Felicitas Pojtinger 🌅Damn those Mythos benchmarks seem very promising
Wild that they don't seem to be making it GA, makes me suspect it's probably actually not as good as they say
Qwen 3.6 is essentially the same as Opus 4.6 now so I guess we'll see how the new generation stacks up?
Steven Deobald@pojntfx have you actually seen qwen perform this well? or are you basing that comment on benchmarks?
i think the mythos benchmarks only have to be "some amount better" at finding 0days than the current public models to justify them waiting on ga... quite a few maintainers are already swamped.
@deobald Yup, I used Qwen 3.6 with Nanobot via OpenRouter, Alibaba was providing it for free for testing until yesterday. Switched to GLM 5.1 earlier - same thing, beats Opus. GLM's weights are even MIT-licensed
@deobald And yeah re:Mythos I'll believe it when I see it, but current-gen models except free is already a massive value IMHO. Sonnet etc. is still very useful despite the other models existing
@deobald I'm pretty happy about mostly working with higher-level, memory-safe languages
@deobald If you'e like to try for yourself I've documented it here: https://gist.github.com/pojntfx/5916ceb7ec35eb010010400447e9c034
@pojntfx are you using nanobot for hacking or were you just pointing me to the provider section?
@pojntfx nod. it does have me thinking hard about other forms of baked-in safety. i'll admit this is the first point in my career where i've ever taken elixir seriously.
(well, ok, not really... @abnv ran a team at nilenso that did some amazing work with it for an quiz app that ran in parallel to a tv show. but i've never previously been tempted to learn it.)
James Just James