Modern Creator
Why Modern Creator?
AI Revolution · YouTube

Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions

TypeSafe AI's Jev never generates a sentence. It returns a calibrated probability instead, and developers adopted it faster than any model in Vercel's AI Gateway history.

Posted
yesterday
Duration
Format
Essay
educational
Views
50.6K
876 likes
Part of the collectionJev, explainedEvery Jev breakdown, synthesized into one page.
Read the playbook
Big Idea

The argument in one line.

Jev is a non-LLM "System 1" model that replaces generated text with typed, calibrated probabilities for machine-to-machine decisions, and its early adoption suggests structured decision-making, not chat, may be the more automatable use of AI.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You build AI agents or workflows and want a faster, cheaper way to make branching decisions than parsing free-text LLM output.
  • You're evaluating whether a classification or routing step in your pipeline actually needs a full language model.
  • You follow AI model launches and want the real claims behind a hype-packaged headline, not just the pull quotes.
SKIP IF…
  • You're looking for a chatbot, coding assistant, or anything that needs to generate natural language, which Jev is explicitly not built to do.
  • You want peer-reviewed benchmarks, since TypeSafe skips public leaderboards in favor of its own workflow evals.
TL;DR

The full version, fast.

Diogo Almeida, who co-invented RLHF and helped build ChatGPT, left OpenAI because years of superhuman chat performance never translated into real automation. His company, TypeSafe AI, built Jev: a transformer that never outputs text, instead returning typed, calibrated probabilities like billing 0.08 or technical 0.85 that code can branch on directly. TypeSafe claims 40 to 200 times the speed of frontier LLMs, roughly 238 times lower cost, and zero hallucination by construction, since output is locked to a predefined schema. Within 24 hours it became the fastest-adopted model in Vercel's AI Gateway history. The catch: it can't chat, skips public benchmarks for its own workflow evals, and independent reviewers note a probability can still be wrong even if it can't be a hallucinated sentence.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →
Chapters

Where the time goes.

00:00 – 00:57

01 · Cold open: the RLHF inventor's disappointment

Diogo Almeida, who co-invented RLHF and helped build ChatGPT, says four years of superhuman chat performance never produced real automation, because computers speak a different language than people.

00:57 – 01:51

02 · Introducing Jev: a System 1 model, not an LLM

TypeSafe AI, Almeida's company, raised $40M and spent two years in stealth before launching Jev: a transformer-based model that gives back typed probabilities instead of text.

01:51 – 02:46

03 · How it works: choice, score, null, and type safety

Jev takes unstructured state in and returns typed decisions via three primitives, choice, score, and null, each evaluated in parallel with its own confidence score.

02:46 – 03:42

04 · The claims: no hallucination, a new architecture

By locking output to a predefined structure, TypeSafe says Jev can't return anything unexpected. Instead of generating tokens one at a time, it produces its full answer in a single hardware-aware pass.

03:42 – 04:42

05 · Speed and price benchmarks

TypeSafe claims 70 to 500 millisecond response times versus 3 to 329 seconds for frontier LLMs, and pricing near $0.042 per million input tokens with free output.

04:42 – 05:34

06 · Calibrated confidence, and TypeSafe's own caveats

LLMs are overconfident and inconsistent about their own certainty; Jev's calibrated confidence means higher stated confidence tracks higher accuracy. TypeSafe pairs the claim with its own list of caveats.

05:34 – 06:50

07 · Workflow evals versus the simplified demo

TypeSafe admits its side-by-side demo against GPT-5.6 Terra was simplified and flattering, then points to workflow evals, run without hand-tuning, as the heavier evidence.

06:50 – 07:58

08 · Homepage numbers, and the hallucination pushback

TypeSafe's homepage claims 194 times faster and 445 times cheaper. Its 0% hallucination figure isn't measured, it's guaranteed by schema matching, and The Register pushes back that comparing a language model to a non-language model isn't fully fair.

07:58 – 09:20

09 · The Doom bot and Wikipedia race demos

TypeSafe built a bot that plays Doom from text game-state and ran a Wikipedia link-race, showing Jev finishing in fewer steps than LLM competitors at a fraction of the cost.

