Hacker News
What a task costs on Opus 5.5
mycall
|next
[-]
x-complexity
|root
|parent
[-]
https://claude.dev/blog/what-a-task-costs-on-opus-5-5/#try-y...
What's discussed is the same that I've already heard from a dozen or so YT channels:
- Maximise cache reads to the fullest extent possible
- Keep to medium/high for almost all tasks. xhigh & max waste tokens.
- Regular compactions before reaching 200k context
KellyCriterion
|next
|previous
[-]
Question:
Where can I see how many tokens the conversation already burned/consumed which is in front of me? Or is this hidden for web UI users?
hodgesrm
|next
|previous
[-]
qazxcvbnmlp
|root
|parent
[-]
hodgesrm
|root
|parent
[-]
Anthropic is SaaS on steroids at least as far as costs are concerned. It’s not entirely obvious how they defend their margins in a competitive market with customers who will be looking hard at bounding costs. If nothing else there’s a market opportunity for cost management products for AI.
simianwords
|next
|previous
[-]
Here are the AI smells
> You don't set out to buy millions of tokens. You set out to build a feature, finish a migration, or run a task. The token count is whatever the model needed to get there.
You can see how forcefully punchy and grandeur it reads. Pangram also flags it as 100% AI written.2
replwoacause
|next
|previous
[-]
taylorfinley
|next
|previous
[-]
hxugufjfjf
|root
|parent
|next
[-]
taylorfinley
|root
|parent
|next
[-]
I work on a carpentry simulator, I'm adding a "scan my slab" feature right now that uses the actual grain from your wood. Here's a real sentence from Opus 5.5 earlier today: "A board turned end for end shows it turned, and one turned a quarter turn shows it turned 90°" I had to stop and ask it what the hell it was trying to communicate. I'm sure we've all got examples like this, but it really is nauseating.
[0] (only link I could find that includes the supposed pattern, maybe don't read if you've ever had suicidal ideation though just to be on the safe side) https://frankreport.com/the-suicide-pattern-of-nlp-was-it-he...
bad_username
|root
|parent
|next
|previous
[-]
Another trait of AI prose is ambiguity. It loves using words or phrases that can are grammatically valid but have two different meanings in two different contexts, and it is up to the reader to look ahead in the sentence and then reinterpret the whole thing. The very common offender is present tense verb vs plural noun (e.g. "tracks" meaning "agrees with" vs "tracks" meaning "plural of track"). This ambiguous English is so statistically prevailing in LLM writing that there has to be a technical reason for this phenomenon.
Reading such text - efficient for LLMs to write, hard for humans to read - is extra labor. And it is no wonder that reader is pissed when suddenly the reading experience becomes an uphill battle.
The last LLM which was fine to read is Sonnet 4.6, but while it writes well, it makes factual errors often.
hxugufjfjf
|root
|parent
|next
[-]
bmitc
|root
|parent
|next
|previous
[-]
I think sometimes it just writes nonsense. As in there are words there, they're all strung together in a grammatically correct way, and the words are "right", but the sentences just don't contain any actual information.
There have been sentences and paragraphs that I have read over and over, and I just cannot extract any meaning from them. It's very, very strange, and I think that's the uncanny valley of it. It's all there, in the sense of words and grammar, but there's nothing there, the sense of knowledge and meaning.
alexgoodhart
|root
|parent
|previous
[-]
And some of the ambiguity I think comes from liability training sets, designed with a preference for not making direct attributions to parties who could claim they were incorrectly characterized. The habit can be seen in like half of Claude messages, where its responses have multiple inert objects or concepts being described through active personalized verbs. Like “the build rejected an undeclared package” as opposed to saying “the build failed because I didn’t update the packages.json.”
fortedoesnthack
|root
|parent
|next
|previous
[-]
I just assume the training data is AI generated hence why it's so uncanny valley. Even though platforms that provide data like mercor and micro1 are supposed to provide "organic" data that isn't output using LLMs I've heard some people I know on those platforms say they generate their work using ChatGPT.