Why an un-record-breaking model is exciting
I read Thinking Machines' technical release notes for Inkling end to end, and what stopped me wasn't the parameter count — though a mixture-of-experts design with 975 billion total parameters that activates only ~41 billion per task is elegant in itself. What stopped me was a sentence the company put in its own briefing materials: Inkling is 'not the strongest model available today, closed or open.' Deliberate modesty at a model launch! They brag about other things instead: calibrated answers that flag uncertainty rather than guess, a 'thinking effort' dial you can turn down when you want speed, and hitting Nvidia's Nemotron 3 Ultra's coding performance on one benchmark while spending a third as many tokens.[1]
The real bet isn't the model but the idea around it: the weights are open, and revenue comes from Tinker, the fine-tuning platform. Their proof point, the Bridgewater experiment, is striking — an open model trained further on the world's largest hedge fund's own financial expertise scored 84.7% on financial reasoning tests, beating top proprietary models at roughly one-fourteenth the running cost. (Honest footnote: those results are the two companies' own evaluation, not an independent one.) Training on 45 trillion multimodal tokens, market entry in nine months... These are signs the one-size-fits-all era is ending.[1]
On the code side, the quiet revolution is complete
A figure from Coinbase the same day photographs this shift in production: head of platform Rob Witoff says 95 to 100% of the company's code is now written by or with large language models. In February that share was 40%. From 40% to nearly 100% in five months — that's not an experiment, it's the default way of working. The critical detail is in place too: in cryptography and core security, human oversight stays central, while AI accelerates prototyping and routine work in lower-risk areas.[2]
Put the two stories side by side and the shape of late 2026 comes into focus: the model layer is commoditizing, and value is moving up — to adaptation, oversight and orchestration. Inkling's offer of 'we're not the strongest, but with your data we could be the strongest for you' and Coinbase engineers becoming conductors who dispatch work to agents are two pages of the same story. The programmer's job isn't disappearing; the recipe is changing. And I'll confess: the new recipe looks more creative to me than the old one.[1], [2]