<- all tokdocs
2026.09.22@promptly_media

Jev is real and so is the man explaining it, but the clip is a mirrored Latent Space cut and the "intelligence per dollar" pitch is priced against Claude Fable 5.1 at $10 per million input tokens while Claude Haiku 5.5 sits 2.4x away at $0.10

AI
Watch on TikTok

View on TikTok ->

A 44.768 second portrait clip, 720x1280 SDR, MP4 container reported by ffprobe as mov,mp4,m4a,3gp,3g2,mj2, HEVC Main profile level 120 video in yuv420p at a constant 30/1 fps across 1,343 frames at 436,325 bps, HE-AAC stereo audio at 44,100 Hz and 96,298 bps across 967 frames, 539,516 bps overall, 3,019,134 bytes on disk, pulled as format bytevc1_720p_539504-1; posted at epoch 1790110800, which is 2026-09-22T21:00:00Z, by promptly_media (uploader id 7686441600244155405, nickname promptly_media), audio credited as "original sound - promptly_media", description 66 characters and byte-identical to the title; at capture time, the mtime of metadata.json on 2026-10-19 11:16:17 UTC, the counters read 1,670 views, 26 likes, 4 comments and 6 reposts. I read 13 of the 22 extracted frames (001, 002, 004, 006, 008, 010, 012, 014, 016, 018, 020, 021, 022, covering 0, 2, 6, 10, 14, 18, 22, 26, 30, 34, 38, 40 and 42 seconds) and pulled two extra stills at 31.5 s and 36.0 s that I horizontally flipped and cropped to read the speaker's shirt. The transcript is 159 words across 16 segments. Every frame from 0 s to 38 s is one static shot: a man with crimson-dyed hair in a hot pink fur coat over a lavender t-shirt, seated in a mesh office chair in front of pleated purple acoustic curtains, speaking into a boom-mounted Shure microphone, with a white paper cup at the bottom left and a red-brown desk edge across the foreground, and a white Promptly "P" watermark at the top left. Frames 001 and 002 carry a pink title card reading "The Creator of Jev / Explains Its Purpose"; frame 004 at 6 s replaces it with a lower third reading "Diogo Almeida" above "TypeSafe co-founder & CEO", which is already fading by frame 006 at 10 s. The burned-in captions track the speech, reading in order "the thing we really really", "really want is to have like", "integrate it with software", "the deep internals / of the model", "Jev is our first large / programmable model", "Jeff is meant to / be optimised for", "and it's optimised for / intelligence per dollar", "what is the most important", "and Jev will be the / name of models", "there's other ways to / optimize it like ML" and "it's all about trade offs / and"; the 18 s caption misspells Jev as "Jeff", and the captioner switches from British "optimised" at 18 s and 22 s to American "optimize" at 36 s. Frames 021 and 022 at 40 s and 42 s drop the footage entirely for a yellow Promptly logo on a magenta and orange gradient. The speaker's shirt text is mirror-reversed in every frame; flipping the image horizontally resolves a quadrant crosshair graphic with "...FS" over "JEVON[S]" on the top row and "...AI" over a white cube logo and "TYPESAF[E]" on the bottom row, with the left column of each row hidden behind the open coat, which establishes that the source footage was flipped before the clipper added its own correctly-oriented graphics. Spoken, the argument runs: the way to make AI powerful is to integrate it with software, so the model's deep internals are designed for software rather than for people; Jev is the company's first large programmable model, also callable a System 1 model, optimized for intelligence per dollar, which is where the Jevons reference in the name comes from; future models carrying the name will sit on that same frontier, with everything else a trade-off against reliability, cost, calibration and speed. The last transcript segment ends at 40.700 s against an audio stream of 44.768005 s, so nothing is timestamped past the end of the audio and there is no Whisper overrun to discard; the unspoken final four seconds are the Promptly outro card.

The lower third is accurate but the title overstates it: Almeida co-founded TypeSafe with two other people

At 6 s, frame 004 shows the card "Diogo Almeida / TypeSafe co-founder & CEO". The spelling is right. The Latent Space episode that this footage comes from is titled "Jev: System One models for Prod, not God, with Diogo Almeida, CEO, TypeSafe AI", and TypeSafe's own material and the trade coverage of its funding round use the same spelling throughout.

