back
11 comments
Watching K3’s thought traces has been fascinating. I have a podcast transcription pipeline that uses an LLM pass to increase quality especially over proper nouns and Whisper hallucinations. Here is the trace for an episode of Deep Cover set in St Louis. Agent had no context except for what was in the transcript. This is the 4th time it had returned to this particular sentence:

> 29. "emo's cup" — hmm, one more thought: "And there was a cup. You're like, that's dad's emo's cup." In context of a car photographed from helicopter... a cup in the car. "Dad's brand cup" — QuikTrip? "QT cup"? "emo's"... Hmm, could it be "that's dad's. You know, cup"? Or "Imo's"! IMOS — Imo's Pizza is THE St. Louis pizza chain! "Dad's Imo's cup" — Imo's Pizza cups! That's very plausible: St. Louis family, Imo's cup in the car. High confidence? Imo's is iconic StL. "emo's" → "Imo's" — phonetically identical. I'd include this one: from "dad's emo's cup" to "dad's Imo's cup". Moderate-high confidence given St. Louis setting. Include with context

Claude 4.6 didn’t get it. A higher level model might have, but 4.6 is the last model that doesn’t emit erratic content policy rejections for this task.

why does newer Claude reject this
It trips a content restriction. It’s vague, but I’m guessing because the podcasts are adjacent to/are copyrighted material.

What’s really interesting is it’s also tied to thinking level. The sweet spot for quality with low refusal is Claude 4.6 high effort, go higher in effort and and it will trip too. 4.7 is hit and miss with high and often with xhigh and above 4.8 is often. I didn’t try Fable. Sonnet wasn’t good enough quality wise.

It also seems to be related to asking it producing what it knows is copied content. I originally had it return full corrected transcripts and that hit all the time. Then I started having it return just correction lists and that works much better.

These big models recall is absolutely astounding. I built the transcript pipeline so I could search and find that episode that I kind of remember part of from a year ago. I asked fable to search the transcripts to find a particular Conan Needs a Friend episode (Katakai, as god made her!). It knew there were actually multiple episodes where that anecdote was told _before_ it had even searched. When Kimi K3 does a correction you can see in the thinking trace that it instantly recalls people and works before it reasons in to whether it’s certain that’s right.

I’m still testing, but Claude is going to get replaced for K3 when I get some time. It’s very good at this work.

> Below is an example of Kimi K3 using this for Unsplash, a free images provider that models commonly use to find hero visuals and other filler images. On top is the chain-of-thought reasoning, and below is the Unsplash CDN image that the model thought about when it generated the ID.

I wonder how many image ids it memorized and whether it would always use those or search for new images?

I would hope that they really didn't move in the direction of having the model memorize image IDs, as that could be easily replaced with an embedding-based image search tool call.

Although, it is possible that it was unintentional as the model was RLHF'ed into memorizing images.

If you train on any form of HTML, or any code that includes URLs, your model is going to learn them somehow.

Stripping URLs would certainly break some code, and also train the model with unhelpful behavior. You want your coding model to write real URLs, not write "return redirect(<|URL_STRIPPED|>)".

Seems Kimi K3 really thinks a lot more than other models
Cranking up the reasoning effort seems be a trend recently and to me it looks desperate, like cranking up the wattage of a processor when you can’t get other, better gains.

Like yeah, you made it perform better, but I could’ve set any other model to xhigh and spent all the tokens and patience.

It’s surely enough to generate buzz for some headlines about “Fable-level performance”, but I would’ve expected people “in the know” to push back at this more.

I feel like everyone does it though. Like who really runs frontier LLMs at `max`?

It feels like a benchmark-only setting, just for ArtificialAnalysis really.

I run fable at max for basically any non-trivial task. Not even for the result of that task, I get so many “I noticed in passing” bullets at the end that have led to productive refactors and caught bugs it’s worth the tokens by itself. Obvious caveats, depends on your language, framework, codebase, prompt… I also have a “no gardening” rule that tells Claude not to bundle unrelated changes or refactors and it really respects it.

Just to add before someone objects, but the tokens! I have only once maxed out a 5x subscription week (which I then upgraded and maxed out a 20x week so it was a big push). I don’t understand how some people are using so many tokens

> Like who really runs frontier LLMs at `max`?

Umm what? Lots of people do