Jev Ultrafast clears 14k stars — JEV Brief · Sep 21
TL;DR. No TypeSafe release today. The signal is the ecosystem: browser-use/jev-ultrafast is past 14k stars and still #1 among new GitHub repos this week/month — a clean Pro split (Jev picks operation + target; a small LLM types only when needed). In parallel, open-source interface clones (SemIf, Kev, jevlike and friends) crossed multi-thousand-star territory within a week of launch. If you only have five minutes, read the deep dive, then the clone caution, then pin your model id.
Deep dive: Jev Ultrafast is the pattern, not the meme
browser-use/jev-ultrafast (live ~14,030 stars as of this morning Pacific; created 16 Sep) is a browser agent with a dynamic, indexed action space. Each observation builds an element table ([1] button…, [2] combobox…). One TypeSafe request returns operation (CLICK / TYPE_TEXT / …) and the matching target head. A small LLM runs only for TYPE_TEXT. Demo claim in the README: Zürich → London on Google Flights in ~7.1 s, with independent outcome checks — not “Jev wrote a selector.”
Why Pros care more than the star count:
- Structured state in, closed options out. No screenshots in the default loop. Model output never becomes CSS selectors, coordinates, shell, or executable JS — the executor resolves observed DOM nodes.
- Speculative fan-out in one RTT. Operation and target questions share state; only the chosen head executes. That is the System One shape applied to computer use, not a chat agent with extra prompts.
- Generation stays tiny and gated. Typing is the only string path. Decide with Jev; draft with an LLM. Same split as support triage or tool routing.
Star trackers (Seismograph / GitNova) still list it #1 among new repos this week and this month; HN covered the launch earlier in the week (~90 pts). Treat “14k in five days” as attention, not a reliability benchmark — the README’s own evidence table is three repeats of one task.
The clone wave: copy the contract, not the calibration
A same-day roundup (Pasquale Pillitteri, 21 Sep) frames what builders already feel on GitHub: the API shape escaped into open weights within a week. Live star checks this morning (approximate; move fast):
| Project | What it is | Stars (approx.) |
|---|---|---|
| TheoLeeCJ/SemIf | Open-model “semantic ifs”; WebGPU browser demo; independent, not affiliated | ~2.8k |
| jaredpalmer/kev | Qwen-based family serving /v1/systemone; point TypeSafe SDK at localhost |
~1.6k |
| vinnylarouge/jevlike | Train a small scorer; Doom/chess demos | ~1.1k |
| wfzyx/von | Local non-autoregressive drop-in claim; on HN today | ~300 |
| razorback16/openjev / ekzhang/openjev-sglang | Compatible servers on open models | ~250 each |
What clones copy: state + typed questions → option probabilities in one forward pass. What they do not copy: TypeSafe’s closed weights, RLCD training data, or production calibration. SemIf’s own README frames agreement numbers as self-reported. Kev’s Hub cards compare checkpoints to hosted Jev on frozen suites — useful, not a license to swap vendors without your harness.
Pro move ranked for today:
- Keep a frozen eval set of your questions before you trust any local server for auto-act.
- Pin versioned ids (
jev-1.13.0on TypeSafe; named checkpoints on Kev). Do not let*-latestbe the only production pin. - Data residency is a real reason to try local (tickets that must not leave the VPC). Latency memes are not.
- Schema-safe ≠ correct. A Choice still picks the wrong department; it just cannot invent a fourth one.
What people are reading (and arguing)
Ranked by how much they change how you ship this week:
- Flavio Copes — deep dive — clean operator map:
jev-1.13.0/ aliases,$0.042/ M input (output free — vendor claim), Vercel AI Gateway idtypesafe-ai/jev, and the decide-then-draft split. - HN threads on the launch — the useful disagreement is linguistic, not tribal: “can’t hallucinate” means schema matching, not factual omniscience. Pros translate slogans before they paste them into design docs.
- Side experiments — drawing-without-vision demos, MySQL “filter by meaning,” latency playgrounds. Color, not contract. Official facts stay on docs.typesafe.ai and the 15 Sep launch post.
Nothing TypeSafe-official shipped overnight that changes the public API. An honest briefing says so.
Jagged edges to keep in the wrapper
- Arithmetic, dates, counts stay in code — documented weakness for
jev-1.13, and the clone suites still struggle on deadline/date arms relative to hosted Jev. - Vision is not in the default path. Ultrafast’s strength is structured DOM state; screenshot-first agents are a different product.
- Vendor latency/price figures ($0.042 / M in; 70–500 ms e2e) remain claims until you measure on your path and region.
- Abstain before auto-act. Write the low-confidence / “leave it” branch before you celebrate a 14k-star demo.
FAQ
Is TypeSafe open-sourcing Jev?
No. Hosted API only (console + gateways). Open repos around it are tooling, agents, and independent interface clones.
Should I replace TypeSafe with SemIf or Kev tomorrow?
Only after your own frozen harness says so. Clones prove the format is copyable; calibration and ops still need your data.
Why subscribe if the blog is free?
The 09:00 Pacific email is this same article, same order — not a teaser.