Benjamin

new blog post! why do LLMs freak out over the seahorse emoji? i put llama-3.3-70b through its paces with the logit lens to find out, and explain what the logit lens (everyone's favorite underrated interpretability tool) is in the process.

link in reply!

This fails to take into account the legal realities. Data can be "owned" in many forms: Copyright, GDPR, Trademarks, ....

Correctly assigning rights in different jurisdictions is an impossible task.

It is possible to expand control over dissemination of knowledge to a dystopian level, of course.

Flawed thinking helps reasoning models learn better!

Meta's RECAP: an RL post-training method that trains models to override unsafe reasoning, reroute to safe & helpful answers, and stay robust - all without extra training cost.

Professor Stiglitz’s Contributions to Debates on Intellectual Property

Monopoly rents from IP total almost $1 trillion per year and account for 40% of all corporate profits in the US. Eliminating it would almost completely reverse inequality growth since 1980 and save the average person $3,000 per year. A 100% annual return on investment! cepr.net/publications...

Video models are zero-shot learners and reasoners Fascinating new paper from Google DeepMind which makes a very convincing case that their Veo 3 model - and generative video models in general - serve a similar role in …

Made some notes on the new DeepMind paper "Video models are zero-shot learners and reasoners" - it makes a convincing case that generative video models are to vision problems what LLMs were to NLP problems: single models that can solve a wide array of challenges simonwillison.net/2025/Sep/27/...

obvs you're on this, but here's some lesser known accounts that are good. i need to spend a lot more time on this go.bsky.app/LFAZcGE

Incidentally, what benefit do you see from that right that outweighs the harms?

And you hope that the AI Act will protect the right to rectification?

That's unfortunately the usual response. You cannot say how people are protected. You can't even name specific rights that you hope are protected.

And yet, you know enough to attack the doubting heretic.

And how would that protect people?

The GDPR was the mother of all trainwrecks. It's weird how it gained such a cult following.

In my experience, tech-literate people don't know what these laws actually say. While law-literate people don't understand how it interacts with technology to produce bad outcomes.

That is all under the assumption that the AI Act and other laws holding back development do actually protect citizens. That is very doubtful.

The economic and cultural harm is already becoming obvious.

They're on here! @arcinstitute.org

You could use 20-year-old technology, but only read ~100-year-old books.
You could have computers and cellphones, but they wouldn't work because the software is still under copyright.

The point would be quickly made, after it would just be about how racist people were 100 years ago.

Weil Probleme lösen schwerer ist, als populistische Reden zu schwingen.

Locality in Image Diffusion Models Emerges from Data Statistics ArXiv link for Locality in Image Diffusion Models Emerges from Data Statistics

A study shows that locality in diffusion models arises from dataset statistics, not network architecture, bridging theory and practice in generative modeling. The new analytical denoiser significantly outperforms traditional methods, enabling novel image generation. https://arxiv.org/abs/2509.09672

I have to think of Liebig's law of the minimum in biology. "It states that growth is dictated not by total resources available, but by the scarcest resource (limiting factor)."

Compute: Matter of money

Algorithm: Same for everyone

Curated data: Variable and hence the obvious limiting factor

The appeal seems to rest on the idea that the teacher acted on behalf of the financier, so that the training that happened outside of Germany can be pinned on the teacher and his club, LAION.

Seems a stretch.

There was some communication between the school teacher who created the database, the university students who created the image diffusion model, and the (UK) financier who funded the training.

LAION created databases with links to images and also computed metadata for the images. It is the latter use of the images that was at issue in this case.

US readers may find that incredible. It explains a lot about Europe's inability to build a Big Tech industry.

I don't think it's likely that the LAION case will reveal much.

Case C-250/25 is the one to watch (per @technollama.bsky.social ).

Because it's about rephrasing news, I wonder if it might have repercussions for Wikipedia, Reddit, ...

"Research organisations and cultural heritage institutions" get special privileges in that directive (Article 3). They do not have to abide by opt-outs, mostly. Curiously, programs seem to be omitted.

Perhaps the AI Act's explicit reference to Article 4 overrides this?

I hope you will be able to tell me that I am being overly dramatic.

2 more things: "Union law on copyright" is an unclear phrase. There's a copyright directive and national law. Maybe directives are to be treated as law for the purposes of the AI Act?

EU Data Act starts to apply in the EU

Data Act by Hamburg and Bavaria DPAs lnkd.in/dyfRkSaU and position on Data Act/GDPR: lnkd.in/d-a6VejQ.
Guidelines on vehicle data lnkd.in/d4e6g47R
Draft MCTs lnkd.in/dpv4-3bv
Draft SCCs lnkd.in/dmnpQEtS

Offers that are not in an EU language are probably out of scope, or maybe if no EU payments are accepted.

Otherwise, the way to be out of scope is to explicitly exclude EU use in the TOS/license and/or geo-block.

That's going to cause so much international drama.

If a model is downloadable, available via API, or chatbot interface, then it's placed on the market "irrespective of whether they are established or located within the
Union or in a third country". It's crazy.

Internal use may count as placing on the market. Pretty sure that platform moderation counts.

Research is excepted, but that only means R&D on AI models before they are released. Not sure how far that exception may stretch.

Only "providers" need to abide. Providers are also responsible for 3rd party scrapers.
The provider responsibilities may shift downward in a few circumstances.

What's quite alarming is the wide meaning of placing on the market.

The EC has released "guidelines" with its take on the Act.

digital-strategy.ec.europa.eu/en/library/g...

Interplay between the DSA and the GDPR: EDPB adopts guidelines www.edpb.europa.eu/news/news/20...