I don't agree. For a start, this blogpost puts way too much weight on the recitals.
EU law is all in the Articles. AFAICT the recitals can be used in determining purpose, when courts have to interpret ambiguity. But so can other facts!
Thanks. My rule of thumb is simply to skip straight to the Articles, when interpreting EU law.
Recitals maybe seem weird because we don't have them in UK legislation. Maybe the nearest thing is Explanatory Notes? If you want to know what a law says, read the law itself. But if the law makes no sense (either as a lay person, or because the law is badly drafted even for legal folks) then the Explanatory Notes may offer some insight.
I see it as the complete opposite. It's not open source because of ass covering.
The biggest barrier to the licence passing open source requirements is point 0 in the above linked evaluation. Namely that it has a whole lot of "you may not use this for [nefarious purposes...]" type of statements. That seems like ass covering so that this can comply with the responsible AI use laws such as the EU AI act.
it claims to be open source because the weights are freely available, and whether or not that conforms to the definition some consortium of folks cooked up for what "Open Source" means, anyone who can put aside their feelings of ire for Meta for 2 seconds can tell that making the weights available is meaningfully different than keeping them locked up, and exposes most of the value to the public to use, for free.
Gemma is called "open weights." This terminology is correct. You don't get to redefine terms just because you happened to do something that other people like. https://github.com/google-deepmind/gemma
at least using llama is less dumb than using the Gemini, ClosedAI, Claude, or xAI apis, since it’s not effectively a one way wiretap like those lemming tier options
Are you a dedicated member of the "LLMs have no legitimate uses" camp?
In case it wasn't obvious, my blog post is meant to be equally about the "Meta/open source/EU AI act" thing and the "look at what you can do with these new long context models that were released in the last few weeks" thing.
As so often is the case with my LLM projects this one wasn't a case of choosing between "read the EU AI act or pipe it through an LLM" - it was a choice between "pipe the EU AI act through an LLM or lose interest in this mild spike of curiosity and go and do something else instead."
> You would rather read all of <law> than pipe it into an LLM to answer a few questions about it?
yes, yes I would.
> How long would it take you to read that thing? I'm a reasonably fast reader and it would take me hours.
For most of the "few questions" you can skim most of it
> Are you a dedicated member of the "LLMs have no legitimate uses" camp?
I'm in the camp of "do not offshore your thinking process to a non-deterministic black box whose whole mode operandi is to always generate plausible-looking answer and then profusely apologize if it was caught to produce invalid output".
---
Also, "five hours to read an important legislation written in a surprisingly clear language is too long and nigh impossible" is the premier reason about so much bullshit disinformation about EU AI Act, DMA, GDPR and plethora of other, less important regulations
I outsourced the "skim most of it" bit to the model. I used an LLM to jump to the bits that mattered, then I confirmed those bits by reading them myself in the original document (and thinking about them).
I fed in the entire act with the articles and the recitals. The full response from Gemini included information from the articles, but I didn't quote that directly in my blog post. Here's that full response: https://gist.github.com/simonw/f2e341a2e8ea9ca75c6426fa85bc2...
Relevant section:
> Article 53(2) provides an exception from the obligations for providers of general-purpose AI models regarding technical documentation (Art 53(1)(a)) and providing information to downstream providers (Art 53(1)(b)) if the models are released under a free and open-source license and their parameters (including weights), information on model architecture, and information on model usage are made publicly available. This exemption does not apply to general-purpose AI models with systemic risks.
I don't agree. For a start, this blogpost puts way too much weight on the recitals.
EU law is all in the Articles. AFAICT the recitals can be used in determining purpose, when courts have to interpret ambiguity. But so can other facts!
Sounds to me like you know a heck of a lot more about EU law than I do! Can you explain more?
Thanks. My rule of thumb is simply to skip straight to the Articles, when interpreting EU law.
Recitals maybe seem weird because we don't have them in UK legislation. Maybe the nearest thing is Explanatory Notes? If you want to know what a law says, read the law itself. But if the law makes no sense (either as a lay person, or because the law is badly drafted even for legal folks) then the Explanatory Notes may offer some insight.
I see it as the complete opposite. It's not open source because of ass covering.
The biggest barrier to the licence passing open source requirements is point 0 in the above linked evaluation. Namely that it has a whole lot of "you may not use this for [nefarious purposes...]" type of statements. That seems like ass covering so that this can comply with the responsible AI use laws such as the EU AI act.
it claims to be open source because the weights are freely available, and whether or not that conforms to the definition some consortium of folks cooked up for what "Open Source" means, anyone who can put aside their feelings of ire for Meta for 2 seconds can tell that making the weights available is meaningfully different than keeping them locked up, and exposes most of the value to the public to use, for free.
Gemma is called "open weights." This terminology is correct. You don't get to redefine terms just because you happened to do something that other people like. https://github.com/google-deepmind/gemma
So they should call them open weights.
at least using llama is less dumb than using the Gemini, ClosedAI, Claude, or xAI apis, since it’s not effectively a one way wiretap like those lemming tier options
> That should fit nicely into Gemini 2.5 Flash (or GPT-4.1 or Gemini 2.5 Pro).
Christ almighty. The act is neither long enough nor hard enough to read and understand yourself
Seriously? You would rather read all of https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:... than pipe it into an LLM to answer a few questions about it?
How long would it take you to read that thing? I'm a reasonably fast reader and it would take me hours.
(https://wordcount.com/character-counter estimates 5 hours and 46 minutes)
Are you a dedicated member of the "LLMs have no legitimate uses" camp?
In case it wasn't obvious, my blog post is meant to be equally about the "Meta/open source/EU AI act" thing and the "look at what you can do with these new long context models that were released in the last few weeks" thing.
As so often is the case with my LLM projects this one wasn't a case of choosing between "read the EU AI act or pipe it through an LLM" - it was a choice between "pipe the EU AI act through an LLM or lose interest in this mild spike of curiosity and go and do something else instead."
> You would rather read all of <law> than pipe it into an LLM to answer a few questions about it?
yes, yes I would.
> How long would it take you to read that thing? I'm a reasonably fast reader and it would take me hours.
For most of the "few questions" you can skim most of it
> Are you a dedicated member of the "LLMs have no legitimate uses" camp?
I'm in the camp of "do not offshore your thinking process to a non-deterministic black box whose whole mode operandi is to always generate plausible-looking answer and then profusely apologize if it was caught to produce invalid output".
---
Also, "five hours to read an important legislation written in a surprisingly clear language is too long and nigh impossible" is the premier reason about so much bullshit disinformation about EU AI Act, DMA, GDPR and plethora of other, less important regulations
I outsourced the "skim most of it" bit to the model. I used an LLM to jump to the bits that mattered, then I confirmed those bits by reading them myself in the original document (and thinking about them).
LLMs are a tool.
Yup, LLMs took you to recitals, not to the articles themselves. This is definitely better than invalid info, I'll grant you that.
I fed in the entire act with the articles and the recitals. The full response from Gemini included information from the articles, but I didn't quote that directly in my blog post. Here's that full response: https://gist.github.com/simonw/f2e341a2e8ea9ca75c6426fa85bc2...
Relevant section:
> Article 53(2) provides an exception from the obligations for providers of general-purpose AI models regarding technical documentation (Art 53(1)(a)) and providing information to downstream providers (Art 53(1)(b)) if the models are released under a free and open-source license and their parameters (including weights), information on model architecture, and information on model usage are made publicly available. This exemption does not apply to general-purpose AI models with systemic risks.