Michiel Bontenbal

accidentally caused Nano Banana's system prompt to leak, but in a fun way

Gemma 3n fully available in the open-source ecosystem! We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co/blog/gemma3n

Do AI Models Perform Human-like Abstract Reasoning Across Modalities? OpenAI's o3-preview reasoning model exceeded human accuracy on the ARC-AGI benchmark, but does that mean state-of-the-art models recognize and reason with the abstractions that the task creators inten...

Do AI reasoning models abstract and reason like humans?

New paper on this from my group:

arxiv.org/abs/2510.02125

🧵 1/10

Seems to me a great pelican, if not the best, but yeah 6 min is too much.

Performing great on the benchmark says little about model quality.

We all have one, but are too ashamed to admit it.

My excuse is 'We do it for the kids'.

Lees ook ‘Active Measures’ van Thomas Rid over de 100+ jaar ervaring van de KGB.

Sorry, that was too quick. Including the tail it was 3 hours!

The demonstration passed my house. The red line demo lasted 2, 5 hours!

AI Agents with MCP Since its release in late 2024, Anthropic's Model Context Protocol (MCP) has redefined how developers build and connect AI agents to tools, data, and each other. AI Agents with MCP... - Selection from AI Agents with MCP [Book]

and the addition of chapter 5, the first of 3 chapters covering the ins and outs of building MCP servers. If you subscribe to @oreilly.bsky.social's learning platform, you can spend your weekend with the book now here: learning.oreilly.com/library/vie...

I haven't posted about Model Spec's in a while, but Dean gave me a shoutout on my earlier writing on them, so its time to say definitively again that every frontier lab should have a model spec. It builds long term trust with users, developers and regulators.

Yes agree that would be nice addition to this figure!

Every week there are new models, but to get further in AI as a community we need open source models, that also disclose the data they are trained on.

#learnAI

Given the ‘strategic autonomy’ debates here in 🇪🇺, I do not see a full takeover happen.

We're announcing a new update to MTEB: RTEB

It's a new multilingual text embedding retrieval benchmark with private (!) datasets, to ensure that we measure true generalization and avoid (accidental) overfitting.

Details in our blogpost below 🧵

OpenAI employees are very excited about how well their new AI tool can create fake videos of people doing crimes and have definitely thought through all the implications of this

YouTube video by Anthropic An experimental new way to design software

Imagine with Claude

generative UI? like for real? as the user clicks buttons the UI manifests itself 🤯

youtu.be/dGiqrsv530Y

if my coding agent is the example... then my commerce agent will buy all sorts of stuff that I do not need... ;-)

Claude Sonnet 4.5 is probably the “best coding model in the world” (at least for now) Anthropic released Claude Sonnet 4.5 today, with a very bold set of claims: Claude Sonnet 4.5 is the best coding model in the world. It’s the strongest model for building …

Wrote up my initial impressions of the brand new Claude Sonnet 4.5 - I think it may live up to Anthropic's claims of being the "best coding model in the world", for the next few weeks at least!
simonwillison.net/2025/Sep/29/...

he will not leave. not by himself.

In a future of AGI, LLM's have their role. It might be an important role, but LLM alone are not enough. (Same with self-driving cars, CNN or ViT alone are not enough, but are one part of the solution).

En de raad van state is niet politiek, dus kan zich niet verdedigen. Ze weet dat ze feitenvrij wat kan roepen.

Artikel van @marchijink.bsky.social.

Groter is niet altijd beter in de wereld van AI OpenAI gooit met gigawatts en strooit met biljoenen, om de volgende versie van ChatGPT nóg slimmer te maken. Maar groter is niet altijd beter.

Groter is niet altijd beter in de wereld van AI www.nrc.nl/nieuws/2025/... (10x gratis). Met @maartengr.bsky.social )

Want to visualize the response format constraints on the LLM when working in a Jupyter notebook?
Then you might be interested in my new project `litelines`.
Litelines lets you visualize the selected path by the LLM.
It supports a Pydantic schema as a response format, as well as regular expressions.

I believe Sutton is right. There is a lot of value in LLM's and many other ML architectures, but it falls short of true intelligence.

Mooie oefening voor mijn it studenten: stel eisen op voor de btw berekening van 1) een losse appel 2) een voorverpakte salade waar ook kip en croutoms in zitten. 3) een broodje met gegrilde groentes. Succes!

🇪🇺🇪🇺🇪🇺

And new paper out: Pleias 1.0: the First Family of Language Models Trained on Fully Open Data

How we train an open everything model on a new pretraining environment with releasable data (Common Corpus) with an open source framework (Nanotron from HuggingFace).

www.sciencedirect.com/science/arti...

@berthub.eu