Building your AI real estate agent @oneplaceai • Previously: Paw (API testing SaaS, exit in 2020), @electricbeastco Tesla rental platform
Europe
Joined October 2009
- Tweets1.7K
- Following1.9K
- Followers892
- Likes19.1K
The only thing more embarrassing than the only European lab @MistralAI being #13, is @spacexai ranking #10, four months after the largest IPO in history.
Mistral Large 4 by @MistralAI has landed in the Agent Arena top 15 labs!
Across +5K real-world agentic sessions, this preview model records a -6.6% net improvement score. It is 11 rankings above the previous variant, Mistral Medium 3.5 (-12.60%).
Mistral Large 4 is #43 overall in the Agent Arena. Open weights are expected at the end of October, but at its current score Mistral Large 4 would rank #13 among open models.
@MistralAI is also the only European lab in the Agent Arena top 15. Congrats to the team!
It’s actually crazy to think that every proprietary software can now legally be re-implemented in an open source version, as long as you have the binary. Every desktop app, every mobile app.
This is the doom I predicted a few days ago, coming for Photoshop. A clean-room open-source reimplementation.
No prizes for guessing that they decompiled Photoshop to source code, processed that to some kind of non-code specification language, then fed the spec to an LLM with an instruction to generate Rust.
Adobe just got nuked. And closed source is dead, dead, dead.
github.com/storytold/photocr…
The Toulouse team is killing it this year while Seattle is still in damage control mode
Here it is: @Airbus’ first A350F taxiing out for its maiden flight! ✈️ Did you know the livery is the result of a design competition won by Dublin-based graphic designer John Feehan and teenage brothers Quinnten and Ellisten Iversen? Track it on FR24: fr24.com/AIB01LD/41e2e9c6
Not gonna miss those parasites
This is 100% correct. I’ve seen human-written slop code way before GPT 3 existed. Still is today, just look at the WordPress marketshare. The code quality of my projects went up, not down, with AI, and by every metric (readability, stability/lack of bugs, tests).
100% my thoughts. But we’ll thank Jev for pioneering this new kind of model. OSS Jev will be awesome.
I don’t think Jev will survive.
@nvidia and @meta will release open-source Jev competitors within the next 30 days.
@OpenAI will add a Jev-like endpoint to their model list in the next 60 days.
And the rest of us will fine-tune custom classifiers that crush Jev and are 10x faster.
In two years we’ll look back at this flash-in-the-pan and say
“remember Jev?”
Yes 100%, Jev is filling a big void.
The problem with @typesafeaiAPI’s API is (1) we need batch inference, much more fitted for data pipelines and (2) for live requests, I think having it on the same datacenter will be a game changer (it’s so fast that network latency matters)
Yes, classifiers existed. BERT, DeBERTa, GLiNER, fine-tuned heads. Most of them were specialized. Pick the labels, train, ship. Changing the rules at runtime meant another training run.
What we did not have was a general-purpose classifier at Terra / Sonnet 5 level intelligence.
You could fake it by asking an LLM for JSON, but that was slow, expensive, and poorly calibrated.
Jev is that missing piece. Labels defined in code, typed answer plus confidence, no extra LLM call.
Classification is not new. What is new is a general-purpose classifier this intelligent, this fast, and this cheap.
- Hetzner + Kamal / Docker
- Postgres is open source + replica + daily backup (Claude configures that in <1 hour)
- Gmail + Google Workspace (sure, that’s a paid SaaS)
I rarely am positive of gov tech programs, but France casually built open source versions of Slack, Google Drive (Docs, Spreadsheets), Gmail/Outlook, Zoom, and a basic ChatGPT. And I think that's pretty cool.
github.com/suitenumerique/
Beijing, start your photocopiers
How is Jev so much faster than an LLM?
It's not about how the model is trained - it's the inference technique.
In fact, you can modify an inference engine to provide a performant Jev-like API with any open-weight LLM.
Say you are trying to ask N multiple choice questions in parallel.
In standard text generation, you decode tokens one at a time to generate a JSON representing those properties, and then parse the JSON.
In a Jev-like system, what you can do is:
- Label choices for each property 1,2,3,...
- Prefill the shared information (output shape and other context)
- Fork this inference context for each property, re-using that cached prefill
- Add an individual question to each parallel context
- Look at the logits (token probabilities) for each one, constrained to the valid choice labels
- Map the labels back to the choice values. Include the softmax of the logit itself to get a combined choice+confidence score
This means you decode one time in parallel across all questions, instead of one at a time - so the efficiency gains roughly scale with the number of questions per request (until compute is saturated).
So what makes Jev's model unique? It's because they train a small model to perform well in this specific inference setting - improving both cost and capability compared to a non-specialized model.
Ultimately though, if you have a use case Jev would be good for, you could do it with open-weight models too!
(Some speculation made here about how Jev's engine might work, but this is one way you could get similar performance)
This is truly incredible!
Prediction: it will be extremely easy & cheap to use Jev to build labeled datasets to create your own classifier models.
Heck, it could even be a library “frozen-jev”, based on RoBERTa kinda models. Jev labels, lib fine-tunes RoBERTa, da da da 🎉
Tested @typesafeai's claim that their new model Jev delivered "comparable... intelligence" to GPT-5.6 Terra on "System 1" tasks. To do this, I compare both models on multiple-choice benchmarks (MMLU, GPQA, etc.). Set reasoning=none for Terra for sys 1. Result: Jev is Terra-tier.
The problem is that Claude Code is a really good harness that works really well with..Claude models. Codex sucks badly. OpenCode is good, but still not at the level of CC. Pi is nice-ish, but no subagents, no auto-mode classifier. I want CC for all models.
This is actually insane. LLMs wrongfully made us believe that all outputs need to be text/JSON. Categorizer models are actually so powerful.
got @typesafeai’s new model Jev to play Super Mario Bros.
fast inference + structured outputs makes it surprisingly good for real time use cases.
I'm excited to see what can be done with these new models!
That shows that autoregressive LLMs are far from the final state of AI. They’re excellent, but still a brute force “hack” version of intelligence. We’ll see many projects like @typesafeai emerge. Maybe @amilabs & @ylecun will have a shot!
