ENPT

AIE CODE Day 1: Michelle Chen on Clef and Clairic, Diogo Almeida on jevolution

Last updated:

Source · Model card: Jev

Two talks from the AIE CODE Day 1 livestream, 11 Nov 2026. This is a recap of what was said on stage. It is not a product launch. The recording runs about nine hours and has no official captions. The wording below follows an automatic transcript and stays conservative where a name came out wrong. Agenda: schedule and sessions.json, read 12 Nov 2026.

Michelle Chen (Cloudflare)

Agenda title “cloudflare is a neolab”, 9:50–10:10 PT, mainstage. In the livestream the talk runs from about 0:48 to 0:59:52 and ends on “what are you going to build today?”.

Sequence she gave:

  • Jev launched on 15 Sep 2026, closed-source, and Cloudflare could not host it.
  • On 18 Sep, three days later, they already had a first experiment.
  • After a week of Neolab deliberation they trained over a weekend (48 hours) and launched the following week as Clef.
  • About three weeks later they launched Clairic. The automatic transcript wrote “Cleric”. The published name is Clairic. She described a System One extraction model: one forward pass, unstructured input (an email, a tweet, Slack, and also an image) in, typed JSON out (extract, compute, estimate, forecast).
  • She said they launched four models in three weeks. This recap names Clef and Clairic because those are the ones she described. It does not fill in the other names.

The catalog date for Clef stays 1 Oct 2026. Her talk gave the sequence, not a calendar date for that weekend. Clairic’s blog on this site is 19 Oct. Scores on the Clef page stay 7.2 / 7.7 / 37. The talk is not a new board result.

She said Clairic shipped as an early preview and that a new launch is coming out for it. That is not GA, and it is not open weights on 11 Nov. Clairic still has no catalog page. The catalog covers decision models (choice, score, boolean). Clairic answers a different question. The 19 Oct news keeps that scope decision; the talk does not reopen it.

CI-triage demo (UnverifiedClaim)

An agent with 40 tools. The typical LLM loop, which she illustrated with GLM 5.3, sends the tools back and forth. The augmented path: Clef Flash prunes the tools (10 of 40), Clairic pulls values (a PR number, an error from 300–600 log lines), and one LLM does the diagnosis and the action.

Figures she stated. They come from the talk. They are not an independent benchmark, and they are not ModelSystem.One measurements.

Setting What she said
Live, on stage 6 LLM calls vs 1; 40,000 LLM tokens vs 1,800
Prepared, offline, not conference Wi-Fi 9.2 s vs 17.2 s, almost 2× faster, 72% less expensive, LLM calls from five down to one, about 94% fewer tokens

The live line said six calls. The offline recap said five down to one. Both stay in the table.

Slides also listed routers, fraud / spam / phishing decisions, and Clairic extracting from web pages and support tickets. Those are examples on the slide.

After about 1:00:13 the next speaker starts, on “Hundreds of pull requests a day”. That tail is not Michelle’s.

Diogo Almeida (TypeSafe)

Agenda title “I love you so jevving much”, 2:05–2:23 PT, mainstage. Livestream from about 5:07. The talk ends around 5:28:54–5:29:50. It was a manifesto. It was not a customer list and not an ARR announcement.

What he argued:

  • LLMs today are optimized for assistance, not automation, in a human-in-the-loop / RLHF framing. The result is jagged: strong on some tasks, useless on others.
  • “Jevolution”: optimize for the task, and for narrow, reliable intelligence. His illustration was Einstein 95% of the time and Mr Bean 5%. Reliability, he said, does not show up well on benchmarks, and that is why people love Jev. The automatic transcript garbled a French phrase. The line is “je ne sais quoi”. This page does not quote the garbled form.
  • Jev as a foundation for smart software, against just-in-time software and a chatbot on the side. The field, he said, “isn’t jeving hard enough.”
  • A blue-LED analogy: a new primitive that composes. He described machine-native intelligence as adding a fourth “semantic logicate” to software. That phrase is his, from the automatic transcript.
  • A pitch for “software research”: engineering, not vibe coding. He said “we have raised a bunch of money” while talking about funding that direction. The sentence has no figure. It is not revenue and it is not ARR.
  • Close: a TypeSafe “stop-gen AI” baseball bat as a prop; a Jev-like model inside a linter, to hold semantic and type-safe properties; later talks would cover automating everything; thanks to early believers.

He did not say TypeSafe had crossed $100M ARR. He did not name production customers. In sessions.json the line “very profitable and above $100M ARR” is Quinn Slack’s Amp talk on 12 Nov, “A Company Built Around the Agent”, 4:30–4:50 PT. Different speaker, different day.

Scores on the Jev page stay 6.0 / 8.1 / 73.

What this stage was not

Gemini Decision 1 was not a talk on this livestream, and Sundar Pichai and Logan Kilpatrick are not on the 11 Nov mainstage agenda. Google’s announcement is the same calendar day and is already on the site: Gemini Decision 1 and the 11 Nov news. This recap does not rank that launch.

The Latent Space AINews put Diogo, later customer talks and Amp’s agenda line in one paragraph. Richard MacManus’s written recap is about what engineers are building. That article is separate from this talk.

The agenda lists later Day 1 mainstage slots whose blurbs discuss Jev in production at Glean (Chinmay Goyal, 2:30 PT), ZoomInfo (Brian Kennedy, 3:06 PT) and Workday (Scott Roberts, 3:24 PT). This page does not report those talks. Read on 12 Nov, sessions.json has no “UiPath” string. The 2:48 PT slot is “Your Automation Is Overthinking It” (Chibi Vikram, Kevin Bernstein). The blurb does not name a company.

On this site