Our Big Dumb AI Gods Were Wrong

Jev was released to much fanfare on September 20, 2026, by Typesafe.ai. It’s one of those moments where there’s a “before” and an “after”. For years, now, we’ve been told that we must have more compute, more storage, more memory, and more datacenters. More, more, more… the AI juggernaut has an insatiable hunger for all things infrastructure. This was the lie promulgated by Sam, Amodei, and to some extent, even Jensen. Jev shows us that there’s another way to do AI. If we focus on automation, trustworthy workflows, and other facets of “composable AI”, then we can actually build agentic systems that work reliably.

One gets the sense that the Sams and the Amodeis of the world weren’t actually enthusiastic about building automation or composable AI – they were obsessed with building their AI sky buddy, the big dumb AI god, the one that would transcend humanity and usher us into the age of the singularity. For those of us who could not give less of a shit about singularities and following cult leaders down a path of doom, we were… less than thrilled about their worldview. Jev has all the markings of making the cultish ravings of the AI-pilled seem, dare I say it, anachronistic and outmoded. In one fell swoop, whatever OpenAI and Anthropic are working on became yesterday’s news.

Almost immediately after the Jev release, there was an onslaught of small decisioning models designed for specific domains of knowledge. The vLLM semantic router project released Open Decision Foundation Models, covering different aspects of data routing decisions. Convai released Laya, an “open source Jev alternative”, along with several other Jev-esque clones in the “decision model” category. Things will not be the same after this. We seem to be transitioning into a UNIX primitive phase of AI, to quote Rachel Chalmers, where lots of autonomous decision bots trained to do one thing, or a few well-defined things, will be working in consort under a common SDLC framework (read: agent swarms) decided on by whoever happens to be running platform engineering. These decision models will follow the old UNIX mantra “do 1 thing, and do it well”. Applications will be composed by piecing them together, hence the term “composable AI”.

For the last 4 years, we’ve been told that the only way to do anything useful with AI was to keep expanding the context window, keep adding more petabytes of data, and keep expanding until – seemingly – the world consists of nothing besides datacenters. And when anyone pointed out the shortcomings of the economics behind the build out, they were shouted down and lectured about “scale”. Yes, Claude Opus and other genAI tools are built for something very different from Jev and it’s ilk, but once a big chunk of use cases have been taken out of your purview, it does lead to the question of how, exactly, the frontier labs are going to justify their massive capital expenditures. And if I were an investor, I would be more than a little concerned about the upcoming IPO.

The response from the frontier labs was that they could take reams of unstructured data and turn it into something that produces whatever you need, be it code, written content, or pretty pictures. But now we know that if you take the time to structure the incoming data, the compute cost for running inference is drastically reduced. Again: what do we need this massive buildout for? What are we getting out of it? Given where technology has been headed for the last two decades, we know the most likely answer: massive surveillance used to sell advertising and other e-commerce and marketing tech. If you strip that away, in addition to the decision automation referenced above, how much is really left? I like generating white papers, videos, and email as much as the next person, but how long can those use cases justify the ongoing expansion? Especially when the economics model behind them is based on advertising sales.

We’re just beginning this journey towards the UNIX model of agentic systems. It’s going to be a fun ride, and I, for one, intend to enjoy it.