deeply. You do need some of this talent, and if you can afford it, you should get some of that talent. B…”
Pineau: Multi-billion-dollar price tags for top AI talent are unnecessary
“Time will tell. I don't think it's necessarily needed to go at that scale, but time will tell.”
Pineau: Having AI talent without compute is a waste of time
“If you have too much talent and not enough compute, you're wasting your time, so usually like an equilibrium between those two pieces.”
Pineau: Simple data labeling is obsolete as AI requires complex tasks
“The days of like having data labelers who can say this is a cat and this is a dog or somewhat over like the easy test the AI can do. So we're getting in a space where we need more specialized tasks.”
Pineau: Human-guided AI training is a permanent necessity, not a phase
“I don't think it's a phase in the sense that I do think this partnership, we'll call it, between humans and machines, where humans provide guidance to machine, like, we are in this for a long time. What will change is the nature of the information that the AI …”
Pineau: AI data market is shifting from labeling to crafting environments
“I think for me, the even bigger trend we're seeing is the moving from just labeling data to crafting environments to produce new, new tasks.”
Pineau: Synthetic image and language data causes model degradation
“So in some domains, if you think like images, languages, like LLMs talking to each other at some point, you definitely get the degradation and that degradation is due to essentially like a loss of diversity of your data.”
Pineau: Synthetic data for code avoids model collapse through injected diversity
“If I think of coding, we can generate synthetic code. You take normal code and we know how to inject diversity into the code. Like I can take a couple of repositories, mix and match, apply an LLM to transform it. And so there's a way to generate synthetic data…”
Joelle Pineau: AI Code Generation Quality Will Be Excellent in 10 Years
“So you think of code generation, like right now we're in the phase we were for image 10 years ago. Yes, there's a lot of bad code that's getting generated. There's a lot of code that will get thrown away, but wait another 10 years and I think the quality of th…”
Pineau: Software engineering will shift from writing code to curation
“What matters now is sort of, you know, picking the quality out of the volume, and so if I fast forward 10 years on code generation, when we have the ability, To generate a ton of code to do a ton of different things. We're going to need some selection mechanis…”
Pineau: Human-AI interaction will expand to voice, gesture, and eye gaze
“It's awfully limited, and, you know, prompts can mean a few different things, but the idea of, like, typing in a box, that to me is very limited, and we're gonna break out of that box already. We're seeing a lot of cases where voice is a lot more natural as an…”
Pineau: Language will remain the primary paradigm for human-AI interaction
“But language is incredibly powerful. So if you think of prompt as being more language as a way to express ideas and communicate with a machine, that's a powerful paradigm. I mean, as humans, we, so much of our communication is based on language. I don't think …”
Pineau: The AI capital bubble will see massive upswings and downswings
“I think about it as a bubble with bigger variants. It's like, you know, the upswing is going to be bigger, and you know, there's going to be big downswings as well, and so there's a lot of variants into the system right now.”
Joelle Pineau: AI benchmarks should be treated as narrow unit tests
“Think of evaluations as, like, unit test for the performance of your system. I mean, software engineers will know what that is, right? Like, you run through that evaluation, and that gives you, like, a signal of how the system is doing in a particular dimensio…”
Pineau: Enterprise clients don't care if AI models win math Olympiads
“We build AI systems that go into enterprise. None of our clients ask about, like, are you able to win the math Olympiad with this model? That's not what they care about. They care about bringing value to their business.”
Pineau: Academic researchers often win best paper awards at NeurIPS and ICML
“You go to the major international conferences, Nureps, ICML, and others, and often the best paper awards are actually won by researchers out of universities.”
Pineau: Universities allow riskier small-scale AI research than corporate labs
“There's a lot of good ideas that you need to test out at small scale. And in a university, you have a lot more freedom to pick pretty risky ideas at a small scale, but still, you know, no one's asking you to justify your research in, in ways that, that often h…”
Pineau: AI in Healthcare and Science Will Make Transformative Progress Within Five Years
“There's a lot of verticals, whether healthcare or scientific discovery that I think have increased Incredible promise where we're going to see real tangible progress within five years that are going to change completely the face of what we can do. So that's pr…”
Joelle Pineau: Blaming AI and platforms for mental illness is an oversimplification
“I think the, we have to be careful about taking shortcuts and saying, you know, because suddenly, you know, we have certain platforms, we have AI, and so on, that is causing that mental illness.”
Pineau: Mark Zuckerberg dives incredibly deep into AI technical details
“He is incredibly deep into understanding the work. Like, he does not coast, you know. When he started getting into AI, just the depth of the question that he'd ask, he just gets really interested in the topic and goes super deep, and that then just informs eve…”
Pineau: Meta's 2019 RoBERTa model hit 20M monthly downloads during LLM hype
“We were in, in the frenzy of large language models, and I pulled the stats on, you know, most downloaded models of last month. We had a model like Roberta from 2019, small language model, was getting twenty million downloads a month.”