Post #1498885

2026-04-11 23:00 UTC

Wrote up some thoughts about Mythos, why I think it&#39;s actually not publicly available, and what it says about LLM scaling https://www.colincornaby.me/2026/04/anthropics-mythos-implies-llm-scaling-has-hit-a-wall/?utm_source=mastodon&amp;utm_medium=jetpack_social

Replies (4)

@bwebster@mastodon.social 2026-04-11 23:06

@colincornaby Colin, I'd like to buy your rock.

Open ##1737814
@stevex@mastodon.social 2026-04-12 00:01

@colincornaby I've seen the security researchers trying to write this off as no big deal, that much smaller models could have done this. Question I have is, why haven't they? Yeah, security is a good way to charge more for an LLM that's difficult to scale. There's probably some of that. But I do believe it's also genuinely more capable of finding these vulnerabilities. Both things can be true.

Open ##1737815
@j_s_j@mastodon.social 2026-04-12 00:07

@colincornaby Great points, but consider an alternative: Huge models can be used to distill smaller, focused models. That’s the nut they’re trying to crack. That’s why Intel and AMD and Apple are rushing to add neural cores. It’s completely untenable economically if the only way they can run useful models is in the datacenter.

Open ##1737817
@rickfillion@mastodon.social 2026-04-12 00:18

@colincornaby solid post. Think there might be a typo though. You say “2 trillion dollar” when I think you mean “2 trillion parameter.”

Open ##1737820