Nov 21, 2024 · 44m · no-priors
No Priors Ep. 91 | With Cohere Co-Founder and CEO Aidan Gomez
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
Cohere Co-Founder and CEO Aidan Gomez discusses the evolution of Transformer architectures, Cohere's enterprise-focused strategy, and the economics of training frontier models. He offers deep insights into scaling laws, inference-time reasoning, and why foundation models remain far from commoditization.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 17.6% of the talking time here. How this is scored →
The hosts as informed peer 5.3 Guest teaching 4.3 Guest disagreement 2.3 The hosts pushing back 3.0
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 21:44 Aidan flatly rejects frontier lab pre-training consensus
When Sarah notes frontier lab leaders claim nobody outside AGI labs should touch pre-training, Aidan bluntly dismisses the claim as empirically wrong.
Hardest push from the hosts ▶ 20:00 Sarah challenges enterprise pre-training feasibilitySarah directly challenges Aidan's framework by arguing that pre-training for enterprises is controversial and rejected by leading AI labs due to compute and data curation limits.
Biggest teaching moment ▶ 24:24 Aidan explains limitations of frozen open-source weightsAidan educates on why open-source models like Llama cannot simply replace proprietary base training, explaining that frozen, cooled-down models with zero gradients offer fewer levers than vertical training data integration.
The host holds their own ▶ 28:35 Sarah synthesizes the economic model shift of reasoning computeSarah demonstrates sharp industry insight by translating the technical concept of inference-time compute into a structural shift from upfront CapEx model training to consumption-based intelligence.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Founding Cohere and the Enterprise-First Mission | 4 | 2 | 1 | 1 | Sarah introduces Aidan and sets up his background from Google Brain to founding Cohere. Aidan amicably explains the founding rationale, targeting enterprise workloads rather than building a consumer ChatGPT competitor. | |
| Resolving Enterprise Pitfalls and Practical AI Use Cases | 4 | 4 | 1 | 2 | Sarah asks Aidan to make enterprise frustrations concrete. Aidan details common RAG pitfalls, model sensitivity to prompt formatting, and walks through practical vertical deployments like longitudinal healthcare record synthesis. | |
| Enterprise Adoption Strategy and Overcoming Adoption Barriers | 5 | 4 | 2 | 3 | Sarah questions the long-term equilibrium between specialized AI apps and in-house enterprise development, as well as whether enterprise AI is entering a trough of disillusionment. Aidan reframes the market using an adoption pyramid and argues the technology is too early for a disillusionment slump. | |
| Model Customization Tiers and Cohere's Cost Efficiency Strategy | 7 | 6 | 5 | 6 | Sarah pushes back against the premise of enterprises doing pre-training, citing frontier lab consensus, and challenges Cohere's supercomputer capital expense given open-source models like Llama. Aidan forcefully rejects the lab consensus as empirically wrong and explains the limitations of fine-tuning zero-gradient open-source weights. | |
| Scaling Laws, Inference-Time Compute, and Reasoning Paradigms | 6 | 5 | 1 | 2 | Aidan explains the plateau in generic vibe checks and the shift toward inference-time compute and reasoning models. Sarah demonstrates domain expertise by framing test-time compute as a transition from CapEx-driven capability gains to consumption-driven models. | |
| Data Limits, Scientific Frontiers, and AGI Realism | 6 | 5 | 4 | 4 | Sarah probes the limits of sequence-to-sequence scaling, AGI discrete milestones, and model representation bottlenecks. Aidan rejects sci-fi takeoff scenarios and checklist AGI definitions, while debunking model commoditization as temporary loss-leading price dumping. |
Statements from this episode (18)
Cohere Will Not Build a Direct ChatGPT Competitor
“We're not going to build a ChatGPT competitor. What we want to build is a platform and a series of products to enable enterprises to adopt this technology and make it valuable.”
