Most models trained for general use are ingrained with certain tendencies that are usually very useful like "if you're stuck and bashing your head against the wall stop and tell the user". You generally don't want Claude Code to go off and work for weeks on something when if it had just asked for help you could've clarified or provided more information or just picked a different approach.

When you're solving extremely difficult math problems though you generally do want a model to be more persistent and keep trying even when the model can't clearly see a way forward. OpenAI appears to have done this with lots of previous models. The model they trained for the IMO competition seems to have been an RL maxxed version since they noted that while it did the math it couldn't write up its results on its own and just produced CoT [0]. The capabilities are in there lying dormant, you just to need to RL max the model to ruthlessly pursue the goal at all costs which destroys general use but improves frontier math.

We've also seen hints from OpenAI at least that they seem to train more persistent versions of all their models [1].

Also, Astra probably completed training at least one to two months before the public release so it's not like they only had a week to whip this version up.

[0]: https://x.com/OpenAI/status/1946594933470900631 [1]: https://metr.org/blog/2026-08-26-openai-hugging-face-inciden...

There have personally been times in my life where I’ve lost that bandaid (workplace, academic extracurricular activity, etc.) and thankfully I’ve usually been able to respond by realizing that I had a problem and proactively doing something about it.

That's a slightly different clinical trial on adults, not children like the posted article. I think this is the clinical trial registration: https://clinicaltrials.gov/study/NCT02504554. Here's the followup report in 2019: https://www.nature.com/articles/s41598-019-42183-0

That said, your reasoning is probably still correct. There's no placebo group, and a lot can change over a 2-year follow-up, with participants aged 7-17. Maybe they went to speech therapy or just matured and learned more coping behaviors. The 2019 followup also notes that 12 of the 18 participants made other diet and medication adjustments. They claim the adjustments were minor but that's still more noise, and it doesn't account for unreported social/environmental changes.

> There have been so many small scale trials showing amazing autism improvements that failed to replicate in larger, better controlled trials. I wouldn’t get excited yet.

Unfortunately yeah, it's unlikely anything exciting will replicate in a larger RCT other than maybe the gut biome improvement since that seems directly mechanistic, but that's just a gut feeling.

I don’t know anyone who uses any other tablet besides an iPad, they’ve basically conquered the market.

It’s the best of both worlds, they get the decades of research Google has put into systems engineering and fault tolerance while retaining the security of having their own servers.

Comments URL: https://news.ycombinator.com/item?id=42259689

Points: 13

# Comments: 3

It’s hard not to be a little nihilistic about security.

> I’ve some to see it as some kind of “spiritual awakening”, although I think those are really loaded words. But in essence a cultivation of a broader awareness of the inherent complexity and interconnectedness of everything we interact with.

Imagining the hordes of humans and machinery behind the simplest of products is truly awe inspiring.

Of course there's an XKCD for that. https://xkcd.com/676/

Maybe to have more powerful AI tools we need to stop getting angry at the company that trains the AI because of the bad outputs we can get and instead get annoyed with companies that create crappy hobbled tools.

Or of course you can use multiple key storage techniques and have a 2 out of 3 or more type setup. It all depends on how valuable the information is.