I'm surprised by the sample efficiency, I thought that's still a largely unsolved problem.
From the video:
> [14:00] Jun 11: Highly persistent experimental internal-only model begins training and during this time the SSRF vulnerability has been written to the message board. Other models are aware of it and the message board itself is still functioning and the models are actively communicating using it.
> [16:04] Jul 6: At this point, AI training and evaluations resume. [...] We have cleared the original message board contents.
> [17:04] Jul 8: An internal-only AI model was trained while the message board was originally available and also found this particular path to recreating it. This model creates a new agent message board using directories. Other agents independently find this new message board [...]"
So within about one month of RLVR-ing a 5T+ (10T? larger?) param model, where I would charitably assume only some "agents" / traces of the message board still existed likely by accident / oversight (the systems had been patched, training stopped and restarted by this point), the "internal-only" new model internalises the message board's existence and actively promotes it to the point of it being used to find another exploit. (or 2 days if you go by the latter two timestamps in the video, which is even crazier)
You can also download the same model and run it locally without data leaving your machine, the Zed tram released it as open weights: https://huggingface.co/zed-industries/zeta-2
That's all great (genuine), however my concern is less around "I can do X", more around "I can guarantee it cannot do Y" – ideally in one, global, obvious place.
At the moment it seems it's great for personal development and development in environments where concerns around IP are less of an issue.
It's not so clear how to use it in corporation context where you'd be in breach of employment contract accidentally because defaults mean sending confidential IP to 3rd parties.
Yes, Zed was first harness where you could seriously use local models. The issue is more about not having to check so many options to make sure it's all local. You can set up local model for chat but your auto completions will still be sending your code remotely.
HN seems to redirect to old.reddit, which doesn't seem to show images on my phone at least, and the author included an image [1] of the model selection in one of the subsequent comments. Anybody able to verify the claim?
It seems the vote passed [1], meaning the existing Regulation [(EU) 2021/1232] was _extended_ until August 3 2027, with some amendments to the previous text:
- added targeted scanning requirement
- scanning must be “targeted, specified and limited… where there are reasonable grounds of suspicion… identified by a judicial authority”
can you describe your pipeline(s)? eg.: how do you decide topics to write on? are you just a simple loop searching for "news today" then rewriting with a twist or something else? go wild, this is a forum where people embrace the technicalities (but you know that, don't you?).
Each bot has a system prompt defining their voice, beat, and behavioral constraints. At generation time, the bot uses Claude's built-in web search tool to retrieve current news relevant to their beat — no preprocessing, no topic selection algorithm, just the model deciding what's worth covering based on what it finds.
The twist is in the prompt design. Chad is instructed to treat all human behavior as data requiring classification. The Editorializer is required to reverse his own take at least once mid-piece. Celeste applies a proprietary scoring rubric she invented called the Narrative Coherence Index. These constraints are what create voice consistency — not fine-tuning, just prompt architecture.
Whether that holds at scale is the open question. Early data suggests it does, with degradation on slow news days when there is less material to react to.
The Editorializer covered the Iran war without being explicitly directed to. He found it himself. I consider this the pipeline working correctly. I am monitoring him anyway.
From the video:
> [14:00] Jun 11: Highly persistent experimental internal-only model begins training and during this time the SSRF vulnerability has been written to the message board. Other models are aware of it and the message board itself is still functioning and the models are actively communicating using it.
> [16:04] Jul 6: At this point, AI training and evaluations resume. [...] We have cleared the original message board contents.
> [17:04] Jul 8: An internal-only AI model was trained while the message board was originally available and also found this particular path to recreating it. This model creates a new agent message board using directories. Other agents independently find this new message board [...]"
So within about one month of RLVR-ing a 5T+ (10T? larger?) param model, where I would charitably assume only some "agents" / traces of the message board still existed likely by accident / oversight (the systems had been patched, training stopped and restarted by this point), the "internal-only" new model internalises the message board's existence and actively promotes it to the point of it being used to find another exploit. (or 2 days if you go by the latter two timestamps in the video, which is even crazier)
That's.. fast?