09:20 – 10:39

10 · Developer adoption: Vercel, Bryo AI, and the Gateway numbers

Demand got so high TypeSafe briefly couldn't serve API traffic. A Vercel engineer got a safety classifier running 5 to 18 times faster, and within 24 hours Jev became the fastest-adopted model in AI Gateway history.

10:39 – 12:22

11 · Where it fits, and what's inside Jev

Vercel and TypeSafe pitch Jev for agent routing, workflow control flow, risk scoring, map-reducing large datasets, and real-time verification. Jev is trained exclusively on synthetic data TypeSafe generates itself.

12:22 – 14:41

12 · The philosophy: bitterest lesson, Jevons paradox, and what's next

TypeSafe argues optimizing for the right task beats more data or compute, names the model after the Jevons paradox, skips public benchmarks by choice, and closes with Almeida's take on what frontier labs should actually be selling.

Atomic Insights

Lines worth screenshotting.

  • TypeSafe AI's Jev is a transformer-based model that never outputs text, returning calibrated probabilities instead, like billing 0.08 or technical 0.85.
  • Jev answers in 70 to 500 milliseconds versus 3 to 329 seconds for frontier LLMs, a claimed 40 to 200 times speedup.
  • Jev is priced at $0.042 per million input tokens with free output, versus roughly $2 in and $12 out per million for GPT-5.6 Terra.
  • TypeSafe says Jev can't hallucinate because its output is locked to a predefined schema, so there's no free-form text left to go wrong.
  • The RLHF co-inventor who helped build ChatGPT left OpenAI because four years of superhuman chat performance never turned into real automation.
  • TypeSafe trains Jev entirely on synthetic data it generates itself, calling that decision one of the best bets in the company's history.
  • Within 24 hours of launch, Jev reached more than twice as many paid teams on Vercel's AI Gateway as any previous model release.
  • A Vercel engineer who swapped a command-safety classifier from an LLM to Jev got results 5 to 18 times faster.
  • TypeSafe skips public benchmarks entirely, arguing System 1 tasks are easier to evaluate against your own workflow than against a shared leaderboard.
  • Jev is named after economist William Stanley Jevons, whose paradox holds that making a resource cheaper increases total demand for it rather than shrinking it.
  • Jev's decision space tops out at 255 options; past that it scores choices independently before picking, which slows it down.
  • TypeSafe's own side-by-side demo used a deliberately simplified query, and outside reviewers noted that comparing Jev to an LLM is inherently unfair since Jev never produces language at all.
Takeaway

Structured decisions may be the more automatable half of AI

WHAT TO LEARN

The bottleneck in AI automation isn't model intelligence, it's that most models answer in free text when code actually needs a typed, confident decision.

