AI leaks, news, projects, etc. Manager of @aitrackerbot 1.6m+ downloads on @huggingface https://xitter.cf/t.co/46uCxrsCpA @Alibaba_Qwen Ambassador
Joined August 2025
- Tweets1.2K
- Following396
- Followers753
- Likes2K
SSH retweeted
Update: deepseek-flash on DeepSeek's API is still the public V4.1 Flash, not the 1017 snapshot. Its outputs match a third-party host serving the open weights as closely as V4 Pro matches itself across the same two hosts 🧪
SSH retweeted
Couldn't even set it up 🥹 Banned in less than an hour. Not a single post sent. @Reddit WHY???? Nothing I did was against their terms, and I was setting it up properly not using browser agents.
SSH retweeted
MISSINGNO. update: the weights are Qwen's, not Xiaomi MiMo's.
We had it copy 30 rare tokens that Qwen never trained but MiMo did. It fails like Qwen: 38% copied, vs 36-53% for Qwen models and 70-83% for MiMo.
SSH retweeted
Yay I got #1 place
SSH retweeted
OPENAI 🔥: A new gpt-rosalind-discovery model has been spotted on the API pricing page. The model hasn’t been publicly announced so far.
The model has the same pricing as gpt-rosalind-research while its exact purpose is yet unclear.
Should it be specifically a drug discovery model?
That would be huge news 👀
SSH retweeted
OpenAI's API pricing page now lists gpt-rosalind-discovery, a GPT-Rosalind model it hasn't announced yet.
It's priced like gpt-rosalind-research: $5 / $25 per 1M tokens. GPT-Rosalind is OpenAI's life-sciences model, open only through its trusted-access program.
SSH retweeted
I'll be real: I have zero faith in Command Code or Ahmad Awais as a founder. It's one of my most regretted investments.
I'm not interested in supporting Command Code. Ahmad is more likely to send passive aggressive emails to our users than he is to ship something useful.
That said, we have ACP registry in the nightly. If that doesn't work, that's a failure in their implementation, and I will not expend any effort into fixing it on our end.
SSH retweeted
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
SSH retweeted
command code is legit extremely sketch…I remember once reading on their TOS that any input you pass into it gives them the right to MODIDY, REPRODUCE, AND DISTRIBUTE “on and through the service” and make it available to other users. Nothing you input to them is safe. : "However, by posting content via the service, you grant us a license to use, modify, publicly perform, publicly display, reproduce, and distribute such content on and through the service. This license also allows us to make your content available to other users of the service, who may use your content in accordance with these terms."
SSH retweeted
I know the company that made this absolute demon.
Union Alpha: Circuit & Chisel, the team behind Unbiased/Pareto.
Their Sept 14 release-note PR (still unmerged) names "union-alpha" as a per-account public name for Pareto.
github.com/circuitandchisel/…
SSH retweeted
A few corrections to the callout:
Our reference was **GLM-5.3-Flash**, whose processor uses a **28px grid**. We also tested irregular sizes like **257×383 and 1152×896**, so the suggestion that our results came from standard image sizes doesn’t hold.
Our evidence includes **26/26 text+image probes**, **13/13 image conditions**, and an exposed **vLLM build associated with GLM-5.3-Flash deployments**. We also recorded Qwen-like counts with cached tokens, so the cache explanation has exceptions.
We reproduce the Qwen-like text counts too. **Qwen is plausible, but ruling out GLM is premature.** Our original post explicitly kept the owner and exact model unconfirmed—that distinction matters.
