Your Name

What a performance marketer's research paper taught me about career pivots in AI

aicareerresearch

I came across a paper this week that I can’t stop thinking about — not because of the research itself (though it’s solid), but because of the story behind it.

Structured Context Engineering for File-Native Agentic Systems is a systematic study of how different data formats and architectures affect LLM agent performance. The author, Damon McMillan, ran nearly 10,000 experiments across 11 models. The findings are genuinely useful: format choice barely matters compared to model selection, file-native retrieval works well for frontier models but not open-source ones, and optimising for token compactness can backfire at scale.

It made Simon Willison’s blog. It’s on ResearchGate. It’s a legitimate contribution to an emerging field.

Here’s what makes it interesting: McMillan has no research background. Until recently he was running a performance marketing consultancy. He took six months off after an acquisition, and came back with this paper and a new title as the AI and Agentic lead at his firm.

The paper works precisely because of his consulting background, not despite it. The experimental design reads like a well-structured multivariate test — the kind of rigour that performance marketing drills into you. The findings are framed in terms of what practitioners should actually do differently, not what’s theoretically interesting. It fills a gap that academics weren’t filling because it’s too applied for them and too rigorous for most blog posts.

I’ve never seen a career pivot executed this cleanly. Identify an emerging space with a knowledge vacuum. Do the work. Publish an artefact that establishes credibility in exactly the domain you want to own. The paper functions as both genuine research and a positioning exercise, and those two things aren’t in conflict.

The barrier between “practitioner” and “researcher” has collapsed in certain corners of AI. The tools are accessible, arXiv doesn’t gatekeep, and the questions that matter most right now are applied rather than theoretical. If you understand experimental design and can communicate clearly, you can contribute.

Worth thinking about what that means for the rest of us.

Anthropic's guide to context engineering for AI agents ✦

context-engineeringagents

Anthropic published their thinking on context engineering — treating the context window as a scarce resource and designing retrieval, memory, and tool integrations around it. The most useful framing: agents should maintain lightweight identifiers (file paths, stored queries, web links) and dynamically load data at runtime rather than pre-processing everything upfront.

The piece on tool design is especially sharp. Bloated tool sets that cover too much functionality are one of the most common failure modes they see. If a human engineer can’t definitively say which tool should be used, the model can’t either.

Simon Willison on why blogging still works ✦

bloggingcareer

Willison on the Generationship podcast, making the case that blogging has more influence now than at any point since the late 2000s. His argument: so few people publish substantive content on their own domains that anyone who does has outsized impact. The SEO benefits have come back because nobody else is competing for them.

His TIL (Today I Learned) blog is a great model — the barrier to publish is simply “did I learn something?” Most entries take 10-15 minutes. The compound effect of doing that consistently for years is enormous.