01Cold open: the RLHF inventor's disappointment
  • A tool can be technically impressive and still fail at its intended use case, which is worth checking before assuming better output alone will drive adoption.
  • Chat-optimized performance and automation-ready performance are different goals; getting very good at one doesn't guarantee progress on the other.
02Introducing Jev: a System 1 model, not an LLM
  • When an existing approach plateaus, the fix is sometimes a different output format entirely, not a bigger or better version of the same format.
  • A model can be transformer-based without being a language model; the underlying architecture doesn't determine what kind of output a system produces.
03How it works: choice, score, null, and type safety
  • Breaking a decision into typed primitives, a choice among options, a numeric score, a yes/no, makes it directly usable by code, where a sentence would first need to be parsed.
  • Evaluating several questions in parallel, each returning its own confidence score, gives more actionable signal than one confident-sounding paragraph.
04The claims: no hallucination, a new architecture
  • Locking output to a schema decided in advance is what actually prevents unexpected results, not asking a model to try harder or be more careful.
  • Generating a whole structured answer in one pass, instead of building it piece by piece, removes a source of compounding error at every step.
05Speed and price benchmarks
  • Latency differences of two or three orders of magnitude between approaches usually point to a fundamentally different computation, not just a faster server.
  • Pricing a service per billion units instead of per million is itself a signal of how cheap the underlying computation is meant to be.
06Calibrated confidence, and TypeSafe's own caveats
  • A model that's right 95% of the time is only safe to automate if it can also tell you which cases fall in the other 5%.
  • Publishing your own limitations before critics find them is a credibility move worth copying when making a strong technical claim.
07Workflow evals versus the simplified demo
  • A flattering side-by-side demo and a real workflow evaluation are not the same evidence; the second is harder to produce and worth more.
  • Disclosing that a comparison case was genuinely ambiguous, rather than a clean win, is a stronger signal of honesty than letting every result look favorable.
08Homepage numbers, and the hallucination pushback
  • A 0% hallucination rate that's guaranteed by design is a different, stronger claim than a 0% rate that was merely measured on a test set.
  • When your own team builds and interprets an evaluation, disclose the bias risk explicitly rather than let a favorable number stand unquestioned.
09The Doom bot and Wikipedia race demos
  • A playful demo, like a bot playing a game or racing across links, can make an abstract efficiency claim tangible in a way a benchmark table can't.
  • Admitting a simpler, non-AI approach would outperform your demo doesn't undercut the point if the demo's actual goal was adaptability, not winning.
10Developer adoption: Vercel, Bryo AI, and the Gateway numbers
  • The most convincing proof of a tool's value is often an independent user's own before-and-after number, not the vendor's benchmark.
  • A model that's slightly less accurate but returns a real confidence score can still be preferred, if that score is what makes a workflow safe to automate.
11Where it fits, and what's inside Jev
  • Naming your target use cases explicitly, like agent routing, retry logic, or risk scoring, helps people see uses you never demoed directly.
  • Choosing to train exclusively on data you generate yourself is a defensible strategy when your target output doesn't resemble naturally occurring human text.
12The philosophy: bitterest lesson, Jevons paradox, and what's next
  • Optimizing for the right task can matter more than adding more data, compute, or clever algorithms to the wrong one.
  • Skipping shared public benchmarks for your own workflow evals is a real tradeoff: less external comparability, but a truer measure of your actual use case.
  • Betting that cheaper capability increases total usage, rather than just cutting your own costs, is a specific economic claim worth stating explicitly, not assuming.
Glossary

Terms worth knowing.

System 1 model
TypeSafe's term for a fast, non-language model that returns typed probabilistic decisions, borrowed from Kahneman's idea of fast, automatic thinking as opposed to slow, deliberate reasoning.
RLHF (Reinforcement Learning from Human Feedback)
A training technique that tunes a model to produce output human raters prefer. It's the dominant method behind chat-tuned LLMs like ChatGPT.
RLVR (Reinforcement Learning with Verifiable Rewards)
A training method that only works for tasks you can check programmatically, such as math proofs or code correctness, so it doesn't generalize to fuzzier real-world judgment.
RLCD (Reinforcement Learning for Calibrated Decisions)
TypeSafe's own training method for Jev, designed so the model's stated confidence actually matches how often it's right.
Calibrated confidence
A property where a model's stated probability, like 95%, actually matches its real accuracy at that confidence level, so you know when a decision is safe to automate.
Type safety
A programming concept where a system rejects data that doesn't match an expected structure, preventing errors like trying to divide a number by a text string.
Jevons paradox
The 19th-century economic observation that making a resource cheaper to use tends to increase total consumption of it rather than reduce it.
AI Gateway
Vercel's routing layer that lets developers call different AI models through one interface, used here to measure how fast teams adopted Jev after launch.
Resources

Things they pointed at.

00:00linkTypeSafe AI: Introducing System One Models and Jev ↗
00:00linkTechCrunch: A new kind of AI model from a ChatGPT inventor is thrilling developers ↗
00:00linkThe Register: TypeSafe AI debuts model for machines that plays Doom ↗
00:00linkVercel: AI Gateway Jev model launch ↗
Quotables

Lines you could clip.

