Video · 2026-07-13 · 56m · 6 moments

Why a Nation Can't Outsource Its Frontier AI - Alistair Pullen (Cosine AI)

✦ AI generated

timeline · colored by role

Fact

Government-backed access to compute on the Isambard supercomputer cluster is what unlocked Cosine's ability to even attempt building a frontier sovereign LLM, since compute is the single biggest blocker for a startup of its size.

Alistair explains that being inducted into the UK government's sovereign AI unit and getting an allocation on the Isambard cluster removed the compute-cost barrier that would otherwise make a project of this scale impossible for a company Cosine's size.

transcript

Alistair Pullen: what that looks like in practice for us is um an allocation of compute on the ismpod supercomputer cluster out in Bristol. Uh and that has allowed us to well honestly it's one of the things that's unlocked our ability to even like have the ambition to do something like this right um fundamentally 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

supports · 1

Mechanism◆

Cosine can compete with US labs on a fraction of their budget because it licenses model weights for customers to run themselves rather than hosting inference, avoiding the massive data-center spend that consumes most of a frontier lab's capital.

Answering how Cosine can do with millions what US labs do with billions, Alistair says Cosine licenses its technology instead of selling inference tokens, so it avoids the huge data-center costs that dominate spending at labs like Anthropic.

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. Um and all of this ties into your question meaning um we don't have to spend a lot of the money that the Americans are having to spend on data centers for inference purposes.

supports · 1

Example

Active parameter count per token, not raw total parameter count, is the main driver of how good a model feels in practice — a fully dense ~123B model with far more active parameters clearly outperforms a sparse MoE model with a similar or larger total size but only ~5B active.

Comparing GPT-OSS-120B (roughly 5B active parameters) against the fully-dense Devstral-123B, Alistair says the difference in real-world feel is night and day, illustrating that active parameters per token matter more than total model size.

transcript

Alistair Pullen: if you take at small scale something like a GPOSS 12B right that is ane I believe it's 5 billion active I can't remember it was a while ago but it was something like that 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 I don't know if you've used them back to back before devstrol feels so much better than the GPOSS model, right?

Mechanism◆

Standard RL treats every token in a long trajectory as equally responsible for the final pass/fail reward, which is nonsensical because only a handful of pivotal decision points actually determined the outcome — a flaw that helps produce slop.

Alistair argues that current RL for coding agents rewards entire trajectories uniformly based on the end result, which he calls 'insane' since only a few high-stakes decisions in a rollout actually mattered — motivating Cosine's work on credit attribution to reduce slop.

transcript

Alistair Pullen: fundamentally this notion of you have a roll out of maybe 256,000 tokens, right? In some extreme cases. Um and that culminated in like a one or a zero depending on what the model did. And what we're saying at the moment in many cases is okay, all of those tokens are equally weighted in getting us to that answer.

Example

Cosine's 'swarm' — an orchestrator delegating to sub-planners who delegate to up to 253 worker sub-agents — can one-shot a complex, novel project that no single frontier model can do out of the box.

Alistair describes Cosine's 'swarm' system — one orchestrator breaking work into sub-planner verticals that delegate to a flat layer of workers (253 in one case) — successfully one-shotting a from-scratch mechanical-watch SDK project that base models like Opus or Gemini can't do alone.

transcript

Alistair Pullen: as soon as you put them in a swarm and for what swarm looks like for cosign is you have one orchestrator at the very top it breaks down a problem into um sub problems for uh basically product managers or or or whatever you want to call them. We call them subplanners but they own verticals of this.

gives example · 2

Claim◆

Now that the US is visibly pulling ahead in frontier AI, the UK sovereign AI effort has no choice but to succeed rather than rely on outsourcing.

Reflecting on the gap opened up by newly announced US models, Alistair says the situation 'boils his blood' more than anyone's and insists Cosine has no choice but to make the UK's sovereign AI project succeed.

transcript

Alistair Pullen: And and and and believe me that boils my blood more than anyone else. Um and I we are going to do everything we can to pull this off. Um you mentioned like how are you going to do this? It's like we are just going to we're just going to make it happen. Um we have no choice but to make it happen.

Highlight slides

How Cosine Competes on a Fraction of the Budget✦ from: Cosine can compete with US labs on a fraction of their budget because it licenses model weights for customers to run themselves rather than hosting inference, avoiding the massive data-center spend that consumes most of a frontier lab's capital.Licensing vs. Inference Business Models✦ from: Cosine can compete with US labs on a fraction of their budget because it licenses model weights for customers to run themselves rather than hosting inference, avoiding the massive data-center spend that consumes most of a frontier lab's capital.The Credit Assignment Problem✦ from: Standard RL treats every token in a long trajectory as equally responsible for the final pass/fail reward, which is nonsensical because only a handful of pivotal decision points actually determined the outcome — a flaw that helps produce slop.Why That's 'Insane'✦ from: Standard RL treats every token in a long trajectory as equally responsible for the final pass/fail reward, which is nonsensical because only a handful of pivotal decision points actually determined the outcome — a flaw that helps produce slop.Cosine's Response✦ from: Standard RL treats every token in a long trajectory as equally responsible for the final pass/fail reward, which is nonsensical because only a handful of pivotal decision points actually determined the outcome — a flaw that helps produce slop.The Gap That 'Boils His Blood'✦ from: Now that the US is visibly pulling ahead in frontier AI, the UK sovereign AI effort has no choice but to succeed rather than rely on outsourcing.No Choice But to Succeed✦ from: Now that the US is visibly pulling ahead in frontier AI, the UK sovereign AI effort has no choice but to succeed rather than rely on outsourcing.

Related episodes

Video · The Cognitive Revolution

Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

→

Video · a16z Podcast

Software in the Age of Agents | The a16z Show

→

Video · a16z Podcast

Ben Horowitz on the Global Race for Tech, Power, and Influence

→