Wow.

Possible ways to kept Meta ad records honest and transparent:

- CCing archive.org

- Store on an append-only system with hashing, hello blockchain use-case ha ;D.. IPFS or even GitHub should do, no crypto payments required.

- Third-party government bodies could require copies.

Currently it says they have released metadata and album art. Is archiving and sharing the textual track metadata alone (no images, no audio) legal in the US, or Europe? By what basis is it legal or illegal?

They have this page for reporting: https://github.com/hackclub/hcb/blob/main/SECURITY.md

Old man yells at cloud: that's not what "literal killer" means.

"literal" and "literally" are both now commonly used oxymoronically as a figure of speech warning of the use of metaphor and/or exaggeration.

It literally drives me nuts

I had the same reaction as them many months ago, the Google Cloud and Vertex AI stuff namespacing is a too messy. The different paths people might take to learning and trying to use the good new models needs properly mapping out and fixing so that the UX makes sense and actually works as they expect.

I'd lower the temperature and try DSPy Refine loop on it (or similar). Using audio encoder-decoder models and segmentation are good things to try too. A length mismatch would be bad. DSPy has optimisers. It could probably optimise well with length match heuristic, there is probably a good Shannon entropy rule.

Simon, sorry I didn't get around to answering your question on post-t5 encoder-decoders from the Markdown Lethal Trifecta prompt injection post. (https://news.ycombinator.com/item?id=45724941)

Since the plain decoder models stole the show, Google DeepMind demonstrated a way to adapt LLMs,adding a T5 encoder to an existing normal Gemma model to get the benefits of the more grounded text-to-text tasks WITHOUT instruction tuning (and the increased risk of prompt injection). They also have a few different kinds they shared on HuggingFace. I didn't get around to fine-tuning the weights of one for summarisation yet but it could well be a good way for more reliable summarisation. I did try out some models for inference though and made a Gist here, which is useful since I found the HF default code example a bit broken:

https://gist.github.com/lukestanley/ee89758ea315b68fd66ba52c...

Google's minisite: https://deepmind.google/models/gemma/t5gemma/ Paper: https://arxiv.org/abs/2504.06225

Here is one such model that didn't hallucinate and actually did summarise on HF: https://huggingface.co/google/t5gemma-l-l-prefixlm

I am very picky, hard to place, but from a quick look at the README, I'd say the API interface on display seemed like the right level of abstraction for having to deal with the messy reality.

Since you're asking for feedback:

- perhaps some of the document type specific dependencies by optional?

- could there be LESS config surface?

- I noticed GitHub CI action has a cross.

It's good to add how to use with Astral "uv" these days, especially anything that might pull in PyTorch dependency hell, which they have mostly solved if used correctly!

Nice work!

Love this kind of feedback, thank you. You nailed it on optional deps and config sprawl; I’m trimming both. CI cross is just coverage noise, and I’ll add uv setup notes it really cleans up the PyTorch mess. Glad the API felt right — that was the hardest part to get “just enough abstraction” right.

Direct quotes: "It pays to be brave." ... "we pledge to distribute to tippers $200 million out of every $1 billion we collect." ... "This is TruthWave. Welcome to the platform and community for those who bring unethical corporations to justice."

The Lethal Trifecta strikes again! Mermaid seems like a bit of a side issue, presumably there are lots of ways data might leak out. It could have just been a normal link. They should probably look further into the underlying issue: unrelated instruction following.

Thanks for the archive link and the very useful term BTW! I also got 503 when trying to visit.

I think they're doing this the right way. You can't fix unrelated instruction following with current generation LLMs, so given that the only leg you can remove from the trifecta is mechanisms for exfiltrating the data.

The first AI lab to solve unrelated instruction following is going to have SUCH a huge impact.

Not even humans can do it perfectly (hence social engineering)

Humans would not do what was reported here!

No, they are doing this the only possible way that doesn't massively restrict it being useful at all. That doesn't make it the right way.

A fundamental vulnerability to prompt injection means pretty much any output can be dangerous, and they have to expose it to largely untrusted input to be useful at all.

Even limiting output to ASCII text only is probably not entirely safe.

The right way at this point would be to not use AI.

Summarisation models that do not follow instructions already exist! Fixing exfiltration is good, low hanging fruit. But for a summarisation task, whole classes of typical instruction following behaviour are totally off target!

Which models do you recommend for that these days?

Has the field moved on much since the pre-LLM-era models like T5?