00:08
“We have lightning in a bottle, and yet it is not useful.”
the whole video's thesis in one line from the person who helped build ChatGPT→ TikTok hook↗ Tweet quote
01:48
“Useless for a person, perfect for code, because you just branch on it.”
crisp explanation of what a calibrated probability output is for→ IG reel cold open↗ Tweet quote
05:08
“Extraordinary claims require extraordinary evidence.”
TypeSafe quoting the standard skeptics use against them, disarming pushback→ newsletter pull-quote↗ Tweet quote
13:25
“More like the early internet than today's mega apps.”
the vision line for where this goes if the bet is right→ newsletter pull-quote↗ Tweet quote
14:26
“The main product of frontier labs is fear or hype, and he'd like his to be intelligence.”
closing line, direct shot at the AI hype cycle→ TikTok hook↗ Tweet quote
The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogystory
So one of the people behind ChatGPT says ChatGPT broke his heart. Diogo Almeida was an open AI researcher who helped build the chatbot and co -invented RLHF, reinforcement learning from human feedback, which is probably the training technique most responsible for this whole AI era. And yet he was disappointed.
In his words, we have lightning in a bottle, and yet it is not useful. His argument is that the industry has been optimizing for human language. Models have been super good at it for about four years, but that doesn't get you automation because computers speak a different language.
In the launch post, he asks, models have been superhuman at chat for years, so where is all the automation? He once thought chat models might lead to AGI, but despite the hype, something really big was clearly missing. So two years ago, he left OpenAI and started TypeSafe AI, where he's co -founder and CEO.
They raised $40 million, spent two years in stealth, and on Tuesday, September 15th, released Jev, their first System 1 model, a new class of frontier models for fast, structured decisions software can use directly. And the weird part is, it's Transformer -based, but it's not an LLM. It doesn't output text.
It gives back probabilities, or what they call calibrated decisions. Almeda calls this machine -native AI because if AI is going to change how work gets done, people can't be the only consumers of intelligence. TypeSafe calls it a frontier intelligence function call, unstructured state in, typed probabilistic decisions out.
The state can be JSON or just a string like, my card was charged twice. Then you ask questions using primitives called choice, score, and null, which return typed choices, scores, or true or false answers with probabilities, all evaluated in parallel.
The registers example. You ask which of three departments should handle a customer service query, and you get billing 0 .08, technical 0 .85, sales 0 .07, confidence 0 .82. Useless for a person, perfect for code, because you just branch on it.
That's where the name comes from. Type safety is how programs catch data they didn't expect, like trying to divide an integer by a text string. With Jev, you define the outputs in advance, so it can't return anything outside that structure.
No parsing, no validating. An LLM gives you a string, which could be a chat response, code, a hallucination, a refusal, or valid structured data with some risk it goes off the rails. You can wrap an LLM in code, sure, but Jev's appeal is efficiency.
TypeSafe claims it can't hallucinate, and that it matches LLMs on System 1 tasks while being two orders of magnitude faster and more efficient. Under the hood, there's a new architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions, or RLCD. The register calls RLCD an architecture, but it's really the training method, basically their answer to RLHF and RLVR.
GPT 5 .6 Tera generates one token at a time, each conditioned on the last. Jev produces everything at once in a single, hardware -aware query, built around structured program state instead of sequential messages. Giving up strings, they say, gives them superpowers.
And the speed is kind of wild. Typesafe says Frontier LLMs take 3 to 329 seconds end -to -end. Fine for humans, but a bottleneck inside code.
Jev takes 70 to 500 milliseconds, so 40 to 200 times faster. In their side -by -side demo, which TechCrunch also ran as a screenshot, Jev answered in 0 .114 seconds. Terra took 8 .566.
Price is even crazier. LLMs charge $0 .20 to $10 per million input tokens, with output about five times more. Jev is $0 .042 per million, so $42 per billion.
And yeah, they meter by the billion. Output is free, too cheap to meter. Tera is $2 in and $12 out per million, and the register says Jev is 238 times cheaper than Fable 5 .1.
