Ramon Astudillo reposted
well, fuck www.anthropic.com/news/fable-m...
Who's Who of AI
Ramon Astudillo
Principal Research Scientist at IBM Research AI in New York. Speech, Formal/Natural Language Processing. Currently LLM post-training, structured SDG and RL. Opinions my own and non stationary.
ramon.astudillo.com
What they're sharing
Statement on the US government directive to suspend access to Fable 5 and Mythos 5 anthropic.com
View on Bluesky →
arXiv.org e-Print archive arxiv.org
Introducing Gemma 4 12B: a unified, encoder-free multimodal model blog.google
View on Bluesky →
Gemini 3.5 Flash: more expensive, but Google plan to use it for everything simonwillison.net
View on Bluesky →
The mysterious Hy3 LLM is topping OpenRouter Model Rankings by a large margin minimaxir.com
Notes from the AI Now Summit by Mistral koenvangilst.nl
Articles & links
👆 A paranoid LLM is ofc worse. This is just tuning a prior belief up or down. I guess you could self distill additional context for the train data e.g. "you know arxiv.org is such and such" or "this is an unknown source" with the hope it generalises (and also injecting some ba…
Ramon Astudillo reposted
12B Gemma 4!!! It's neat on its own, image and audio in, but architecturally it's super cool: instead of a visual or audio encoders, they directly train the model on vision and sound, with images using a "lightweight embedding module" and audio 'projected into the same space a…
Ramon Astudillo reposted
My notes on Gemini 3.5 Flash - 3x the price of Gemini 3 Flash but Google are planning to use it for many of their own products simonwillison.net/2026/May/19/...
new mystery model minimaxir.com/2026/05/open...
if you are wondering about Mistral koenvangilst.nl/lab/mistral-...
AI Weekly's analysis
→
- Mistral owns data centers directly, backing its full-stack infrastructure claim with physical compute assets beyond model licensing.
- Domain-specific smaller models outperformed general-purpose alternatives in speed and efficiency across Mistral's enterprise demonstrations.
- Mistral signed anchor enterprise deals with ASML, BNP Paribas, and Amazon Alexa+ as evidence of European sovereign AI demand.
Read full analysis →
Their own posts
Recent commentary
Competing against a local gpt-oss-120b 10 sample ensemble at paper understanding and, man, it's not looking great for humans
You can see how LLMs still lack a lot of implicit context. For example, when reading a document, they are bad at guessing if the document can be trustworthy. They read an arxiv paper with grandiose unsupported claims and they repeat them to you as if it were its own judgment. 👇
There is this new meme out there that is something like "AI costs more than human employees". Seems like totally the wrong take. It costs much less for the things they can do, but you can't run an org w/o human employees (for now). 👇
Now there are three levels of alerts in generative code: errors, warnings and errors and warnings that you pass to the LLM agent and don't bother about.
Got reminded about OpenAI 5 and now I see much more timelines with decent probability mass, that are pretty far from where we are now. We could call them the "no Radford" timelines.
An LLM being bad at an underspecified problem or consuming lots of tokens seems like a signal of benchmaxing
5y ago Demerzel would have felt like a completely wrong portrayal of an AI. Now it somehow feels pretty realistic.
It seems the suspected 5T models from Anthropic and OpenAI are kinda close in cybersec skills?
Their network
In Ramon Astudillo's orbit
Center = Ramon Astudillo. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.