fxwin

Earlier this week, we detected and responded to an intrusion into part of our production infrastructure. This one was different from anything we had handled before in one important way: it was driven, end to end, by an autonomous AI agent system - and we detected and dissected it largely with AI of our own. (https://huggingface.co/blog/security-incident-july-2026)

We are living in crazy times

im not sure which part of my summary you take issue with then? the reason for study 1b is not relevant to answer the question asked in the comment i replied to

You can read the study here: https://osf.io/preprints/psyarxiv/5y6m4_v1

TLDR: They studied both cases (Access to a LLM/Chat interface which gives a wrong answer when asked, and access to pregenerated (wrong) answers). Both experiments yielded similar results.

definitely the latter, it is even referenced in the foreword:

Its goal is not to be exhaustive, but rather minimalist and easy to read. For this reason, it follows the format of The Little Book of Deep Learning [Fleuret 2023]. Its tone, however, is closer to that of a blog post, as the book is built around a single narrative thread. Its structure broadly follows that of Sutton and Barto’s Reinforcement Learning: An Introduction [Sutton et al. 2018], which remains the canonical reference on the subject.

looks like they have since fixed the dataset, hope they release an updated report (either without the leaked evals or a new checkpoint/weights)

https://huggingface.co/datasets/AIML-TUDA/QA-base/commit/2f9...

I think the real challenge is to figure out wtf you're actually supposed to do from all this flowery over the top claudespeak

The reality is people don't always care if a human poured their heart and soul into something.

That's fine, but I don't think the author would suggest writing e.g. library documentation by hand. It's clearly advice for the creator side of the problem of low signal-to-noise ratio in the digital space and how to stand out/signal, rather than a general rule

That was a fun read! I caught myself almost skimming the first part until i got to the mirrored paragraph, and slowed down significantly after that to read more deliberately.

I'm not sure how much actual advice one can take from this essay though beyond "use personal commitment (e.g. time or presence) to signal importance/care" and "go offline" (aka touch grass)

For four months, no frontier model beat Claude Opus in our production evals

What do you think that is referring to?

The Pi 3B+ seems just enough to render some simple animations.

I think a Pi 4 might be a good sweet spot between processing power and price,

I know this isn't exactly a serious product and more of a gadget/gimmick but man we are off by a factor of like 3-4 here :D

I would expect they have production based datasets they evaluate new models against.

Tried exploring a small project i built with CC, but i don't see anything in the tree/terrain view (The edits/reads/writes do show up in the timeline). The project itself doesn't exist on my drive anymore, is that a requirement?

fwiw o1, o3 and 4.1 also give the correct answer (without web search)

I'm not sure which chatbots you used, but OpenAI's o3, o1 and 4.1 get it right first try (used in the Playground without web search or any other tools).

Sure, but that's the nature of language (which is also why i put "understand" in quotation marks. I usually follow it up with "whatever that means" lol) . I think in this case, it carries with it implicit properties (abstract semantic representation) that i think models possess, which a decoder doesn't.

I think "(intelligent) language understander" is an apt term. It contains within it the fact that these models are mainly trained on text, and "understand" it beyond a simple token-by-token level (i.e. their latent space maps to more and more complex concepts).

It also separates them from "world understanders" since any understanding they might have about the world comes from text (or images if we include multimodal models). They do not gather experience, memories or other "qualia" that many people (me included) would probably include in a definition of human experience/intelligence.

(fwiw i think artificial intelligence is a good, broad term, but it is both too broad to describe the current sota, and too loaded nowadays to be using in nuanced discussions)

Have they retracted it? My understanding was simply that they released results with more recent data, not that this study itself was flawed (and their website doesn't mention a retraction either)

And i would consider this as an agent doing things under human instruction

I'm not sure which part of my comment comes across as fence sitting to you, but to clarify:

- I think there are good and useful ways to use AI in art creation

- I also think that AI (especially end to end creation of full songs) is a massive problem for both content moderation and business model aspects of this type of platform

- I don't think a blanket ban on AI in <Field X> is a good idea (even a bad idea in almost all cases)

You can disagree with these three points and that's fine, but it's not fence sitting when someones position isnt 100% one way or the other

I work with coding agents every day, I don't think they have ever started working on a project without me telling them to

Laws and rules exist to serve humans, not machines.

Machines don't go out on their own to create and upload music, they do so under human instruction, so their output should be policed the same way we police other machine generated output directed by humans.

The issue i have with it will depend heavily on implementation, i can see cases where songs that i would consider "produced and written" by people don't qualify for royalties under Tidal's guidelines. (I intentionally left out the "performed" part, since digital music production is way past the point where this was an easy and/or meaningful distinction)

Tidal will accept AI-generated music.

Tidal will hold AI-generated music to a higher standard of content integrity. We will not tolerate AI-generated music that exploits an individual’s or group’s music, name or likeness, deceives listeners, or diminishes the quality of our service.

I think this is a very reasonable approach, and probably also the best way to treat AI-powered copyright infringement as a whole. Just like we don't penalize artists for consuming content unless they produce actually infringing content, we should set the same focus for AI systems.

Starting today, AI-generated music will not be monetizable. We are only in the beginning of the era of AI-generated music.

Don't really agree that this follows from the stated principle here ("... ensuring royalties go to original works produced, written and performed by people"), but will definitely help with spam etc.

fwiw aleph alpha have been around since 2019

Those are specs that belong on a laptop or a lowend console offering like the Series S.

Unfortunately, valve (and we consumers) have to recalibrate our understanding of which prices qualify as "insulting".

Me, a german, looking at my CAX11: this doesn't seem too bad

Americans:

I think it helps his credibility that he has been working with and speaking positively about AI assisted mathematics (especially for formalizing proofs) for over a year now . I'm sure he isn't unbiased, but as far as spokespeople in the AI space are concerned I'd count him among the less biased ones.

The advantage of frameworks isn't that they make it easier to write the actual agent, it's tooling + observability + ... Even Langchain, for all the (deserved) criticism it gets made this very clear very early: It might be easy/easier to write your own chatbot from the ground up, but what happens if you have to add observability/tracing? Being able to just add one environment variable and instantly have a UI where i can nicely go through all of my traces with basically 0 additional effort is something a hand rolled solution just can't really compete with

here: https://llmstxt.org/ and obviously it doesn't automatically produce markdown, it's something the website needs to provide (e.g. https://pydantic.dev/llms.txt)

Well similar to how turing machines are a sufficient theoretical model to make all kinds of arguments about runtime complexity of classical computers without relying on their actual physical implementation, we have theoretical models for the way we are approaching quantum computation that do the same thing (Namely the quantum circuit model)