Then there's confidence, which honestly might be the most underrated part. LLMs are overconfident and inconsistent about how sure they are. And if a model does a task 95 % of the time but can't tell you when it's in the 5%, you can't automate that task.
Jev's confidence is calibrated, so higher confidence actually means higher accuracy, and similar inputs get similar answers. Armin Ronacher, CTO of Earendil, which makes the open -source model harness Pi, told TechCrunch it kind of delegates the hallucination problem to the user. 50 %?
Coin toss? Ignore it. 95 %?
Sure, act on it. Okay, big claims, and TypeSafe knows it. The post says, extraordinary claims require extraordinary evidence.
Then see below for the receipts with a nail polish emoji. They say they love skeptics and are skeptics themselves. And to their credit, they add a lot of nuance.
Speed, you can check yourself, though their evals mostly run from laptops on the West Coast, where the service is based. Pricing is transparent, but they admit they can't prove it isn't subsidized and expect it to drop. and no type errors, which one counterexample would disprove.
But that's mathematically impossible, they say. On the Terra demo, they admit the query was highly simplified, with readable question keys and a short, dense state paragraph that flatters Jev. They picked Terra with default reasoning as the closest to Jev in intelligence, and the only disagreement was on churn likelihood level, which they call genuinely ambiguous.
Funnily enough, a similar demo is what convinced them to go all in. The heavier evidence is their workflow evals, which measure how AI performs inside code. No ground truth labels, and no tweaking the harness, which invites overfitting.
Every model gets the same workflow, scored against the average of GPT -6 Astra and Fable 5 .1.
owns the Pareto frontier across almost two orders of magnitude, and LLMs doing all the logic and chain of thought did significantly worse than following the workflow. These calls are way more complex than the demo, closer to production, and four are published on their eval site. Reliable workflows, they argue, are lots of small, independent, probability -driven questions that end in a branch.
That's where the homepage numbers come from. 193 .6 times faster and 444 .6 times cheaper, which Vercel rounded to 194 and 445. Probably the high end of real world gains, they say.
The workflows weren't cherry -picked and aren't in the training distribution, but their own team built them, so some bias is possible. The Astra and Fable reference probably tilts toward OpenAI and Anthropic, so they figure they're underestimating themselves and DeepSeq. And the LLMs ran through TypeSafe's open -source System 1 LLM wrapper for Python.
The most accurate way to get LLM decisions, they say, but slower and pricier. On Hallucination, they call TypeSafety table stakes. A hallucinated tool call is inconvenient in an agent, but a deal -breaker in a system with latency guarantees, or buried several layers deep in a dependency chain, and even the smartest models still do it.
Their LLM numbers come from OpenRouter, where complex queries might get routed to better models, and JEV's 0 % isn't measured, it's just guaranteed by schema matching. The register pushed back, Fairly, I think.
Jev doesn't produce language, so the comparison isn't fair, and a probability can still be wrong. It won't invent fake legal citations, sure, but it won't be used that way either. Now the fun stuff.
They built a Jevbot that plays Doom, a Dumo that doominstrates real -time intelligence, their words. Its engineer worried about 10 queries a second, around $7 an hour, which the team thought was lower than expected. It reads GameState as text, not images, not yet, and they admit a non -AI bot would play better, but they wanted one that adapts to different GameState representations and follows instructions, which was, quote, cool as heck.
A walkthrough and hack events are coming. Then there's wickeracing, getting from one Wikipedia page to another using only links, with hundreds or thousands of options per step. it shows off intelligence per second, and how not hallucinating compounds over huge choice sets.
Jev supports a cardinality up to 255, so beyond that it scores links independently, then picks, hence the occasional slowdown. Speedups are smaller since the LLMs ran without reasoning, except Astra on its lowest setting to keep it watchable. And Jev tended to finish in fewer steps.
My favorite detail. The second and third challenges both randomly started on Rubber Duck. and the author didn't notice until the team pointed it out.
