Alistair Pullen

What they argue

Alistair Pullen is an AI thinker who focuses on the architectural and strategic shifts required to build specialized, high-utility AI systems. He explores the transition from simple chat interfaces to complex agentic systems, focusing on model fine-tuning, sovereign AI strategies, and the technical levers of the developer platform.

1. Platform-Based Development

   OpenAI is shifting toward a platform model where research is abstracted into APIs. This allows developers to build high-fidelity, specialized applications using tools like Realtime APIs and Model Distillation.

   From Building AGI in Real Time (OpenAI Dev Day 2024) (episode)

   • Says much the same Sam Altman — The Platform-First ModelOpenAI is shifting from being a primary creator of benefits to a provider of tools that allow others to create benefits. This positions the AI as the 'engine' or…GPT-4o
   • Says much the same Greg Eisenberg — The Shift to Personal Agent PlatformsThe AI landscape is shifting from a 'chat' interface to a 'personal agent' platform, exemplified by OpenAI's Dots. This creates a massive distribution channel where…OpenAI DevDay: Dots, Agents & $100B Opportunities

2. Realtime API and Persistent Sessions

   The Realtime API uses persistent WebSocket connections to enable continuous, low-latency voice and text interaction. This removes the friction of traditional back-and-forth turn-taking.

   From Building AGI in Real Time (OpenAI Dev Day 2024) (episode)

3. Process over Product Training

   Training data should focus on the incremental knowledge discovery and step-by-step decision-making of a human engineer. This includes capturing intentional errors and iterative corrections rather than just providing final working code.

   From Is finetuning GPT4o worth it? — with Alistair Pullen, Cosine (Genie) (episode)

4. The Genie 4-Stage Workflow

   To build an autonomous agent, a workflow should mimic an 'LLM OS' stack. This includes planning, code retrieval, execution, and verification.

   From Is finetuning GPT4o worth it? — with Alistair Pullen, Cosine (Genie) (episode)

5. License, Not Inference Model

   Smaller entities can compete by selling model weights rather than providing hosted APIs. This allows customers to deploy weights on their own GPUs or private clouds to remove inference costs.

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

   • A related idea Dylan Patel — Commoditization of InferenceSoftware optimizations for running models are increasingly becoming open-source and a commodity, making them harder to defend as a unique advantage. Because the…No Priors Ep. 127 | With SemiAnalysis Founder and CEO Dylan Patel

6. Architectural Differentiation

   The path to sovereign AI for smaller entities is found in architectural differentiation rather than competing on raw inference scale. This involves utilizing sovereign compute and specialized Reinforcement Learning (RL) for domain-specific models.

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

7. Sovereign AI Competitive Framework

   To compete with billion-dollar labs, entities should follow three pillars: infrastructure leverage, architectural differentiation, and specialized RL. This includes securing government-backed compute to offset a lack of private capital.

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

   • Says much the same Dylan Patel — The Infrastructure MoatThe primary competitive advantage in the current AI era is the ability to operate and own highly efficient, custom-built infrastructure that competitors cannot match.…The State of Silicon and the GPU Poors - with Dylan Patel of SemiAnalysis
   • A related idea Simon Willison — The Three Pillars of CustomizationTo build successful AI applications, developers must leverage three distinct layers: the System API for stateful interactions, Dialogue, and other layers of…AGI is Being Achieved Incrementally (OpenAI DevDay w/ Simon Willison, Alex Volkov, Jim Fan, Raza Habib, Shreya Rajpal, Rahul Ligma, et al)

Every idea names the summary it came from. Written by a local model from this mind’s own work, never from memory.

Interviews 3

• Is finetuning GPT4o worth it — with Alistair Pullen, Cosine (Genie) · Latent Space — To build a world-class coding agent, you must move beyond prompting and fine-tune models on the actual 'process' of…
• Building AGI in Real Time (OpenAI Dev Day 2024) · Latent Space — OpenAI is shifting from a research-first laboratory to a developer-centric platform, prioritizing real-time…
• Watching America Run Away With AI - Alistair Pullen (Cosine AI) · Machine Learning Street Talk — Alistair Pullen explains how Cosine AI is building a sovereign UK LLM by leveraging a 'license-not-inference' business…