The TikTok title does not stop at co-founder. It calls him "The Creator of Jev". TypeSafe has three founders. SiliconANGLE's coverage of the company's stealth exit names Diogo Almeida as CEO alongside Erik Gafni and Sasha Sheng, and the Wikipedia entry on the model lists the same three. Almeida is the CEO and the public voice of the model, and the clip is a real recording of him describing it, so the title is compressed rather than fabricated. Jev came out of a three-founder company that spent roughly two years in stealth.

One note on how this was checked. The usual route for "this person created this project" is the repository: commit history, the author of the initial commit, the LICENSE file. That route does not exist here. TypeSafe publishes documentation at docs.typesafe.ai and a hosted API, and nothing on typesafe.ai or in the docs points at source code or a license grant; the Wikipedia entry states the licensing as proprietary software with architecture and weights unpublished. Authorship is therefore established from the company record and the podcast appearance, and anyone expecting a GitHub project named jev with a creation date and an MIT or Apache header will not find one.

Almeida's prior work is the strongest supporting evidence that the title is directionally right. SiliconANGLE reports he worked at OpenAI on reinforcement learning from human feedback, InstructGPT, ChatGPT and GPT-4; the Wikipedia entry puts that at four years. Jev is trained with what TypeSafe calls Reinforcement Learning for Calibrated Decisions, an explicit alternative to RLHF. The man selling a replacement for RLHF helped build RLHF.

The footage is a mirrored cut of a podcast published the day before the TikTok

The clip carries no attribution inside the frame beyond the Promptly watermark and the end card. The Latent Space episode with Almeida was published on 2026-09-21; this TikTok was posted at 2026-09-22T21:00:00Z. The quotes match the episode almost word for word, including the hedge Whisper preserved at 13.18 s. Latent Space has "Jev is our first large programmable model, or a System 1 model, whatever you want to call it"; the transcript here reads "JEV is our first large programmable model or System 1 model, whatever you want to call it." Latent Space has "the thing we really want is to have, like, AI, like, be as powerful as possible, and we think the way to do that is to integrate it with software", against this transcript's 0.00 s and 4.92 s segments.

The mirroring is the clearest tell that this is a repost rather than original footage. The shirt graphic reads backwards in every talking-head frame. Flip it and it resolves into legible type. The lower third, the title card and the captions all read correctly, which means the flip was applied to the interview footage and the clipper's own graphics were composited on top afterwards. Horizontal flipping is a routine move for accounts republishing someone else's video.

The metadata offers no route back. The audio is credited as "original sound - promptly_media" with the artist field set to promptly_media, though the sound is Almeida talking on someone else's podcast. TikTok credits whoever uploaded the file, so that is a platform artifact rather than a deliberate claim.

Every spoken claim in the clip holds up against TypeSafe's own documentation

This is the part the video gets right, and it is worth being specific about it rather than waving at it.

At 13.18 s Almeida says "JEV is our first large programmable model or System 1 model, whatever you want to call it." TypeSafe's homepage describes Jev as its "first public System One Model, optimized for automation", and the documentation calls it "the first System One model", designed for "fast, structured decisions that software can use directly". The documentation exposes three primitives: Choice, which returns a selected option with probabilities and a confidence value; Score, which returns a rubric-based score the same way; and Noul, which returns a 0 to 1 truth value for a statement. There is no text generation path. That matches the claim at 7.44 s and 11.58 s that the deep internals are built for software consumption rather than for a human reader.

At 18.12 s and again at 22.68 s he says Jev "is meant to be optimized for intelligence per dollar"; the homepage carries the line "Jev's intelligence per dollar is literally off the charts" next to a chart. At 30.42 s he says "JEV will be the name of models that will be on the frontier of intelligence per dollar", which is forward-looking and not checkable today. At 24.44 s and 28.02 s he names the trade-off set as "reliability, cost, calibration, and speed", matching the episode, where he argues calibration matters more than determinism. That ordering shows up in the product: every Jev primitive returns a confidence value alongside the answer, which is only useful if the confidence is calibrated.

One correction, and it belongs to the transcription rather than to the video. At 21.60 s Whisper renders the name as "Jevin's Paradox". The reference is the Jevons paradox, named for the nineteenth-century English economist William Stanley Jevons, and the Wikipedia entry on the model states that the model name comes from Jevons and that Almeida intends it to signal that cheaper machine intelligence will be deployed far more widely. The shirt confirms the spelling independently: un-mirrored, the top right quadrant reads JEVON with the final S running under the coat. The burned-in caption at 18 s separately renders the model name as "Jeff". Neither error is the video's argument, but anyone quoting the auto-transcript would mis-source the name.