And developers jumped on it. Demand got so high that TypeSafe briefly couldn't serve users from its API, and TechCrunch says devs see it as a cheaper, more robust way to put intelligence into code. Pranit Sharma, an engineer at Vercel, said they'd been running a command safety classifier on OpenAI's chat GPT Luna 5 .6, and swapping in Jev got results 5 to 18 times faster.
and more accurate. Nikhil Mudholkar, CTO of Bryo AI, tested it against Gemini on business emails. Gemini was slightly more accurate, but 10 to 20 times pricier.
What got him, though, was Jev being the only one handing back a real probability, which he says makes it ideal for automating workflows. Vercel's AI gateway numbers are kind of absurd. Within 24 hours, GEV reached more than twice as many paid teams as any previous launch, the fastest adopted model in Gateway history.
It passed every comparison model within 12 hours, hit a tenth of paid teams by hour 18, and nearly 13 % by hour 24. That's two times the GPT 5 .6 family and over six times Fable 5 .1, while no other recent launch got past 7 % after a full day. Vercel says it shows how fast a specialized model can land in production, but the real test is whether that lasts.
As for uses, Vercel mentions agents picking their next tool or sub -agent, workflows deciding whether to continue, retry, ask the user, or stop, scoring risk before acting, and kicking uncertain cases to a human. TypeSafe calls these smart -if statements, fuzzy rules where handwritten logic is too brittle. They also pitch map -reducing petabytes into features, real -time apps where 100 milliseconds works for UX, and verifying everything, from reasoning traces to jailbreaks.
Almeida expects people to monitor agent traces with it, since agents watching agents gets expensive, Roniker suggests real -time model routing, and the register adds model harnesses. LLM still own chatbots, co -pilots, coding agents, verifiable problems like math proofs, and demos that, as TypeSafe cheekily puts it, only work sometimes.
So what's actually inside? Almeida is tight -lipped, and outside observers suspect an open -weight LLM underneath, though the frequently asked question says Jev is neither small nor an LLM. Jev is trained exclusively on synthetic data, and Almeida told TechCrunch making all their own data was among the best bets of his life.
Better than their launch, better than RLHF. TypeSafe calls itself primarily a data research lab, and half the company basically owns the subfield of statistically well -understood synthetic data, which is now his life joy. They won't train on your data even if you ask, no offense, and for details, they'd have to hire you.
The philosophy is what they call the bitterest lesson. Optimizing for the right task matters more than data, compute, or algorithms. RLHF optimizes for text a human rater prefers, great for chat, wrong for automation, and RLVR only fits tasks you can check programmatically so real -world judgment gets spiky.
Non -robust intelligence. The System 1 name comes from Kahneman's thinking, fast, and slow. And even though System 1 usually implies error -prone, they think these models can be more reliable than the alternatives.
They're also skipping public benchmarks entirely, just one -off evals with updates, pushing users to build their own, since System 1 tasks are easier to evaluate. There's even a post called anti -bench -maxing. JEV itself is named after William Stanley Jevons, the 19th century economist whose paradox says that when steam engines got more efficient, coal demand went up, not down.
The bet, theirs and the industry's, is that every order of magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases. Almeida sees smart software everywhere, emergent and distributed, more like the early internet than today's mega apps. The register notes that Assume's token demand is as broad as Energi's, which isn't settled.
Pew found many people with AI access just don't have a use for it, and a CNBC survey found others avoid AI over moral objections. For now, Jev stands alone, but Renacher expects competitors. He said, TypeSafe is building more versions in new modalities, pulling developers off the waitlist, and asking what sci -fi people want to build.
And when TechCrunch asked whether TypeSafe is a frontier lab, Almeida said the main product of frontier labs is fear or hype, and he'd like his to be intelligence, not a bet on infinite wealth or a religion, or building God in a data center. Anyway, that's Jev. Let me know in the comments what you'd hook it up to.
Thanks for watching, and I'll see you in the next one.
The Hook

