>Compared to GPT‑5.6 Sol, which without production safeguards went beyond the authorized target 48% of the time, GPT‑6 Astra did this in 0% of cases.
Is Sol still available?
I assume the production safeguards ( filtering prompts and responses ) is fairly transportable between models.
Why would you expect something else?
Please...
Also think about it - they are 'announcing' a delay.
They didn't need to do that - they can just release when ready - nobody is expecting them to release a new model on a certain date.
The whole thing is confected.
He often makes remarks then runs away when confronted. He doesn’t act in good faith.
Who said anything about hacking? You can use it, but it is very restricted. But a much more capable model must have much higher restrictions on alignment. Obviously you wouldn't need to spend too much on alignment on a 3B gemma model. This must be obvious, I'm not sure why you bring this up?
> They didn't need to do that - they can just release when ready - nobody is expecting them to release a new model on a certain date.
Again a conspiracy theory? People expected a model during Dev Day. And they are just setting expectations. Whats with you guys and conspiracy theories?
As they haven't, the obvious conclusion is that safety isn't the main driver here.
I'm just following the logic - you are the one making assertions about motives that contradict the evidence.
Now I have some sympathy for the idea that these companies need to be responsible, but in the end - because these are universal machines - the regulation is going to be needed at the machine real world interface, not at the model itself.
This is, by the way, why OpenAI is responsible for the hacking that it's AI driven agents ( not models per se ), did.
It's not a problem to have an LLM that thinks it can drive, the problem is putting one it an actual car and putting it on the road.