The Startup That Ripped the Chatbot Out of the Chatbot
TypeSafe AI's Jev doesn't write a single word. It just answers yes/no and picks-one questions 40 to 200 times faster than an LLM, for a fraction of the price. The catch: it can only ever be as good as the multiple-choice question you hand it.
Am I The Idiot — the strongest case against our own take.
The strongest case against this piece: I'm treating TypeSafe's core technical claim — that Jev is meaningfully "calibrated" in a way ordinary reward models aren't — as an open question, when a harsher read is that it's just marketing wrapped around a normalization trick. A softmax over options always produces numbers that sum to one; calling that "probability" doesn't make it calibrated, and I only have TypeSafe's word, filtered through outside speculation, that their Brier-score training objective actually delivers real-world calibration rather than confident-sounding numbers that are just as unreliable as any classifier's output. No independent benchmark of calibration (as opposed to speed and cost) exists in anything I found. I'm giving the company benefit of the doubt on its most interesting claim by even describing it in technical detail.
Second weak point: I lean on one third-party number — Vercel's 18x — as more trustworthy than TypeSafe's own 40-400x, but it's still a single, loosely sourced claim from AI-industry trade coverage, not a benchmark I can reproduce or a statement I confirmed with Vercel directly. It could just as easily be a curated success story a vendor was happy to have repeated.
Third: the framing of Jev as a bet "against" the industry's chatbot-centric direction is my synthesis, not something Almeida or TypeSafe explicitly claimed in those terms in any source I read. It's a reasonable reading of what he told Forbes and TechCrunch about disliking conversational AI's uselessness, but I should be honest that the framing of Jev as structurally opposed to the OpenAI/Anthropic roadmap is an editorial interpretation, not a quote. A reader could fairly say I'm building a bigger narrative than the facts, on their own, support — TypeSafe may simply be building a narrow, useful tool without any grand thesis about the rest of the industry being wrong, and Almeida's quotes about disliking chat products may be founder color rather than evidence of a coherent alternative bet.