TypeSafe AI, which is developing a model that outputs numerical responses with probability estimates to help businesses assess reliability, raised a $40M seed
Former OpenAI researcher Diogo Almeida says most intelligence is going to live inside software. But to automate any real work, it has to be correct.
ForbesRashi Shrivastava
Context & Ripple Effects
TypeSafe AI’s funding accompanies the debut of Jev, its typed probabilistic-decision model, positioning probability estimates as an interface for software and other AI models rather than as conversational output alone. Diogo Almeida’s OpenAI background gives the company a recognizable research pedigree as it enters that market.
The launch lands alongside AI Underwriting Company’s audit-and-certification funding, highlighting two different approaches to operational AI assurance: embedding confidence estimates in model outputs and evaluating agents through an external assurance layer.
First-order effects
The $40M seed gives TypeSafe AI resources to develop and distribute Jev to software builders that need numerical decisions accompanied by probability estimates.
Developers and AI-model operators gain a model interface designed for typed decisions, allowing them to incorporate reported confidence into automated workflows rather than relying solely on free-form responses.
Second-order effects
AI Underwriting Company will need to distinguish the value of external agent audit and certification from model-native probability estimates as buyers evaluate how to assess reliability.
Enterprise AI buyers will face a more explicit choice between confidence signals supplied by a model such as TypeSafe’s and independent assurance processes, making the provenance of a reliability claim more important in procurement.
Third-order effects
If reported probabilities prove useful in production, operational AI assurance may divide into a model layer that quantifies uncertainty and an external layer that tests or certifies whether those signals can be trusted.
That division would favor AI systems designed as composable software components, where automated workflows can route decisions based on typed outputs and confidence thresholds.
The trend: AI deployment is shifting from general-purpose text generation toward measurable, software-native decision components with explicit reliability signals.
We believe that the future is code + AI, so made workflow evals to reflect that Jev costs: $42 / BILLION input tokens ($0.042 / MTok) and output tokens are free (forever - they're too cheap to meter with our new architecture) Jev is named after Jevons paradox and off the intellig…
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I've spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x [vid…
don't get one-shotted by this it's not a general language model and can't generate free form text it's probably a specialized diffusion model and it can only output a few different primitives and requires definitions of the output format
diogo is an immensely creative guy and is working on really different types of models with the principle of building composable, programmable AI systems from layers of small inferences. it's a weird and ambitious idea thats worth tinkering with
we almost never test new foundation models but we've been testing this for ~a week @every and it's pretty wild. the kind of things that will be obviously indispensible in 6-12 months it doesn't produce words as output, it produces probabilities. so it can efficiently act as a jud…
Another brilliant launch for developers: and its 20-200x faster than LLMs because it skips token-by-token generation entirely. TypeSafe AI just launched Jev, > 20-200x faster >40-400x cheaper (w/ output tokens free) > Frontier composable intelligence optimized for decisions So Je…
Game: race from one Wikipedia page to another using only links Challenge: choosing between hundreds to thousands of links Shows not just intelligence-per-second, but also the compounding benefits of not hallucinating with high-cardinality choices [video]
I've genuinely never seen a model claim this low of a hallucination rate. This company was founded by Diogo Almeida, one of the researchers behind the instruction following work that led to ChatGPT, and instead of building an autoregressive chatbot that generates strings token by…
i don't think people realize what just happened... 20-200x faster. 40-400x cheaper. output tokens free forever. the man who co-invented ChatGPT just dropped a new kind of AI called Jev and the twist is it can't write a word, it only makes decisions. he's calling it the shortest p…
A few weeks ago someone sent me a screenshot of the InstructGPT paper (that led to ChatGPT) with a name highlighted, and asked if I wanted to test a new language model Diogo was working on that doesn't output text... Couldn't resist, so here's my vibe check: https://every.to/...
This is a big deal. Right now, people use LLMs a lot as classifiers in production systems. But they're slow, expensive, and hallucinate. This model is supposed to be >100x faster, at $42/BILLION tokens, and generates not only type-safe structured output but also calibrated confid…
This is incredibly cool. Completely new form of AI models—output tokens are so cheap to meter, they're literally free. The deflation of intelligence continues. 👇
Extraordinary claims require extraordinary evidence so check out our release blog for more technical info: https://typesafe.ai/... Join our waitlist for early access: https://typesafe.ai/ Have technical chats and meme with us on Discord (rumors are good memers skip the line): htt…
The gains aren't free: Jev can't generate text Comparing Jev vs LLMs side-by-side makes the trade-off clear Fun fact: replacing sequential computation with parallel is the same way Transformers leapfrogged RNNs [video]
Anthropic and OpenAI are already free to “pace the frontier” and should do so for business reasons, instead of first demanding a preferred regulatory framework
This is extremely intriguing and might portend a future where LLMs perform many of the tasks they're currently good at vastly more efficiently and cheaply
These claims are downright bonkers. By *not* generating text, models become dramatically faster, cheaper, and more precise. — It's not an understatement to say that if this turns out to be true it would change so much about building with AI, and will lead to a dramatic spike i…
as far as i can tell, it's purely an RL feat — seems like they — 1. take a fully pretrained LLM (decoder-only??) — 2. slap a new 255-element classifier head on it — 3. RL for Calibrated Decisions (RLDR) — i'm sure there's a strict format for the prompt, and during RL it…
Jev: Fable-level model that doesn't charge for output tokens because they're too cheap to meter — it's not general though, it only makes decisions, doesn't generate text, but input tokens are measured by the billion ($42/btok) — typesafe.ai/blog/introdu... [image]
This shit is fucking crazy - I have set Astra with the link on a journey to train our own. If all goes well Erislab might shit out a vibed version of whatever the fuck they're doing, if they've make the mistakes of leaking enough bits for me to figure out what it is they are doi…