On Jev, the capabilities overhang, and imagination + amenability as the bottlenecks
September 2026Around 2014, a team at Google led by Scott Jenson started working on something called “The Physical Web”.1 The idea was as follows - BLE (Bluetooth low energy) beacons, NFC tags and QR codes were becoming cheap and so any object could broadcast a link. And any link was just a website that could utilise all the wonderful things the web can afford you. You would walk up to a bus stop and your phone would offer you a small web page with that stop's timings, some local information, and suchlike.
It didn't survive. Google shut it down a few years later, partly because of something like “hey Maps already knew the bus times”, generalised across many existing services.
But I think the instinct behind it is one that we can generalise: once a technology becomes mature enough and general enough,2 the obvious next move is to try it everywhere and see where it sticks. The web by then, a few years into smartphones, etc. had become a rich, general runtime, and the Physical Web was an attempt to push that runtime out into physical space.
It failed, but it was the first time I viscerally felt and appreciated the idea of taking a sufficiently general thing to its natural end, which is everywhere.
Every now and then, you find a thing that naturally fits into almost everything else, and when you do, the only correct response is to go looking for all the places it fits.
The electric motor was something similar: a general thing that, in hindsight, belonged in nearly every piece of steam or hand-driven apparatus in factories and homes.
I think we're still early in the same process with LLMs. We're already seeing it in some capacity: LLM-style models are taking over recommendation systems3 that used to be a patchwork of ad-hoc models, and customer support bots that used to be something approximable to if-else trees. In both cases the old system was a hand-built guess at what the user meant, and LLMs are simply better at that. Wherever the old system was trying to understand some variant of “intent”, a language model does better, as it operates on that general layer of language instead of operating on a surrogate layer that tries to capture the dynamics of the underlying one.
All the Jev4 demos motivated me to finally post this idea that I’ve been thinking about for a while. You give Jev some context state and a bunch of questions ("which direction?", "how dangerous is this?", "which of these available tools?"), and it answers them all at once as probabilities, with extremely low latency. People have made Postgres extensions that filter rows by plain English descriptions, and browser agents that pick what to click. In each case the model only makes the judgement call, and ordinary code does everything else. The reason Jev feels so special is that you’re essentially pretraining a language model, loading it with all the intelligence one gets from learning facts and contents from a corpus like webtext (the Jev team says it had a synthetic pipeline), and then making it such that you have an if-else statement that is imbued with all that intelligence.
Another example that fits in is the recent imagegen work from the Reve team5: Imagegen models pass the model's intent to the renderer as a latent; the Reve team augmented it with a scaffold: a list of regions, each with a bounding box and a written description. An LLM can write that natively, so all the planning moves into the language model, and diffusion becomes a renderer that turns the plan into pixels. The beautiful consequence is that image generation now trains like a language model and thus inherits everything language models are good at, including reasoning, tool use and scaling.
I’ve had personal success trying to use models as an ad-hoc recsys to get Spotify reccos based on what I like, or optimising my commute based on my day. The endstate probably looks something like apps and services being imbued with intent and knowledge. Not intent and knowledge as monoliths operating on apps and services. Though that's plausible too. Depends on how much centralisation pays off from a consumer UX and market dynamics POV. Those monolithic agents may well end up working, and I personally believe in them, but there's clearly more to do.
I think it’s really hard to correctly reason about all the downstream transfer that comes from more and more knowledge about the world (an improving learning process that bigger pretrains with dropping losses represent) actually buys you.
One can be ambitious: how far are we from something stranger, like an LLM that tries to figure out what you're trying to do and schedules system memory around it or something crazy like that? Maybe it recognises you importing files from your camera, and thus prefetches the project files. Huh, this is too conservative, why are we thinking about speedups and not unlocks? Maybe the computer knows what the import is, a client shoot, and so sets up the project, names the files by what's in them, and flags the blurry ones.
Software, and the way of the world in general, is full of defaults. Many of those defaults represent an average-case tradeoff i.e. a decision most could not afford to make each time. Cheap judgement changes that.
It’s pretty beautiful that most of the cases I discussed in this post were trivially operationalisable on say GPT-4 levels of intelligence, but it’s 2026 and our imagination has yet to catch up with “capabilities”. None of this needed unlocks like test-time compute, multi-agent collaboration, long-horizon capacities, etc.
Winding back to electric motors, they were good by the 1880s or so, but factories got little productivity out of them for the next 40-50 years. Economist Paul David6 found out the reason: old factories were mostly built around a single, huge steam engine driving a central shaft, with belts running to every other machine. At first, owners just swapped the steam engine for one big electric motor and then kept everything else. The real gains only started to come when someone realised you could put a small motor in every machine. That meant redesigning the building, the workflow and the jobs, and it took ~a generation to do that. Pretty reminiscent to me of chasing monolithic “center-of-your-life” agents like Muse7 (that are so superbly capable BTW), not being imaginative enough about where else to stick the thing, and then wondering something like “why is the GDP not growing”.8 I think this is a fantastic time to think about alternative "intelligence-seeding" approaches in particular or about bottlenecks to growth in general.
I think the ultimate realisation here is to just appreciate how much of a capabilities overhang we are in. What we've already got from lowering CE loss on webtext, while letting (often slow) diffusion and adoption play out their game, is sufficient to lead us into an explosion in productivity. I think the proverbial “if model progress stopped completely today, the world would still look radically different” is directionally super true. What intelligence actually buys us is limited much less by how smart models get than by how good we are at making the world amenable to them. “Amenability as the bottleneck”.