Théo

Building benchsift. Mostly debugging.

[Théo @theo0px]

@theonecid It could be him

[Théo @theo0px]

AI browsers are one of the most useless things ever made, at least in the way they were designed.

OpenAI's Atlas didn't even last a year: OpenAI folded it into the ChatGPT app and now pushes its Chrome extension instead.

Perplexity's Comet is just bad. Its appeal at first was letting AI control the page for you. Impressive at the time, but never really improved, and now that control even runs on paid credits. The core problem: AI is still much slower than us at doing an action, and when speed is what matters, that's not worth much.

Today, extensions like Claude's or ChatGPT's are much better designed.

The only AI browser that's still decent is Dia, because they actually thought the concept through.

Perplexity should do what OpenAI did and shut Comet down. Given how few updates it gets, it feels like they're already slowly abandoning it.

[Théo @theo0px]

@GoogleDeepMind @GeminiApp Why don’t you develop Phone Use for Android with Gemini? Like, if it can use a computer, why not a phone? It’s the device we use the most, after all.

[Théo @theo0px]

@Ankur_Sehrawatt Okay, that’s not what your website says, actually, but I just wanted to make sure because I’m really interested in fileOS.

[Théo @theo0px]

I was offered the option to enable data sharing on Grok bot in exchange for a reset. I refused, and ever since, my limit has been melting away it was literally used up in one day. Is this normal, @spacexai @bot ?

[Théo @theo0px]

@guilippert_v4 At least you have Gemini 3.8 Flash; some users are still on Gemini 3.6 Flash.

[Théo @theo0px]

@SamsungMobile @Samsung @SamsungMobileUS @SamsungUK @SamsungFR The One UI 9.5 beta is coming, and once again it will be limited to 6 countries (Germany, South Korea, US, India, Poland, UK). It would be really appreciated to open it to everyone, like Apple does with its public betas.

Lots of users are asking for this, and more testers won't change anything given how long final releases already take to ship. I don't care if some menus are in another language: all I want is the betas.

[Elon Musk @elonmusk]

We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.

Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.

Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.

[Théo @theo0px]

@bot I was reset before this reset so I can't have a reset banked

[Théo @theo0px]

Starting next month, I'm coming back to Hermes and leaving Grok bot.

The reason I'm waiting until next month is that I'm looking for a VPS, and I found one at a good price, but it's only available from then. I already have my plan in mind for what I'm going to use it for.

I hope this return to Hermes will be awesome!

[Théo @theo0px]

@jul1anodev CUA are still the best, btw.

[Théo @theo0px]

Claude’s computer use refuses to interact with a browser beyond read-only access. I don’t understand why.

So I tried CUA, and Claude can do it just fine through that. Why does it work with CUA but not with its own computer use?

@AnthropicAI should change this. I don’t see why using a browser is a problem. This restriction is completely stupid. As things stand, @trycua is still better.

[Théo @theo0px]

@jul1anodev in the desktop application

[Théo @theo0px]

@jul1anodev It was just a matter of opening a web page, and Claude himself told me that Computer Use wouldn’t let him use it and that he was stuck in read-only mode.

[Théo @theo0px]

@HermesWatcher Thanks to @trycua

[Théo @theo0px]

@ryanvogel Hello

[Cua @trycua]

1/ Today we're launching Cua Driver Bench, together with Scale AI's AgentEnv.

Can an agent do a desktop task in the background while you keep working undisturbed? On macOS, Cua Driver fully passed 58%, Codex computer use 33%, Vercel's agent-desktop 20%, browser-use 3%.

[Théo @theo0px]

@trycua @earakely @scale_AI And doesn’t Claude Code have computer use?

[Théo @theo0px]

Picking an AI model shouldn't take 5 tabs. BenchSift compares 1500+ models on benchmarks, speed, price and context, plus coding-agent leaderboards with cost and time. Free, open source, FR/EN.
https://t.co/F4On4Ljmrq