Sep 1, 2025 · 1h 14m · 20vc
Cohere Founder, Nick Frosst: How To Compete with OpenAI & Anthropic, and Sam Altman’s AI Disservice · 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, Cohere co-founder Nick Frosst joins host Harry Stebbings to discuss Cohere’s pragmatic, enterprise-first AI strategy, critiquing the hype of existential AGI threats while addressing real-world economic impacts, labor shifts, and the path to building a generational tech company.
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 22.3% of the talking time here. How this is scored →
Harry as informed peer 3.7 Guest teaching 5.4 Guest disagreement 3.2 Harry pushing back 3.1
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:05:18 Calling Sam Altman's rhetoric academically disingenuous
Nick directly attacks Sam Altman, stating his false predictions on AGI and world tour warnings about existential threats were academically disingenuous and did a disservice to AI.
Hardest push from Harry ▶ 49:11 Harry defends market pay disparities via Adam SmithHarry forcefully refuses Nick's egalitarian view of labor, arguing that society and Adam Smith's invisible hand place definitively higher value on his interview skills than on entry-level grill work.
Biggest teaching moment ▶ 15:47 Debunking public benchmark utilityNick systematically breaks down popular benchmarks like HellaSwag and Arc AGI, demonstrating to Harry how easily gamified and completely detached they are from real enterprise ROI.
Harry holds his own ▶ 49:00 Harry asserts the economic value of specialized talentHarry challenges Nick's stance on worker value by citing his own 10 years of disciplined craft and higher market compensation compared to entry-level social media or restaurant workers.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Working with Geoffrey Hinton at Google Brain | 2 | 4 | 1 | 1 | Harry asks standard exploratory questions about working alongside Geoffrey Hinton at Google Brain. Nick warmly describes Hinton's intuitive, physical analogies for research and reflects on why transformer creators left Google. | |
| Cohere's Enterprise Focus & Synthetic Data | 2 | 5 | 1 | 1 | Harry asks about Cohere's enterprise focus compared to consumer models. Nick explains how training on synthetic business data and tool APIs differs from training conversational models. | |
| The Limits of Scaling Laws and Compute Hype | 5 | 4 | 3 | 4 | Nick turns the tables by asking Harry if he thought GPT-5 was better than GPT-4. Harry defends his view that it felt worse due to cumbersome model auto-selection and UX friction. | |
| Defining AGI and Automating Enterprise Workflow | 3 | 6 | 2 | 2 | Nick highlights the difference between personal automation and enterprise workflow automation. He defines AGI as treating a computer like a person, which current LLMs do not achieve. | |
| Value Capture: Infrastructure vs. Application Layer | 4 | 6 | 2 | 2 | Harry cites specific players like Anthropic and Cursor to ask about value capture across layers. Nick educates on the spectrum between 2015 task-specific neural nets and broad general language transformers. | |
| Debunking AI Benchmarks and Evaluation Metrics | 4 | 7 | 4 | 5 | Harry directly asks if popular benchmarks are bullshit and why Cohere lags on public evals. Nick breaks down benchmarks like HellaSwag and Arc AGI, explaining why they are easily gamified and irrelevant to enterprise utility. | |
| The War for AI Talent and Model Timelines | 4 | 5 | 3 | 3 | Harry brings up sensationalized headline compensation figures for AI researchers. Nick expresses skepticism about $100M compensation packages and emphasizes stable culture and purpose over PR hype. | |
| The Hype of Existential Threat vs. Real Risks | 2 | 6 | 4 | 1 | Nick argues that existential threat claims are misleading and obscure real immediate risks. Harry listens as Nick explains how doom-mongering distracts from pragmatic societal impact. | |
| AI as an Augmentative Tool vs. Job Replacement | 5 | 6 | 7 | 7 | A heated clash occurs when Harry argues junior marketing managers will be replaced by AI agents because they aren't brilliant. Nick forcefully disagrees, arguing statistical sequence models lack intuition and human context. | |
| Labor Impact: Historical Parallels and Future Policy | 3 | 6 | 3 | 2 | Nick draws historical parallels to the Industrial Revolution, emphasizing that labor policy and unions emerged to protect workers. He argues AI deployment requires thoughtful public policy to navigate labor transitions. | |
| Addressing the Rise of Income Inequality in the AI Era | 3 | 5 | 3 | 3 | Harry asks about open vs closed models and Meta's strategy. Nick explains Cohere's commercial vs research weight-releasing strategy and warns against obsessing over competitor moves. | |
| The Evolution of Prompt Engineering & Model Mechanics | 3 | 6 | 2 | 2 | Nick details how prompt engineering was initially just tricking web-trained sequence models with formats like 'In summary:'. He explains that prompting as a specialized discipline is disappearing as models align with natural speech. | |
| Fundraising Dynamics, Enterprise Clients, and Efficiency | 3 | 6 | 2 | 2 | Harry asks how Cohere allocates its $600M raise between compute and talent. Nick explains Cohere's focus on training efficiency, designing models to fit on just two GPUs. | |
