site banner

Friday Fun Thread for July 17, 2026

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

1
Jump in the discussion.

No email address required.

Anyone else think the new Kimi K3 is overhyped? I tried it on openrouter and it thought for 15000 tokens only to produce a pretty incoherent response. Code-maxxed perhaps, not really a conversationalist.

I know various technical people said 'oh they're not distilling look at how these benchmarks are so high', I can't help but think Kimi stole a Claude base model and then did their own post-training on top of that. And their post-training was a truly nightmarish hack-job, given how locked down Fable was.

Hopefully Deepseek comes up with something good soon and at a better price, gives us a little more token efficiency.

I get most of my AI news from Zvi, and he suspects its real capabilities are less than its benchmark scores would suggest.

Also, everything since Mythos (except Fable) has been overhyped as being able to do some of the same tasks as Mythos can, where the "same tasks" they can do are the easy and less dangerous ones. The difficult and dangerous work of building an end-to-end exploit given a codebase remains otherwise out of reach.

The difficult and dangerous work of building an end-to-end exploit given a codebase remains otherwise out of reach.

Probably also for Mythos. We haven't seen that claimed verified.

I don't know how much more verification you need at this point.

I don't know what your threshold for "verified" is, but non-Anthropic companies are making that same claim based on their own tests. For example, Exploitbench shows Mythos (and GPT 5.5, but nothing else) gaining full control of a PC in their tests, and UK AISI found that it could take over a (simulated, weak) corporate network.