Video · 2026-07-13 · 56m · 6 moments

Watching America Run Away With AI - Alistair Pullen (Cosine AI)

✦ AI generated

timeline · colored by role

Context

Getting a compute allocation from the UK government's sovereign AI unit on the Isambard supercomputer is what unlocked Cosine's ability to even attempt building a frontier-scale sovereign LLM, since compute is the single biggest blocker for a startup of their size.

Alistair explains that being inducted into the UK government's sovereign AI unit and getting compute allocation on the Isambard cluster removed the biggest barrier (compute cost) that would otherwise make a project of this scale impossible for a startup.

transcript

Alistair Pullen: one of the biggest blockers for like a startup of our size or smaller to be honest on being able to approach work like this uh is a huge part of it is the compute like if you raised you know 50 to 100 million bucks like a good chunk of that would go on compute on doing a project like this and to have allocation um come from the soft unit is is huge because it genuinely does enable it.

Mechanism

Cosine can compete with labs spending billions because it licenses model weights to customers rather than hosting inference itself, avoiding the massive inference-serving costs that are straining even Anthropic.

Alistair argues Cosine sidesteps the huge inference infrastructure costs that plague US labs because its business model licenses model weights for customers to run themselves, rather than serving tokens at scale.

transcript

Alistair Pullen: what that means in practice for cosign is that we license the technology that we build. We don't actually make a margin on tokens or anything like that... I think that one of the biggest reasons you've seen people like Anthropic struggle recently uh and the reason they've signed the deals they have with like the Colossus cluster and so on is because of inference and not because of training.

Example

Active parameter count, not just total parameter count, is a primary driver of real-world model quality, as shown by a small dense model outperforming a sparse MoE model with far fewer active parameters.

Comparing GPT-OSS 12B (about 5B active params) against the fully dense Devstral 2 123B, Alistair says the dense model feels dramatically better, attributing much of the gap to active parameter count per token.

transcript

Alistair Pullen: if you take at small scale something like a GPOSS 12B right that is... 5 billion active... and then you take a like devstrol 2 123b which is I believe like a similar architecture to llama 70b so it's fully dense... devstrol feels so much better than the GPOSS model... I think that a huge part of that just comes from the fact that one you have 120 billion active parameters per token, the other you have five.

gives example · 1

Mechanism◆

RL for coding agents wastes efficiency by weighting every token in a trajectory equally based only on a final pass/fail signal; credit should instead be attributed to the specific pivotal decision points, the way a teacher should mark exactly which sentences of an essay were good or bad rather than just returning a single grade.

Alistair uses an essay-grading analogy to argue that current RL wastes signal by treating every token in a long trajectory as equally responsible for the final reward, when fine-grained credit attribution to key decision points would make training far more efficient.

transcript

Alistair Pullen: say you were doing your English A level um and you'd written... a 2500word essay... 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.

extends · 1

Anecdote

Cosine's 'swarm' architecture, a hierarchical orchestration of hundreds of sub-agents (an orchestrator delegating to subplanners who delegate to workers), let a mid-tier open model complete a complex one-shot task—building a full mechanical-watch simulation SDK—that even frontier models like Opus and Gemini could not do alone.

Alistair describes Cosine's 'swarm' feature—up to 253 hierarchically organized sub-agents—successfully building a mechanical-watch compiler SDK in one shot, a task that even Opus and Gemini fail at without such orchestration.

transcript

Alistair Pullen: that is something that out of the box Kimmy cannot do... even Gemini 3.5 and Opus and 5.5 can't really do it. But as soon as you put them in a swarm... for that problem we use 253 sub agents which I think is more than you tend to see in a claude code session... when you do that it is possible and you can do that entire project in one shot.

extends · 1

Prediction◆

Recent US export controls and the sudden shutdown of a leading Chinese model made Alistair feel the UK has become a technological second-class citizen relative to the US, and this frustration is now driving Cosine's urgency to deliver a UK sovereign model no matter what.

Alistair admits that being cut off from cutting-edge US models makes him feel like a second-class citizen for the first time, and says this frustration is fueling Cosine's determination to deliver the UK sovereign model regardless of the obstacles.

transcript

Alistair Pullen: for the first time I feel like a secondass citizen because they are going faster than I am and I really hate that... that boils my blood more than anyone else. Um and I we are going to do everything we can to pull this off... we have no choice but to make it happen.

Highlight slides

RL Wastes Signal on Coding Agents✦ from: RL for coding agents wastes efficiency by weighting every token in a trajectory equally based only on a final pass/fail signal; credit should instead be attributed to the specific pivotal decision points, the way a teacher should mark exactly which sentences of an essay were good or bad rather than just returning a single grade.The Essay-Grading Analogy✦ from: RL for coding agents wastes efficiency by weighting every token in a trajectory equally based only on a final pass/fail signal; credit should instead be attributed to the specific pivotal decision points, the way a teacher should mark exactly which sentences of an essay were good or bad rather than just returning a single grade.Cut Off From the Frontier✦ from: Recent US export controls and the sudden shutdown of a leading Chinese model made Alistair feel the UK has become a technological second-class citizen relative to the US, and this frustration is now driving Cosine's urgency to deliver a UK sovereign model no matter what.Frustration Fuels UK Sovereign Push✦ from: Recent US export controls and the sudden shutdown of a leading Chinese model made Alistair feel the UK has become a technological second-class citizen relative to the US, and this frustration is now driving Cosine's urgency to deliver a UK sovereign model no matter what.

Related episodes

Video · The Cognitive Revolution

Getting a compute allocation from the UK government's sovereign AI unit on the Isambard supercomputer is what unlocked Cosine's ability to even attempt building a frontier-scale sovereign LLM, since compute is the single biggest blocker for a startup of their size.

Cosine can compete with labs spending billions because it licenses model weights to customers rather than hosting inference itself, avoiding the massive inference-serving costs that are straining even Anthropic.

Active parameter count, not just total parameter count, is a primary driver of real-world model quality, as shown by a small dense model outperforming a sparse MoE model with far fewer active parameters.

Recent US export controls and the sudden shutdown of a leading Chinese model made Alistair feel the UK has become a technological second-class citizen relative to the US, and this frustration is now driving Cosine's urgency to deliver a UK sovereign model no matter what.

Fable's Back, AI Engineer Recap, & SambaNova

AI Factories Explained: What’s Actually New?

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