ProgramAsWeights compiles English into a neural program you run locally
Two things landed this weekend that point the same direction: the cost of using a model is falling faster than the cost of training one, and some of the people funding frontier labs have started saying so out loud.
University of Waterloo researchers published a method that turns a plain-English description of a function into a small neural program that runs on your own machine, with no model call per use. "Compile by Training" is the second compiler behind ProgramAsWeights: you describe the behaviour once — classify urgent emails, rewrite a paragraph in a given style — a hosted service has teacher models synthesize training examples from that description, those examples fine-tune a LoRA adapter on a shared 0.6B interpreter, and the finished program is packaged so every later call executes locally, CPU included. The build is the slow part: a cold compile took 50.9 seconds on a B300, 68.2 seconds on an H200 and 99.2 seconds on an RTX card, measured at the service's May launch. On FuzzyBench-Hard — a task set where the project's earlier single-pass compiler produced no exact matches — compile by training reaches 83.6% semantic accuracy. Authors are Yuntian Deng and Pengyu Nie at Waterloo with Stuart Shieber.
The interesting claim is not the accuracy number, it is the packaging. A compiled function can be saved, distributed and composed with ordinary code, and it stops costing money every time it fires — the same separation of deciding from generating we described in The Jev idea was published a year ago — open, and unnoticed, and the same shape as Cua's 2.8 MB open model fills forms without an LLM call. The authors are upfront about the ceiling: compilation itself is hosted and sends your specification to teacher APIs, synthetic supervision inherits teacher errors, and anything requiring guaranteed correctness should keep a deterministic control path or validate outputs.
King Charles told the AI executives he convened at Dumfries House that the "existential dangers" of the technology falling into the wrong hands carry an urgency that needs addressing. "Those who have created these technologies are now increasingly warning that AI risks developing darker capacities — perhaps even to take life," he said, adding that the task is "not merely to advance technology, but to ensure that it remains firmly in the service of humanity." Nvidia's Jensen Huang, in the room, said safety was "paramount" and that a company should hold a product back and "keep engineering" if it is not safe, and argued open models stop people and countries being left behind. Demis Hassabis called AGI "probably only a few short years away" with an impact "ten times that of the Industrial Revolution," put the chance of a mistake at "definitely non-zero," and argued for "a sensible middle way." OpenAI's chief financial officer Sarah Friar said no single company or government can answer the safety questions alone.
We wrote before the summit that the only head of state convening the frontier is one with no authority to bind it, and nothing said in the room changed that — the discussion was about a "shared set of principles," not a mechanism. What is now on the record is the split inside the industry: Huang, who sells the capacity, wants products regulated and no new laws; Hassabis, who builds the models, says the catastrophe probability is non-zero and wants time to mitigate it.
GSR Ventures managing partner Zhu Xiaohu says this may be the last year foundation models are the story, and that the bubble is in individual valuations rather than the industry. In an interview published this weekend, he argued compute centres sell as soon as they are built and memory capacity gets absorbed, so there is no industry-level bubble — but "trillion-scale valuations are certainly a bubble," and humanoid robotics is a market he cannot read, with close to 200 Chinese robot companies differing very little and most orders coming from governments or research labs. On the model layer his call is blunt: models commoditise into a utility, DeepSeek-style fully open weights become the standard within roughly a decade, and within a year open models handle 99.9% of tasks — so application companies should not train their own. He is equally blunt about the API business, where a competitor at 10% of the price can take your customers "in one night," which is why he thinks the durable money sits in consumer products and office agents — and concedes that on the office side the platforms own the entry point.
Read against the financing debate we covered in The $1 trillion AI build-out hits a wall cash can't fix, the notable part is which half of the stack he thinks is safe: not the frontier models, not the robots — the compute landlords and the application layer.
What to watch: whether compile-by-training's accuracy holds outside the authors' own benchmark, and whether any lab publishes what it actually agreed to in Ayrshire.
Would you run a compiled model on your own machine if it meant being 10% behind the frontier — or is that good enough for the work you actually do? Tell us in the comments.