Overestimating LLM Flexibility Causes Repeated Enterprise RAG Failures
“Well, I think all language models are quite sensitive to prompts, to the way that you present data. They all have their own individual quirks. The way that you talk to one might not work for the way that you talk to another. And so when you're building a syste…”
Enterprises Should Buy Commodity AI and Build Only Proprietary Differentiators
“What we've pushed organizations to do is have a strategy that encompasses that full pyramid. Yes, you need the generalist standard stuff. Maybe there's some industry specific tools that you can go out and buy, but then if you're building, don't build those thi…”
AI Developer Adoption Will Take Two to Three Years to Permeate
“Eventually developers will become more familiar with building with this technology. But I think it's going to take another two or three years before it really permeates.”
Current LLM Tech Alone Requires Five Years of Economic Integration Work
“Even if we didn't train a single new language model, like, okay, all the data centers blow up. We can't improve the LLM. We only have what we have today. There's a half decade of work to go integrate this into the economy, to build all these things, to build t…”
Fine-Tuning Cannot Effectively Add New Languages to Large Language Models
“There's just no way you can do that without intervening on pre-training. You can't like fine tune or post train Japanese into a model effectively. And so you have to start from scratch.”
GPT-4 Class Enterprise Models Now Cost $10M to $20M to Train
“What we've seen is today you can build a model that's as good as GPT-IV in all the things that enterprises might care about. For ten million dollars, twenty million dollars, like just orders of magnitude less than what was spent to develop that model. And so i…”
Cohere Strategically Lags Frontier Labs by Six Months to Avoid $7B Burn
“So that's the strategy is don't lead. Don't burn, you know, three, five, seven billion dollars a year to be at the front, be six months behind. And offer something to market to enterprises that actually fits their needs at a price point that makes sense for th…”
Fine-Tuning Open Source Models Lacks Levers of Full Vertical Training
“Taking those models and trying to fine tune them It's just, it's not as effective as building it yourself and you have much fewer levers to pull than if you actually have access to the data and you can change the data that goes into that process.”
AI Scaling Curves Are Flattening and Casual Vibe Checks Are Failing
“We're starting to enter into a sort of flat part of the curve and we're certainly past the point where if you just interact with a model, You can know how smart it is. Like the vibe checks, they're losing utility.”
Inference-Time Compute Lets Labs Scale Intelligence Without Doubling Supercomputers
“I don't need to go double the size of my supercomputer to hit a requisite intelligence threshold. I can just double the amount of inference time compute that my customers pay for.”
Inference-Time Compute Does Not Require Densely Interconnected Supercomputers
“If we have a new avenue, which is inference time compute, That doesn't require this densely interconnected supercomputer. It's fine to have nodes. You can do a lot more locally and less distributed.”
Reasoning Models Will Become Highly Robust Within Two to Three Years
“I think right now it's extremely inefficient and it's quite brittle, similar to the early versions of language models. But over the next two or three years, it's gonna become incredibly robust and unlock just a whole new set of problems.”
AI Will Eventually Run Experiments, but Scaling Will Take Many Years
“At some stage we're gonna have to give these models the ability to run their own experiments to fill in areas of their knowledge that they're curious about. But I think that's quite Quite a ways away. And it's going to be tough to scale that. It will take many…”
Humanity Will Achieve Broadly Capable AGI but Will Not Build God
“We will build AGI if what you mean is very useful, generally capable technology that can do a lot of the stuff that humans can do and flex into a lot of different domains. If what you mean is, you know, are we gonna build God? No.”
Falling AI API Prices Stem From Price Dumping, Not Commoditization
“I don't think that models are actually getting commoditized. I think what you see is you see price dumping. And so you see people giving it out for free, giving it out at a loss, giving out at zero margin. And so they see the prices coming down and they assume…”
Global AI Refactor Will Take 15 Years With Few Key Players
“I think in reality, the state of the world is there's a total technological refactor that's going on right now and will last the next 10 to 15 years. And it's kind of like we have to repave every road on the planet. And there's like four or five companies that…”