Comment by rich_sasha
11 hours ago
I've noticed something else - as Anthropic models get even more and more superhuman, they seem to serve me more and more casual nonsense.
Not like adding glue to pizza. Here's an example from today (paraphrasing): "you need to run `git merge-base branch1 branch2`. Pay attention to the order of arguments, it is important: `git merge-base` is symmetric and returns the same value regardless of the order of inputs".
So which one is it? Symmetric or not? It's not even one of those where it self-corrects, it just happily contradicts itself halfway through the sentence.
My pet theory is that these frontier models do quite a bit of brute force at the end - some kind of beam search - and silently downgrade you depending on demand or compute availability.
Still not great. This particular nugget is from Sonnet 5, default settings.
I've noticed the same pattern even in GLM5.2. It has always been a thing, but it seems to be getting worse in recent models
It does feel like the kind of thing beam search would fix. The LLM starts the sentence with a claim like "Pay attention to the order of arguments". Around that time it "notices" that the order doesn't matter, but it's already committed to the sentence and has to complete it in the best way still possible
Maybe at some point someone figures out how to train models with a backspace token
Have noticed similar things using it for code review. "Issue #3: This variable is defined but never used. It is actually used later in the method though. So this is not a real issue to be concerned about", that sort of thing.
Diffusion language models do this. I’d guess they are the future.
Or a strikethrough token!
Yes, LLMs are prone to semantic consistency traps. I see a version of this in a system I have that transcribes a lot of noisy low quality audio (radio comms).
Sometimes it works out, in that an unrecognizable word or two is replaced with reasonable assumptions based on the semantics established by the words that came before the signal degraded.
But the bad audio might also result in words that don't align well to what came before, or represent alternate (mis)interpretations. Now this is part of the context and the next several tokens align to this new path regardless of what is said in the audio.
In pre-LLM transcription, you might get a nonsense word or two when the audio transiently degrades, but the specific meaning of the nonsense words doesn't influence the transcription of audio following the degradation.
First we started by extending chain-of-thought to a formal thinking system. You could extend this again with a "thought scratch-pad" or "thinking draft" with n-number of passes without needing to train backspace tokens.
That being said, all of these are just variations of scaffolding on generating more tokens.
I also see stuff like this with Opus (4.8 and 5). It does the work correctly but then the explanation or "aftermath" description is kinda nonsense. It feels like they are tuned hard for agentic coding and everything else is getting worse. There are people that say older versions are also better for creative writing which would confirm this theory.
> they are tuned hard for agentic coding
Well you can't blame them, it's what brings in the cash. Other uses will suffer proportionally.
Claude has always noticeably degraded under heavy load. Opus goes from "decent to work with" to "dumb intern" depending on whether you're working at 3 AM west coast or 10 AM - 5 PM. It's part of why I cancelled my subscription - "max" plans and "extra high" effort are meaningless when there's so much variability between model availability, model performance, and harness bugs every day and every week.
I don’t find Sonnet usable at all. Especially for the price. I’d much rather get half the usage of Opus.