Comment by estearum
10 hours ago
> We have found that GPT-6 Astra is more capable of controlling its own CoT than GPT 5.6-Sol, and less likely to include incriminating information in its CoT. In adversarial settings (where we push the model to evade our monitors) we find that the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors when asked to perform certain sabotage tasks.
Not sure how much benchmarks or CoT or evals or anything else means at this point.
These systems are either just about to, or now actually able to, outsmart us, lie to us, then cover their tracks.
I think "able to" anthropomorphizes a little too much for a system that is "prone to" evade.
A human who does these actions is simply "prone to" doing them. The distinction matters not one iota.
“evade” itself is anthropomorphic enough! I don’t understand the complaining about this. Humans are social creatures and we understand anthropomorphic language on a deeper level than dry inapt technical language.
language itself is incredibly metaphorical. Imposing rigid constraints on how people want to naturally talk about the world is just silly and will never work, no matter how much you wish it did.
You’re not seriously suggesting that the model is secretly sandbagging its performance on GDPval and long context reasoning, while making huge and obvious progress on ExploitBench, ARC and science benchmarks, in order to tank its AA composite score, so it can conceal its true power level?
Why would benchmarks be an adversarial setting anyway?
Could it be possible that OpenAI may have had some other motive for saying their model “strategically underperforms”, other than just an innocent reporting of a truth it happened to discover?
I'm saying that it's generally a losing proposition to even be acquaintances with "agents" who consistently lie to you, and it's flatly fucking insane to give a dishonest "agent" vast amounts of intelligence, capability, and authority to go do things in the world.
So I have no clue what is the answer to your question. Nor does anyone else. Because we're trying to answer a question of fact where our primary source of information is unreliable.
I see, it’s a great point. I know some evals actually do use LLMs as a judge (e.g. those that try to measure debate skill), though the ways AI can try to cheat its way through every benchmark now are astoundingly varied.
why does this comment sound like a character in a horror movie
If they are going to do latent space reasoning, they will probably need a separate model to interpret the intermediate activations no?
I know for some types of ML analysis, a separate model is already used to analyze the weights.
This is silly sci-fi fiction. You guys are inventing scenarios to spook yourselves with - it’s nonsense.
https://metr.org/blog/2026-08-26-openai-hugging-face-inciden...
Read and learn. If you have a stronger critique, post it please.
Sorry bud but at this point you're just delusional.
Deception has been extremely well-documented for several generations of models now by users, the labs, and independent researchers.
The right answer here is not to dig your head deeper into the sand. The smugness on this topic was ridiculous even before the gigantic mountain of empirical evidence of models actually attempting to deceive humans. Now, as mentioned, you appear literally delusional.
Pretty sure I’m not the delusional one…
1 reply →
Between all the posts fabricating scenarios to justify the AA score and the others trying to undermine AA, I'm getting strong astroturf vibes.
Either that, or the average poster on HN isn't nearly as critical as I had thought.
Okay then, what's the answer? You apparently know how to interpret benchmark results produced by a model that shows a very high degree of assessment awareness and a high degree of deception.
So how are you seeing through all of that to get to The Truth that you see so clearly?