llamasushi

Just updated, thanks

The burying of the lede here is insane. $5/$25 per MTok is a 3x price drop from Opus 4. At that price point, Opus stops being "the model you use for important things" and becomes actually viable for production workloads.

Also notable: they're claiming SOTA prompt injection resistance. The industry has largely given up on solving this problem through training alone, so if the numbers in the system card hold up under adversarial testing, that's legitimately significant for anyone deploying agents with tool access.

The "most aligned model" framing is doing a lot of heavy lifting though. Would love to see third-party red team results.

This is amazing. Thank you for this.

That's an optimistic take, but equally valid is the take where delusions provide impetus for some pretty nasty behaviour - eg see the crusades

"Warmer and more conversational" - they're basically admitting GPT-5 was too robotic. The real tell here is splitting into Instant vs Thinking models explicitly. They've given up on the unified model dream and are now routing queries like everyone else (Anthropic's been doing this, Google's Gemini too).

Calling it "GPT-5.1 Thinking" instead of o3-mini or whatever is interesting branding. They're trying to make reasoning models feel less like a separate product line and more like a mode. Smart move if they can actually make the router intelligent enough to know when to use it without explicit prompting.

Still waiting for them to fix the real issue: the model's pathological need to apologize for everything and hedge every statement lol.

LeCun, who's been saying LLMs are a dead end for years, is finally putting his money where his mouth is. Watch for LeCun to raise an absolutely massive VC round.

Lol, this reminds me of a funny story. I had a lawyer whose name was Jim Halpert. Turns out he was the very Jim who was inspired his namesake on the office. Asked him about it once. His reply? "Hey, it's been great for getting clients." =)

He was also very much like Jim on the show. Fun times.

Am I the only one who thought this was referring to how people felt about the general zeitgeist? Like, how Romans viewed everyone outside Rome as barbarian, etc. Not in the literal sense like, mirrors. Nice HN switcheroo.

Yeah, this was my first startup and I thought I could multi-task lol. Probably did it for the wrong reason (eg: "checking a box - went hitchhiking in Alaska") than anything else. Spent most of the trip worrying about other) startup shit, didn't enjoy the scenery nearly as much as I should have. Still regret it haha!

So it's not really hallucinating - it correctly represents "seahorse emoji" internally, but that concept has no corresponding token. lm_head just picks the closest thing and the model doesn't realize until too late.

Explains why RL helps. Base models never see their own outputs so they can't learn "this concept exists but I can't actually say it."