| Competing with Tech Giants and the Value of Forward Deployed Engineers | 4 | 6 | 4 | 6 | Harry asks if deploying forward deployed engineers implies the underlying technology is weak. Nick rejects the premise, explaining that enterprise deployment naturally requires custom integration support. | |
| Valuation, M&A Offers, and the Philosophy of Legacy | 6 | 5 | 7 | 8 | Harry aggressively claims his interviewing work is paid millions because society values it more than grill workers or junior managers. Nick firmly rejects Harry's market determinism, pointing to his own experience as a cook to defend the value of all labor. | |
| Co-Founder Disagreements, PR, and Sovereign AI Infrastructure | 4 | 5 | 3 | 3 | Harry asks why Cohere founders aren't constantly in front of cameras like Perplexity or OpenAI leadership. Nick reframes the need for media hype as a consumer app strategy that does not apply to B2B enterprise infrastructure. | |
| Geopolitics, Geographies, and Future Hardware Interactivity | 4 | 5 | 2 | 2 | Harry asks about sovereign AI models. Nick explains why being a Canadian company acts as a geopolitical asset for global customers wary of US government intervention in tech. | |
| Loneliness, Technology, and Human Connection | 4 | 6 | 3 | 3 | Harry voices deep concern over youth mental health, loneliness, and influencer culture. Nick offers historical perspective, citing Greek worries about writing and past panics over newspapers. | |
| Sam Altman's AI Rhetoric and Existential Threat Claims | 5 | 6 | 8 | 5 | Nick delivers his strongest attack, calling Sam Altman's AGI and doom predictions academically disingenuous. Harry asks if provocative claims are simply necessary for fundraising. | |
| Post-Round Founder Ritual: McDonald's | 3 | 4 | 2 | 2 | Nick reveals his lighthearted co-founder ritual of eating McDonald's Junior Chickens after closing funding rounds. He also critiques benchmark-based regulation. | |
| AI Workflows, Cursor, and Global AI Companies | 4 | 5 | 2 | 2 | Nick discusses using Cursor for coding and predicts a trillion-dollar AI company emerging from Canada. He candidly shares a past scientific mistake doubting RLHF data efficiency in 2020. |
Statements from this episode (38)
Frosst: Sam Altman's AGI timeline hype does a disservice to the world
“I don't think Sam Altman has done a service to the world by talking about how close AGI is.”
Frosst: Sam Altman's AI predictions were obviously wrong when made
“I think he has made several predictions now that are wrong, and that were obviously wrong at the time he made them.”
Frosst: Sam Altman warning leaders about AI threat was disingenuous
“He did a world tour where he spoke to every major leader the world over to tell them, hey, this technology is gonna pose an existential threat. And I think that was academically disingenuous, and I think did a disservice to the technology he loves.”
Geoffrey Hinton develops AI algorithms using physical intuition rather than equations
“I think I was very surprised at how creatively and playfully he approaches research. When we would discuss like algorithms or like optimizers or loss functions, we would discuss them often in like through physical analogy. So we'd spend a lot of time talking a…”
Frosst: Google failed to quickly commercialize or scale the Transformer
“It wasn't then commercialized very quickly within Google. It wasn't scaled up very quickly within Google. A lot of that work had to be done elsewhere and years later.”
Frosst: All creators of the Transformer architecture eventually left Google
“It is interesting that all the people who worked on the transformer left To continue to work on the transformer.”
Frosst: Fewer than 20 companies globally build foundational large language models
“Maybe there's like 10 companies in the world that are building, like, large language models. In the, yeah, maybe the, in, maybe there's like 15 in, I don't know, we'll have to figure out how, there's a few that have popped up recently. But there's some number,…”
Frosst: Core transformer architecture has remained largely unchanged for a decade
“The models themselves, like transformer architecture, which is the original model that, yeah, that was introduced in Hasn't changed very much, right? Like all, the whole industry is still using transformers. We've changed the way we train them, but the model a…”
Frosst: Data quality, not algorithms, is the primary bottleneck for AI progress
“But the algorithms I think are not the bottleneck in terms of making those models more useful. I do think a lot of it is still getting good quality data and then making good quality synthetic data From your good quality real data.”
Frosst: Current large language model technology will not lead to AGI
“I think when people are talking about building towards AGI like, I don't think this technology gets us there.”
Frosst: Creating the Best AI Model Requires Training Directly on Its Interface
“And if you want to make the best model for a given, you know, interface, it's best to be training the model on that interface.”
Frosst: Enterprise customers rarely request math reasoning capabilities from LLMs
“None of our customers ask the model to do math reasoning. That doesn't come up in the workplace that often that comes up in a few workplaces where mathematicians work, but there aren't a ton of people out there making a living doing math reasoning.”
