Video · 2026-07-13 · 56m · 6 moments

Britain Is Losing The AI Race To America - Alistair Pullen (Cosine AI)

✦ AI generated

timeline · colored by role

Mechanism

Cosine can compete with vastly better-funded US labs because it licenses model weights for customers to run on their own infrastructure rather than hosting inference itself, so it avoids the massive data-center inference spend that burdens companies like Anthropic.

Pullen explains that Cosine avoids the huge inference/data-center spend dominating US labs' budgets because it licenses model weights for customers to deploy themselves rather than serving tokens through its own hosted API.

transcript

Alistair Pullen: Most of the time now nearly all of the time we are not hosting the model ourselves. A customer isn't hitting you know cosign slash you know APIVv1 and then hitting a chat completions endpoint or something like that from us. They are either taking the model weights um that we give to them deploying them on their own GPUs.

supports · 1

Data

Based on independent analysis, Anthropic's Sonnet model likely has around 1.3-1.5 trillion total parameters with 100+ billion active, while Opus is likely 1.5-1.8 trillion total with 150-180 billion active parameters.

Drawing on an outside empirical analysis, Pullen estimates Sonnet at ~1.3-1.5T total parameters (100B+ active) and Opus at ~1.5-1.8T total (150-180B active) — numbers that directly informed Cosine's own architectural decisions for its sovereign model.

transcript

Alistair Pullen: the likelihood is and like again no one really knows outside of the outside of anthropic that something like sonnet is in the I believe like 1 1.3 to 1.5 trillion total params um and has probably 100 100 plus billion active right and then an opus is probably in the 1.5 to 1.8 eight range and probably has 150 to 180 billion active

Mechanism◆

Standard RL training rewards an entire trajectory as pass/fail without indicating which specific decisions were good or bad, which is like a teacher grading an essay 'B' without saying what made it a B — attributing credit to the specific pivotal decisions within a trajectory would make RL training far more efficient.

Pullen argues today's RL rewards treat every token in a long rollout as equally responsible for the final outcome, which wastes compute; isolating and grading the specific high-stakes decision points in a trajectory — like a teacher circling the exact sentence that earned a grade instead of just handing back a B — would make training dramatically more efficient.

transcript

Alistair Pullen: say you were doing your English A level and you'd written a 2500 word essay on a question and the teacher just gives you right okay that's a B thank you so much and you're like I don't what made it a B... it would be far easier if the teacher just sent, you know, circle the sentence and be like, this is rubbish. Don't say this, right? And that is fundamentally the principle we're trying to bring into like RL across the board.

supports · 1

Example

Cosine's 'swarm' feature — a hierarchical orchestration of up to 253 sub-agents, with an orchestrator delegating to subplanners who delegate to workers — let a model with no horology knowledge build a complete, physically-accurate mechanical-watch design SDK with a 3D viewer in one shot, something no single frontier model could do out of the box.

Pullen describes Cosine's "swarm" — an orchestrator that breaks a task into subplanner-owned verticals which delegate further to worker agents (253 in one case) — successfully building a full mechanical-watch design SDK with a physics-accurate 3D viewer in one shot, a task no single model, including Opus or Gemini, could do alone.

transcript

Alistair Pullen: A swarm is genuinely like a real a lot of a lot of sub agents running at the same time in a hierarchical way... for that problem we use 253 sub agents which I think is more than you tend to see in a claude code session and so on... when you do that it is possible and you can do that entire project in one shot.

Claim

Every current approach to AI agent memory is fundamentally a hack — essentially a vector database or bolted-on tool — and it's hard for agents to know when to query it, whether it's stale, or whether to trust it over fresh reasoning.

Pullen says memory remains one of the hardest unsolved problems in agentic AI — every current implementation is essentially a hacky external tool, like a vector DB, that agents struggle to query correctly or keep updated, sometimes overriding correct behavior with stale stored information.

transcript

Alistair Pullen: Yes, memory is very hard to get right. Um, we've tried a bunch of different approaches. Um, fundamentally, I think every approach to memory that exists right now is a bit of a hack, right? It's like a tool. Um, and it's in many cases like a vector DB or like an embedded version of some tidbit of knowledge, but they're very hard for the agents to know when to query.

Anecdote◆

For the first time in my life, I feel like a second-class citizen because people in America have better AI and are moving faster than I am, and that makes me furious.

Pullen says US export controls (like the sudden banning of the Fable model) and America's faster frontier progress have, for the first time, made him feel like a second-class citizen in AI.

transcript

Alistair Pullen: Yeah. I mean for the first time I feel like a second-class citizen because those those folks over there in America they've got better AI than I do. They are going faster than I am and I really hate that.

explains mechanism · 1

Highlight slides

RL Rewards Whole Trajectories, Not Decisions✦ from: Standard RL training rewards an entire trajectory as pass/fail without indicating which specific decisions were good or bad, which is like a teacher grading an essay 'B' without saying what made it a B — attributing credit to the specific pivotal decisions within a trajectory would make RL training far more efficient.The Essay-Grading Analogy✦ from: Standard RL training rewards an entire trajectory as pass/fail without indicating which specific decisions were good or bad, which is like a teacher grading an essay 'B' without saying what made it a B — attributing credit to the specific pivotal decisions within a trajectory would make RL training far more efficient.Credit Assignment Would Boost Efficiency✦ from: Standard RL training rewards an entire trajectory as pass/fail without indicating which specific decisions were good or bad, which is like a teacher grading an essay 'B' without saying what made it a B — attributing credit to the specific pivotal decisions within a trajectory would make RL training far more efficient.Feeling Like a Second-Class AI Citizen✦ from: For the first time in my life, I feel like a second-class citizen because people in America have better AI and are moving faster than I am, and that makes me furious.Root Cause: Export Controls✦ from: For the first time in my life, I feel like a second-class citizen because people in America have better AI and are moving faster than I am, and that makes me furious.

Related episodes

Video · The Cognitive Revolution

Fable's Back, AI Engineer Recap, & SambaNova

→

Video · The Cognitive Revolution

LTX CEO on Video Gen + Can we beat AI superforecasters?

→

Audio · All-In Podcast

Elon's Anthropic Deal, The Next AI Monopoly?, "FDA for AI" Panic, Trading the AI Boom

→