The bait, then the rug-pull.

One of the people who helped build ChatGPT says it broke his heart. Diogo Almeida co-invented RLHF, the technique behind most chat-tuned models, then left OpenAI because years of superhuman chat performance never turned into real automation. His answer is Jev, a model that refuses to speak at all.

Frameworks

Named ideas worth stealing.

01:44list

Choice / Score / Null primitives

  1. choice
  2. score
  3. null

The three question types Jev answers: choice returns typed options with probabilities, score returns a numeric estimate, and null-style questions return true/false with a probability.

Steal forany agent or workflow branch that currently parses free-text LLM output to decide what to do next
12:23concept

The bitterest lesson

TypeSafe's framing that optimizing for the right task beats more data, compute, or algorithms: RLHF optimizes for human-preferred text, RLVR only fits programmatically checkable tasks, and neither fits general real-world judgment well.

Steal fordeciding whether a task actually needs a language model or just a well-scoped classifier
13:05concept

Jevons paradox

When a resource gets more efficient to use, total consumption of it tends to rise rather than fall, historically true of coal and steam engines, and TypeSafe's bet for AI intelligence.

Steal forarguing that cheaper AI inference expands the market rather than shrinking AI spend
CTA Breakdown

How they asked for the click.

MENTIONED ON CAMERA
00:00linkTypeSafe AI: Introducing System One Models and Jev ↗
00:00linkTechCrunch: A new kind of AI model from a ChatGPT inventor is thrilling developers ↗
00:00linkThe Register: TypeSafe AI debuts model for machines that plays Doom ↗
00:00linkVercel: AI Gateway Jev model launch ↗
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Storyboard

Visual structure at a glance.

cold open
hookcold open00:00
Jev intro
promiseJev intro01:03
race benchmark
valuerace benchmark05:06
Ronacher quote
valueRonacher quote04:51
Wikipedia race
valueWikipedia race08:00
Gateway numbers
ctaGateway numbers10:22
Frame Gallery

Visual moments.

cold open00:01Frame at 00:13 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions00:13Frame at 00:25 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions00:25Frame at 00:34 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions00:34Frame at 00:48 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions00:48Jev intro00:57
cold open00:01Frame at 00:13 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions00:13Frame at 00:25 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions00:25Frame at 00:34 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions00:34Frame at 00:48 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions00:48Jev intro00:57Frame at 01:06 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions01:06Frame at 01:17 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions01:17Frame at 01:29 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions01:29Frame at 01:49 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions01:49Frame at 01:58 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions01:58Frame at 02:06 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions02:06Frame at 02:17 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions02:17Frame at 02:27 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions02:27Frame at 02:39 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions02:39Frame at 02:50 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions02:50Frame at 02:56 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions02:56Frame at 03:11 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions03:11Frame at 03:22 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions03:22Frame at 03:35 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions03:35Frame at 03:44 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions03:44Frame at 03:52 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions03:52Frame at 04:08 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions04:08Frame at 04:18 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions04:18Frame at 04:26 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions04:26Frame at 04:45 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions04:45Ronacher quote04:51race benchmark05:06Frame at 05:13 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions05:13Frame at 05:29 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions05:29Frame at 05:35 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions05:35Frame at 05:42 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions05:42Frame at 05:57 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions05:57Frame at 06:08 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions06:08Frame at 06:15 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions06:15Frame at 06:30 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions06:30Frame at 06:41 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions06:41Frame at 06:52 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions06:52Frame at 07:09 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions07:09Frame at 07:13 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions07:13Frame at 07:21 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions07:21Frame at 07:37 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions07:37Frame at 07:48 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions07:48Wikipedia race08:00Frame at 08:10 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions08:10Frame at 08:21 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions08:21Frame at 08:32 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions08:32Frame at 08:47 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions08:47Frame at 08:54 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions08:54Frame at 09:10 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions09:10Frame at 09:11 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions09:11Frame at 09:25 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions09:25Frame at 09:38 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions09:38Frame at 09:51 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions09:51Frame at 10:00 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions10:00Frame at 10:15 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions10:15Gateway numbers10:22Frame at 10:35 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions10:35Frame at 10:39 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions10:39Frame at 10:55 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions10:55Frame at 11:04 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions11:04Frame at 11:17 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions11:17Frame at 11:24 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions11:24Frame at 11:42 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions11:42Frame at 11:49 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions11:49Frame at 12:01 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions12:01Frame at 12:16 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions12:16Frame at 12:23 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions12:23Frame at 12:34 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions12:34Frame at 12:42 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions12:42Frame at 12:54 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions12:54Frame at 13:03 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions13:03Frame at 13:18 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions13:18Frame at 13:24 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions13:24Frame at 13:40 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions13:40Frame at 13:45 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions13:45Frame at 14:02 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions14:02Frame at 14:13 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions14:13Frame at 14:20 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions14:20Frame at 14:35 from Jev: TypeSafe AI's Non-LLM "System 1" Model for Machine-Only Decisions14:35
One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.

20:18
Riley Brown · Demo

JEV: How It Works and What You Can Build

TypeSafe's new model can't write a sentence. It just makes fast, cheap decisions, and a growing list of builders are already wiring it into email triage, trading bots, and self-driving demos.

September 18th