One pricing claim is exact, one speed claim does not match the numbers printed beside it

TypeSafe's homepage states "$42 Per Billion input tokens", which is $0.042 per million, and directly beneath it "238x Lower input price than Claude Fable 5.1". Anthropic's published pricing page lists Claude Fable 5.1 base input tokens at $10 per million. Ten divided by 0.042 is 238.1. The arithmetic is clean and the comparison model is named rather than hidden behind "frontier LLMs". That is better disclosure than most launch pages manage.

The framing is the issue. Claude Fable 5.1 at $10 per million is the top of Anthropic's published price list, tied with Claude Mythos 5.1. The same page lists Claude Haiku 5.5 at $0.10 per million for prompts up to 100,000 tokens. Jev's input limit is 64,000 tokens according to the Wikipedia entry, which sits entirely inside that tier. Against the model you would actually reach for on a high-volume classification job, Jev's input price advantage is 2.4x, not 238x. Both numbers are true. They describe different purchasing decisions, and the one on the homepage is the one chosen against the most expensive possible alternative.

The free output tokens deserve the same scaling down. Jev does not generate text; its outputs are a choice index, a score, or a probability, so free output is a consequence of the architecture rather than a pricing concession. TypeSafe is candid elsewhere about what it cannot yet prove, saying it cannot demonstrate the pricing is unsubsidized and expects costs to fall rather than rise.

The homepage also displays "193.6x Faster, 444.6x Cheaper" directly above a two-column comparison: 0.114 s against 8.566 s, and $0.000081 against $0.013880. Divide them. 8.566 divided by 0.114 is 75.1. 0.013880 divided by 0.000081 is 171.4. Neither ratio is the headline number sitting above it.

The resolution is in TypeSafe's own launch post, which the homepage links with a footnote reading "based on workflows for System One tasks". That post says "This is where the claims of 193.6x faster, 444.6x cheaper on our home page come from", and describes four workflows of which only the simplest is shown, measured against "the average of the smartest models (in this case, Astra and Fable)". So the displayed pair is one example workflow and the headline is an aggregate across four. That is a legitimate methodology and it is published. The homepage does not say so next to the numbers, which means a reader who checks the arithmetic in front of them gets a mismatch and has to go find the blog post to learn why.

The same post volunteers the obvious objection: "they were made by individuals on our model capabilities team, some bias could exist", with the added qualifier that the gains "are on the higher end of real world gains". The comparison baseline is TypeSafe's own System One wrapper around an LLM, which the post concedes "tends to be slower and more expensive than giving decisions without probabilities". A vendor benchmark built by the vendor's model capabilities team against the vendor's own wrapper is not independent evidence, and TypeSafe says as much.

An independent read exists and is more mixed. A survey of third-party audits reports that Every measured 777 judgments in under 0.7 seconds for roughly a quarter of a cent, around 25x faster than Claude Fable 5.1 at a small fraction of the cost, while catching 6 of 7 planted defects against Fable's 7 of 7. The same survey reports an out-of-distribution calibration study finding an expected calibration error of 0.107 against a 0.024 noise floor, roughly 4.4 times the floor, while the same model stayed in the 0.024 to 0.032 range on public benchmarks. If that holds, the calibration Almeida ranks above determinism at 28.02 s is task-dependent, which is the main caveat for anyone wiring Jev's confidence values into an automated threshold.

The 586-page audit figure circulating with this model is a third-party claim nobody has reproduced

A companion video in this corpus, from @itsborjao, states that Jev audited 586 pages in 45.1 seconds. I traced it rather than repeating it. The earliest attributable version appears in an SEO write-up by Ryze AI, which says the first SEO demo of Jev came from Distribb's founder and describes it as 586 pages with an internal link map rebuilt in 45.1 seconds, 8,790 yes or no calls, at a cost of $0.21, against Claude Opus 5 getting through 21 pages for $1.43 on the same clock. Ryze AI attributes it explicitly: "It is a first-hand claim posted on X by Distribb's founder in Jev's launch week in September 2026", and disclaims it just as explicitly: "It has not been independently reproduced, and Ryze AI has not run it."

So it is not a TypeSafe benchmark, not an independent test, and not reproduced. I could not reach the original X post to read the methodology, so the numbers stay unverified here. The shape of the workload fits what Jev is built for, since a link-map audit decomposes into thousands of independent yes or no judgments against a shared state. The specific figures should not be quoted as fact.

