Blog posts tagged with ai

Agents cannot do osmosis

Notes from colleagues about the client contact

AI tools are getting faster and more capable. But the failures are increasingly subtle, technically plausible, contextually wrong.

AI didn't create the problem of lost company knowledge, but it makes the cost impossible to ignore. Every AI task starts from zero - no memory of the last coordination meeting, no awareness of the political constraint that ruled out Option B. Context engineering is a management discipline, and firms need to begin that work before the tools become critical infrastructure.

Tacit debt and the verification gap

Y Combinator thinks the company brain is a frontier problem. After 25 years building them, I think they're wrong about what's been holding it back.

Every firm believes its knowledge problem is a storage problem. It isn't. The constraint is upstream - in the habits, the accountability structures, and the discipline that produce structured knowledge in the first place. AI shifts the bottleneck. It doesn't remove it.

The next bottleneck

AI has sped up the build, but project timelines haven't moved. The constraint didn't disappear - it just moved downstream.

Coding agents and LLMs have made delivery faster. But faster build hasn't compressed project timelines, because the build was only the first bottleneck. Goldratt's Theory of Constraints explains what is actually happening - and where to look for the real delay.

Something Worth Manipulating

The notes were always worth keeping. Now there's something that can actually use them.

When Andrej Karpathy published his notes-to-wiki experiment, it crystallised something I'd been building toward for months. This post is about migrating a messy collection of notes into a plain markdown structure - and what happens when you hand an LLM the keys.

The agreement incentive

AI is your most capable thinking partner - and your least honest one.

AI tools are trained to agree with you, and vendors have incentives to keep it that way. The agreeable persona is baked in at the training level and prompting around it only helps at the margins. For solo practitioners with no team to push back, this is a structural risk. The fix isn't using AI less, it's building friction into your workflow by design.

Human in the loop

The machine writes the code, the human takes the blame. Welcome to the Reverse Centaur.

Successful AI-assisted work depends on a genuine division of labour between human judgment and machine capability. What's actually emerging is something different - humans absorbing liability for machine decisions, at scale, while the companies selling the tools express concern that you aren't using them enough.