Frosst: Enterprise clients will never demand Arc AGI pixel manipulation features
“Stuff like the Arc AGI challenge is a benchmark that people talk about, but that's like a pixel manipulation challenge. It's like, you know, taking in like a grid of pixels and based on rules, predicting the next one. That's not a thing any of our customers ha…”
Frosst: Benchmark scores measure training dataset overlap rather than true model utility
“They're a reflection of how much the model had been trained on those benchmarks.”
Frosst: AGI Hype Is the Most Damaging Rhetoric in AI
“I think the hype around AGI is the most damaging and confusing.”
Frosst: Claims of AI Posing an Imminent Existential Threat Are Incorrect
“The idea that, oh, this technology is like poses an imminent existential threat to humanity writ large was not a, was incorrect.”
Stebbings: AI agents will replace junior marketers and SDRs within 12 months
“Most 25, twenty-six-year-old marketing managers or SDRs, I'm sorry to say it, they're not brilliant. They do not love the craft. They are not better than a phenomenal agent will be in the next 12 months. They will be replaced.”
Frosst: Large language models have made zero independent scientific or intellectual breakthroughs
“There has been no independent breakthrough that an LLM has made. Right. There has been no, like nobody has seen, nobody asked in LLM, Hey, solve this problem. No one's solved before and get the answer. The breakthroughs are still people.”
Frosst: LLM architecture makes independent scientific breakthroughs fundamentally impossible
“No, no, that's not a matter of time. That's fundamentally the way that sequence models work. Like we are training statistical models of text.”
Frosst: Existential AGI risk rhetoric distracts from real issues like inequality
“I think those existential threat questions made it harder to talk about the real things, you know, like income inequality.”
Frosst: AI will exacerbate income inequality without proactive labor policy
“And I'm worried that technology has the potential to exacerbate that without being deployed correctly and without having good Yeah, without having good policy around, around employment.”
Frosst: AI founders obsessing over competitor benchmark gains distracts from customer needs
“I think there's a lot of people who suffer from way too much of it and obsessing over the minute details of like, you know, how, Well, so-and-so got .2% better on this thing, or like, you know, is constant, like constant small changes in businesses out there. …”
Frosst: Prompt engineering will be replaced by understanding how LLMs actually work
“So I think the idea of saying like, oh yeah, you got to learn how to prompt is going to go away. I think the idea of saying you need to learn how language models work and you need to know what they can and can't do in the same way you had to learn how a comput…”
Frosst: Cohere's Command A models are all built to fit on two GPUs
“Our model command day and the command day reasoning model, which we just released. Command A Vision model, which we just released. Those, they're all trained to fit on two GPUs.”
Frosst: Cohere spent orders of magnitude less on models than competitors
“We have spent Orders of magnitude less on creating foundational models than some of the other foundational model companies out there. Truly orders of magnitude less.”
Frosst: Enterprise workers do not need or want image generation models
“Nobody in the workforce is really wanting to generate images as part of their work for the most part. It doesn't, but as a consumer, it's very fun. It's very cool to be like, oh, here, give me a picture of this or something. So we, the types of models we train…”
Frosst: US and Chinese AI models cannot replace sovereign, culturally native models
“And I think just using a model that is built, you know, by China or built within America might not set your country and your economy up as well as having a model that understands the context built in that language, in that dialect, in like, you know, has the c…”
Frosst: Cohere's non-US identity is a major asset in enterprise sales
“So I think there's a lot of companies in Canada and around the world that are interested in working with non-American tech companies. And I would say that's been an asset for us, right?”
Frosst: AI language models are critical national infrastructure like power plants
“I think it's a good idea for countries to have infrastructure within their countries. Like, I think it's a good idea for people to have power plants in the country. You know, I like that Canada has several nuclear power plants and has several water power plant…”
Frosst: VR headsets will fail mass adoption due to real-world isolation
“And I think, you know, I was really excited about VR for a while. And then I realized I actually don't want to strap a computer to my face. I'm not interested in being disengaged from the world more. I want to be engaged in the world more than I am. I don't wa…”
Frosst: Cohere co-founders celebrate every signed funding round at McDonald's
“We go to McDonald's, yeah, after we signed it. The three of us go to McDonald's.”
Frosst: AI benchmark fixation is unhelpful for regulation because benchmarks are easily gamed
“Fixation on particular benchmarks, which can be gamed and can be trained either to do way better on or way worse on. Are not helpful for establishing how the technology can be used and misused.”
Frosst: China has not yet built AI models that beat top US models
“They haven't yet. Right. They've made good models. Definitely good models, but I don't think they've made models that are like beating, you know, the other models out there.”
Frosst: By 2026, AI models will autonomously perform tasks like filing expenses
“In 20, 26, you'll be able to open up a computer You know, log into North or whatever application you're using and say, file my expenses. And then the model will, yeah, figure out, you know, what expense policy it is and what the, where the photos are, like, do…”