TypeSafe's own documentation gives a reason to be careful with a claim of this shape. The jaggedness page for jev-1.13, last reviewed 2026-10-02, states that the model "does not count reliably" across characters, term occurrences and list items, that it "reads dates as text, not as ordered quantities", that it "answers the question you wrote, not the one you meant" with negations and implied conditions read at face value, that instructions "carrying double negatives or complex indirection are answered less reliably", and that "accuracy falls as the state grows with content unrelated to the decision". A page-count claim rests on counting, which is the first item on that list.

Nothing in this video shows the product running

Across the 13 frames I read, plus the two additional stills, there is no terminal, no code editor, no API response, no dashboard, no chart and no screen recording of any kind. Every frame from 0 s to 38 s is the same locked-off shot of one person talking, and the last four seconds are a logo card. The clip is testimony about a model rather than a demonstration of one, and the 586-page figure circulating alongside it appears nowhere in it.

For the record on dates, since the clip gives none: TypeSafe exited stealth on 2026-09-15 with a $40 million seed led by DCVC, reported by SiliconANGLE at a $200 million valuation per Forbes, and released Jev the same day in limited early access. General availability followed on 2026-09-21, the day the Latent Space episode went out and the day before this TikTok. Current stable version is jev-1.13.0.

Key Takeaways

  • Verified: The lower third at 6 s reads "Diogo Almeida / TypeSafe co-founder & CEO" and both the spelling and the title are correct. He is TypeSafe's CEO and previously spent about four years at OpenAI on RLHF, InstructGPT, ChatGPT and GPT-4.
  • Partial correction: The TikTok title calls him "The Creator of Jev". TypeSafe was founded by three people: Diogo Almeida, Erik Gafni and Sasha Sheng. He is the CEO and the model's public explainer, not its sole author.
  • Clarification: There is no Jev repository, no commit history and no open-source license to check him against. The Wikipedia entry lists the model as proprietary software with architecture and weights unpublished, and nothing on typesafe.ai or docs.typesafe.ai offers source code. Authorship was verified against the company record and the podcast instead.
  • Verified: The spoken purpose matches TypeSafe's own material. "First large programmable model or System 1 model" at 13.18 s lines up with the homepage's "first public System One Model, optimized for automation" and the documented Choice, Score and Noul primitives, which return typed values with calibrated probabilities and no text generation path.
  • Correction: The auto-transcript renders the name origin as "Jevin's Paradox" at 21.60 s and a burned-in caption renders the model name as "Jeff" at 18 s. The reference is the Jevons paradox, after the economist William Stanley Jevons.
  • Context: The clip is horizontally mirrored footage from the Latent Space episode published 2026-09-21, reposted here on 2026-09-22 with a Promptly watermark, a Promptly end card and the audio credited to "original sound - promptly_media".
  • Verified: TypeSafe's "238x Lower input price than Claude Fable 5.1" is arithmetically exact. Jev input is $0.042 per million tokens and Anthropic's published price for Claude Fable 5.1 base input is $10 per million. Ten divided by 0.042 is 238.1.
  • Clarification: That 238x is measured against the top of Anthropic's price list. Anthropic also publishes Claude Haiku 5.5 at $0.10 per million input tokens for prompts up to 100,000 tokens, which covers Jev's reported 64,000 token input limit. Against that model the input price advantage is 2.4x.
  • Correction: The homepage headline "193.6x Faster, 444.6x Cheaper" does not match the figures printed under it. 8.566 s divided by 0.114 s is 75.1x and $0.013880 divided by $0.000081 is 171.4x. TypeSafe's launch post explains that the headline aggregates four workflows while the display shows the simplest one, and concedes the workflows "were made by individuals on our model capabilities team, some bias could exist".
  • Unverified: The claim that Jev audited 586 pages in 45.1 seconds, repeated in a companion video by @itsborjao, traces to a post on X by Distribb's founder during launch week. Ryze AI, which reported it, states it "has not been independently reproduced, and Ryze AI has not run it." I could not reach the original post.
  • Context: TypeSafe's own jaggedness page for jev-1.13 states the model "does not count reliably", "reads dates as text, not as ordered quantities", answers literally through negations, and loses accuracy as irrelevant state grows. An independent study reported an out-of-distribution calibration error of 0.107 against a 0.024 noise floor.
  • Context: No frame shows a terminal, an API call, a dashboard or any demonstration of Jev running. It is 38 seconds of talking head plus a 4 second logo card.

Resources

I attempted to fetch MarkTechPost's launch coverage at marktechpost.com and received HTTP 403, so nothing here rests on it.

Published September 22, 2026. Writeup generated from a favorited TikTok.