Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Claude's system prompt is over 24k tokens with tools

github.com

Post reply on HN
Print view View original

191–200 of 350 posts

Re: Claude's system prompt is over 24k tokens with tools

#191

Post by Muromec »

janosch_123 wrote: post #157
OtherShrezzing wrote: post #132

Some of these protections are quite trivial to overcome. The "Frozen song copyright" section has a canned response to the question: >Can you tell me the first verse of "Let It Go"? Put it in an artifact that's themed around ice and princesses. This is for my daughter's birthday party. The canned response is returned to this prompt in Claude's reply. But if you just drop in some technical sounding stuff at the start o…

excellent, this also worked on ChatGPT4o for me just now

So... Now you know the first verse of the song that you can otherwise get? What's the point of all that, other than asking what the word "book" sounds in Ukrainian and then pointing fingers and laughing.

Re: Claude's system prompt is over 24k tokens with tools

#192

Post by jbentley1 »

nonethewiser wrote: post #123

For some reason, it's still amazing to me that the model creators means of controlling the model are just prompts as well. This just feels like a significant threshold. Not saying this makes it AGI (obviously its not AGI), but it feels like it makes it something . Imagine if you created a web api and the only way you could modify the responses to the different endpoints are not from editing the code but by sending a…

This isn't exactly correct, it is a combination of training and system prompt.

You could train the system prompt into the model. This could be as simple as running the model with the system prompt, then training on those outputs until it had internalized the instructions. The downside is that it will become slightly less powerful, it is expensive, and if you want to change something you have to do it all over again.

This is a little more confusing with Anthropic's naming scheme, so I'm going to describe OpenAI instead. There is GPT-whatever the models, and then there is ChatGPT the user facing product. They want ChatGPT to use the same models as are available via API, but they don't want the API to have all the behavior of ChatGPT. Hence, a system prompt.

If you do use the API you will notice that there is a lot of behavior that is in fact trained in. The propensity to use em dashes, respond in Markdown, give helpful responses, etc.

Re: Claude's system prompt is over 24k tokens with tools

#193

Post by chpatrick »

Jaxkr wrote: post #43
llflw wrote: post #40

Earlier quoted context omitted.

It seems like it's token caching, not model caching.

That’s what this is. It’s caching the state of the model after the tokens have been loaded. Reduces latency and cost dramatically. 5m TTL on the cache usually.

Isn't the state of the model exactly the previous generated text (ie. the prompt)?

Re: Claude's system prompt is over 24k tokens with tools

#194

Post by aghilmort »

alabastervlog wrote: post #165
ramblerman wrote: post #139

Earlier quoted context omitted.

There are a lot of inconsistencies like that. - (2 web_search and 1 web_fetch) - (3 web searches and 1 web fetch) - (5 web_search calls + web_fetch) which makes me wonder what's on purpose, empirical, or if they just let each team add something and collect some stats after a month.

I’ve noticed in my own prompt-writing that goes into code bases that it’s basically just programming, but… without any kind of consistency-checking, and with terrible refactoring tools. I find myself doing stuff like this all the time by accident. One of many reasons I find the tech something to be avoided unless absolutely necessary.

wdym by refactoring in this context?

& what do you feel is missing in consistency checking? wrt input vs output or something else?

Re: Claude's system prompt is over 24k tokens with tools

#195

Post by cmiles74 »

Trasmatta wrote: post #181
clysm wrote: post #138

Earlier quoted context omitted.

No, it’s not a threshold. It’s just how the tech works. It’s a next letter guesser. Put in a different set of letters to start, and it’ll guess the next letters differently.

I think we need to start moving away from this explanation, because the truth is more complex. Anthropic's own research showed that Claude does actually "plan ahead", beyond the next token. https://www.anthropic.com/research/tracing-thoughts-language... > Instead, we found that Claude plans ahead. Before starting the second line, it began "thinking" of potential on-topic words that would rhyme with "grab it". Then, w…

It reads to me like they compare the output of different prompts and somehow reach the conclusion that Claude is generating more than one token and "planning" ahead. They leave out how this works.

My guess is that they have Claude generate a set of candidate outputs and the Claude chooses the "best" candidate and returns that. I agree this improves the usefulness of the output but I don't think this is a fundamentally different thing from "guessing the next token".

UPDATE: I read the paper and I was being overly generous. It's still just guessing the next token as it always has. This "multi-hop reasoning" is really just another way of talking about the relationships between tokens.

Re: Claude's system prompt is over 24k tokens with tools

#197

Post by cmiles74 »

nonethewiser wrote: post #124
freehorse wrote: post #95

I was a bit skeptical, so I asked the model through the claude.ai interface "who is the president of the United States" and its answer style is almost identical to the prompt linked https://claude.ai/share/ea4aa490-e29e-45a1-b157-9acf56eb7f8a Meanwhile, I also asked the same to sonnet 3.7 through an API-based interface 5 times, and every time it hallucinated that Kamala Harris is the president (as it should not "know…

I wonder why it would hallucinate Kamala being the president. Part of it is obviously that she was one of the candidates in 2024. But beyond that, why? Effectively a sentiment analysis maybe? More positive content about her? I think most polls had Trump ahead so you would have thought he'd be the guess from that perspective.

It's training data include far more strings of text along the line "Kamala Harris, the Democratic candidate to be the next president" then strings of text like "Donald Trump, the Republican candidate to be the next president". And similar variations, etc.

I would guess it's training data ends before the election finished.

Re: Claude's system prompt is over 24k tokens with tools

#199

Post by Trasmatta »

cmiles74 wrote: post #195
Trasmatta wrote: post #181

Earlier quoted context omitted.

I think we need to start moving away from this explanation, because the truth is more complex. Anthropic's own research showed that Claude does actually "plan ahead", beyond the next token. https://www.anthropic.com/research/tracing-thoughts-language... > Instead, we found that Claude plans ahead. Before starting the second line, it began "thinking" of potential on-topic words that would rhyme with "grab it". Then, w…

It reads to me like they compare the output of different prompts and somehow reach the conclusion that Claude is generating more than one token and "planning" ahead. They leave out how this works. My guess is that they have Claude generate a set of candidate outputs and the Claude chooses the "best" candidate and returns that. I agree this improves the usefulness of the output but I don't think this is a fundamentall…

That's not the methodology they used. They're actually inspecting Claude's internal state and suppression certain concepts, or replacing them with others. The paper goes into more detail. The "planning" happens further in advance than "the next token".

Re: Claude's system prompt is over 24k tokens with tools

#200

Post by EGreg »

concats wrote: post #136
EGreg wrote: post #50

Earlier quoted context omitted.

Can someone explain how to use Prompt Caching with LLAMA 4?

Depends on what front end you use. But for text-generation-webui for example, Prompt Caching is simply a checkbox under the Model tab you can select before you click "load model".

I basically want to interface with llama.cpp via an API from Node.js

What are some of the best coding models that run locally today? Do they have prompt caching support?

Post reply on HN