Commentary on X
@Lunexalith @teortaxesTex lol you are joking right? I see GLM reaching fable before Qwen
5:10 PM · Jul 15, 2026@teortaxesTex Peter did it wrong, don't listen to this video.
5:28 PM · Jul 15, 2026@AtaeiMe unrealistic but here's to hoping
4:53 PM · Jul 15, 2026pretty good video. reaffirms my (updated upwards) expectations of 55ish AA for K3. Around Opus 4.8 or GPT 5.5. my main takeaway though isn't "wow Kimi is behind the true frontier" (I knew it is). It's "wow even Fable and Sol make so much SLOP". Peter glosses over it but Fable adds tons of pseudo-interactivity, half-baked gizmos not really connected to anything… Sol routinely displays terrible decision taste… Like Nano Banana, look closer and it starts to peel. It's a bit sobering. And most of the gap seems to be on the level of effort level and prompt interpretation. Kimi doesn't so much fail at stuff (it does fail, but rarely) as it makes less interesting and detailed stuff, doesn't go the extra mile. For human-in-the-loop agentic coding with objective tasks, K3 should come unreasonably close. I can even buy that it's strictly worse at zero-shot generation of complex environments in Three.JS, "make no mistakes, animate dinosaurs like they do rawr". But from this evidence, I can equally buy that it'd be at parity in writing new game engines, and that's a lot more exciting for me. Of course, that part depends on multi-turn work, which is not covered adequately in such tests. Lisan will probably say I'm coping, because the tone of the video is "sorry kid, it's not REALLY comparable… heh, but for the money, maybe…". I'm actually pleasantly surprised. This is way less of a gap than I imagined at first. Not sure if it's shortened from 6-8 months, seeing as Mythos-Preview exists and according to my sources was around Opus 4.8 in general tasks by February. But as I've said before, every Chinese model that merely *maintains* the gap in the current regime is a heroic feat. "Necessity is the mother of invention" is peasant deepity! Ant and OpenAI aren't sitting on their thumbs! They are racing, pouring vast capital into their post-training, they have more compute than God, vastly larger teams packed with experienced and ambitious luminaries, *years* of user interaction trajectories and datacenters full of synthetic stockpiles, polished recipes… and every Chinese open model and paper from their competitors across the Pacific. It's simply absurd that we're not yet seeing the gap widen dramatically (if at all), and a humbling testament to the ferocious execution by top Chinese startups, and the strength of the Mainland talent pipeline. These are my impressions. I've yet to test K3 myself. Knowing the team, I expect it'll hold up well, but not clear yet.
4:10 PM · Jul 15, 2026Combined views
46.8K
5 posts, first seen 2h ago