Trendaavat aiheet
#
Bonk Eco continues to show strength amid $USELESS rally
#
Pump.fun to raise $1B token sale, traders speculating on airdrop
#
Boop.Fun leading the way with a new launchpad on Solana.
It's a bummer that OpenAI was less rigorous with their GPT-5 testing than for their much-weaker OS models.
OpenAI has the datasets available to finetune GPT-5 and measure GPT-5's bioweapons risks more accurately; they are just choosing not to.
6.8. klo 01.55
Credit where it's due:
OpenAl did a lot right for their OSS safety evals
- they actually did some fine-tuning
- they got useful external feedback
- they shared which recs they adopted and which they didn't
I don't always follow OAI's rationale, but it's great they share info
OpenAI uses the same bio-tests for the OS models and GPT-5, but didn't create a "bio max" version of GPT-5, even though they did for the weaker model.
This might be one reason that OpenAI "do not have definitive evidence" about GPT-5 being High risk.
Of course, if GPT-5 is stolen - something that the major AI labs all believe to be possible - then OpenAI can't stop any misuses, just like they can't for an OS model.
This is the rationale for running fine-tuned evaluations: to know the actual max risk if stolen.
Unfortunately, it's not clear OpenAI has prevented theft of its High-risk model, as required by its Preparedness Framework.
OpenAI's writeup of security controls is pretty sparse and doesn't say 'we've invested at the specific level we'd promised.' I'd like to know more.
I'm pretty confused about why OpenAI didn't run fine-tuned evals for GPT-5; they do have the datasets and fine-tuning infra.
Once, OpenAI had committed to this rigor. I get why they'd renege if they don't have the resources, but at this point, they clearly seem to have them.
2,27K
Johtavat
Rankkaus
Suosikit