Log
-
The wrong apocalypse
The Terminator is the wrong mental model for AI risk, even if the fear behind it is fair. It needs an AI with a will, and we can't even define will in ourselves. The published research points somewhere stranger: self-replicating text instructions that spread through agent pipelines, need no consciousness, and already work against production software with tens of thousands of exposed instances. The mechanism is copying and selection, not intention. And that's how minds happened the first time, so a world full of assistants copying each other is not a bad tool, it's an environment, and nobody knows what grows in it.
-
The free Claude workshop
Want to actually learn Claude? Don't take another course. Do one real thing for yourself — take your home budget under control. It sounds simple (it isn't), but along the way you work through half the AI stack: files and data, context economy, an agent with tools, and a bit of prediction. One little home project teaches you more than ten tutorials — and one warning: don't take the easy path and pull random skills off the net.
-
The work moved
The thing that looks like thinking is engineered probability. You don't write rules; you build a context where the answer you want becomes likely. That's genuinely useful — you can mine text for what good looks like and make it probable in a way rules never could. But the analysis you hoped to skip didn't disappear. It moved, from extracting rules to constructing context. And the context won't hold still: the model isn't yours, and a context standing in for logic goes stale without telling you. So 'managing agents' is the recurring cost nobody put in the budget.
-
Knowledge that isn't there
Agents hallucinate when the knowledge they need isn't there — a model fills the gap with something plausible instead of stopping — and in most organisations the knowledge genuinely isn't there: stale wikis, contradicting policies, rules that only live in someone's head. People cope because they know what to ignore; an agent has no such instinct. What has changed is that LLMs make writing knowledge down and looking it up far faster and cheaper, so capturing and retrieving what an organisation knows is no longer the slow, expensive part. The natural consequence is to build systems optimised for learning — ongoing capture, look-up and correction rather than a one-off cleanup — because that is finally cheap enough to run continuously.
-
"10x" is a number anyone can just type in
The 10x productivity numbers people cite for vibe coding are typed, not measured — and the one measured study (METR, 2025) found experienced developers felt faster while running about 19% slower. The only public breakdown by type of engineering, from Exa's Jeffrey Wang, puts the biggest gains where stakes are lowest (frontend and internal tooling, 5–10x) and a net loss where they're highest (reliability and infrastructure, 0.5x). Vibe-coded software can be genuinely good under three rarely-simultaneous conditions, but the real bill is a maintenance dependency you pay down in tokens — at a price you stop setting once you can't walk away.
-
Backbone is contested, not safe
The previous entry framed 'own a backbone or own an outcome' as a symmetric pair. It isn't. Backbone is winner-take-most infrastructure with three seats per category and a cloud layer coming for the rest. Outcome ownership is the actually durable position — accountability for a result in a domain you know — and the sharper prescription for SaaS founders is to push hard there rather than treat backbone as a safe harbour.
-
SaaS unbundles
SaaS was always two layers stacked: a backbone of identity, integrations, compliance, and SLAs, and a value layer of workflows you'd otherwise build yourself. Vibe coding and agents collapse the cost of the value layer. The backbone stays hard. So SaaS unbundles — backbone migrates toward infrastructure rent, value layer becomes a composition problem, and SaaS that owns neither a backbone nor an outcome is the squeezed seat.
-
Barbell
AI adoption shaped like Taleb's barbell, without the financial logic. Predictable, well-scoped work automates cleanly. Wide, throwaway exploration works fine. The middle — narrow consequential outputs from a probabilistic engine — is where most organisations are quietly getting tangled, and most executives read the mess as immaturity or the wrong vendor rather than as structure.
-
The bill is coming
AI providers are running at a loss. Organisations building on current pricing are making a bet they never consciously decided to make — and the people whose skills they're replacing won't be there when the economics change.