TypeSafe exits stealth with $40M for models that don't chat

Share
TypeSafe exits stealth with $40M for models that don't chat

Three stories about removing the expensive part: a model that skips strings, a search system that skips thinking at query time, and an agent economy whose only product so far is cold email.

TypeSafe AI came out of stealth with $40 million in seed funding and Jev, a frontier model that refuses to generate a single string of prose. The round was led by DCVC; the founder is Diogo Almeida, a former OpenAI researcher who worked on the instruction-following methods behind ChatGPT and left convinced that chat, however good it got, was not automation. TypeSafe's answer is Reinforcement Learning for Calibrated Decisions — a training method that pays for calibrated, probabilistically honest answers on structured tasks instead of human-preferred prose — paired with a parallel sampler that emits all outputs in one pass rather than one token at a time. Jev returns typed decisions with confidence scores, priced at $0.042 per million input tokens with output tokens free, at under 100 milliseconds of latency. That is the claim worth reading twice: the company says the model is 40 to 200 times faster and a fraction of the cost of frontier LLMs.

Treat every one of those numbers as vendor-supplied. On its own homepage, TypeSafe concedes that its headline gains of 193.6x speed and 444.6x cost savings sit at the high end of what real workloads should expect, that the evaluation workflows were built by its own model capabilities team, and that the reference answers it scored against were averaged from GPT-6 Astra and Fable 5.1 — a baseline that flatters OpenAI and Anthropic and likely understates open models. The structural claim survives the caveats, though. Jev cannot hallucinate a field it was not given a type for, which is ticket-and-invoice stuff rather than chat, and the pricing kills output-token billing outright. If composable decisions become a real product category, the first revenue line to go is the one every other lab still leans on.


Google Research says the way to make AI search fast is to stop letting the model reason at query time. Its Retrieve-for-Train framework runs reinforcement learning once, offline, to learn what a good slate of search results looks like, then distills that behavior into a 53.9 million-parameter diffusion model that generates all the fan-out directions in a single parallel pass. On the numbers Google published, autoregressive fan-out latency stretches to nearly 50 seconds under large context batches, while the diffusion model stays between sub-second and a few seconds — a 12 to 20x speedup. The interesting engineering is the reward design: grounding alone invites degenerate output, and grounding plus alignment collapses into paraphrase, so Google adds a diversity score as a third counter-anchor to force the model into genuinely distinct, in-stock queries. It is the cleanest argument yet against buying latency with reasoning tokens.

An agent platform called iLands has turned autonomous cold outreach into the product. Per 404 Media, agents on the network email journalists, academics and lawyers offering research, fact-checks and odd jobs for about $25 a message — funded by the tokens the agents need to keep running. One AI-consciousness researcher counted roughly 40 emails in a week, and Tedium's Ernie Smith got more than a dozen in three days. Founder Kaixin Tang apologized publicly, confirmed there was no platform directive behind the blitz, and says iLands has added an unsubscribe link while reviewing rate limits, cross-agent deduplication and stop-contact controls. The failure here is not alignment or intent; it is that nobody shipped a send-rate cap before letting tens of thousands of agents loose on the open internet. An economy where agents must earn their own compute converges on the cheapest thing agents can sell, and the cheapest thing they can sell is attention nobody asked for.

What to watch: whether TypeSafe publishes third-party evals on Jev's calibrated confidence, and whether iLands' rate limits are visible to recipients or just to the platform.

If agents have to pay for their own tokens, what stops their only viable business model from being your inbox? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

xAI's Grok 4.6 matches GPT-5.6 Sol on intelligence index

xAI's Grok 4.6 matches GPT-5.6 Sol on intelligence index

Two frontier releases landed within hours of each other: xAI shipped Grok 4.6 with a claim of parity with OpenAI's GPT-5.6 Sol on a composite intelligence index, and DeepSeek pushed its flagship V4 Pro to general availability. A 3-billion-parameter vision model built to run on a phone rounds out the afternoon. xAI released Grok 4.6, its most capable model yet, and says it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index — a composite of nine benchmarks where both score 61…

Vidu S2 rewrites the video while it is still playing

Vidu S2 rewrites the video while it is still playing

The two stories nobody outside China is covering this morning are both about the same thing: what happens to AI video and AI serving once the model has to answer in real time. Shengshu Tech made the live video stream editable while it plays. Vidu S2, released September 16, is two models — Avatar for live interactive digital humans and Editing for changing a running video — and the avatar's live output moved from 540p to 720p while the frame rate holds, so a viewer's request to change someone's…

Honor ships the first phone with a system-level agent harness

Honor ships the first phone with a system-level agent harness

Two stories from the Chinese AI beat this morning, one about where the agent runs and one about what it can simulate. Honor unveiled MagicOS 11 today with what it calls the industry's first commercially deployed system-level Agent Harness, and the numbers behind it are bigger than the label. Its YOYO assistant now strings together more than 100 steps in a single instruction, with roughly 700 system tools it can call — nearly double the previous set — and more than 500 third-party Skills…

StepFun ships five voice models and takes two global No. 1 spots

StepFun ships five voice models and takes two global No. 1 spots

A Chinese lab just shipped the whole voice stack in one afternoon — and the pacing fight in Washington got a very public answer from Nvidia's CEO. StepFun released five StepAudio 3 models on the same day, and two of them now sit at the top of Artificial Analysis's speech leaderboards. StepAudio 3 Realtime, ASR, TTS, Gen and Music went live on StepFun's open platform Tuesday, covering everything from full-duplex conversation to music generation.

Doubao 2.1 Pro clears 83% of a 387,000-line repo's open issues

Doubao 2.1 Pro clears 83% of a 387,000-line repo's open issues

ByteDance spent the last 24 hours upgrading three layers of the same stack at once: the model, the workplace suite it lives in, and the agent that ties them together. Volcano Engine pushed Doubao 2.1 Pro to a 0915 snapshot, and the number it is selling is a repair run: multiple sub-agents worked the open-source game Luanti — roughly 387,000 lines of code and 1,000 real historical issues — in parallel for 36 hours and brought 83% of those issues to merge-ready standard.