I knew these system prompts were getting big, but holy fuck. More than 60,000 words. With the 3/4 words per token rule of thumb, that's ~80k tokens. Even with 1M context window, that is approaching 10% and you haven't even had any user input yet. And it gets churned by every single request they receive. No wonder their infra costs keep ballooning. And most of it seems to be stable between claude version iterations to…
Surely the system prompt is cached across accounts?
You can cache K and V matrices, but for such huge matrices you'll still pay a ton of compute to calculate attention in the end even if the user just adds a five word question.
I knew these system prompts were getting big, but holy fuck. More than 60,000 words. With the 3/4 words per token rule of thumb, that's ~80k tokens. Even with 1M context window, that is approaching 10% and you haven't even had any user input yet. And it gets churned by every single request they receive. No wonder their infra costs keep ballooning. And most of it seems to be stable between claude version iterations to…
That's usually not how these things work. Only parts of the prompt are actually loaded at any given moment. For example, "system prompt" warnings about intellectual property are effectively alerts that the model gets. ...Though I have to ask in case I'm assuming something dumb: what are you referring to when you said "more than 60,000 words"?
The system prompt is always loaded in its entirety IIUC. It's technically possible to modify it during a conversation but that would invalidate the prefill cache for the big model providers.
The acting_vs_clarifying change is the one I notice most as a heavy user. Older Claude would ask 3 clarifying questions before doing anything. Now it just picks the most reasonable interpretation and goes. Way less friction in practice.
Haven't had a chance to test 4.7 much but one of my pet peeves with 4.6 is how eager it is to jump into implementation. Though maybe the 4.7 is smarter about this now.
The eating disorder section is kind of crazy. Are we going to incrementally add sections for every 'bad' human behaviour as time goes on?
Even better, adding it to the system prompt is a temporary fix, then they'll work it into post-training, so next model release will probably remove it from the system prompt. At least when it's in the system prompt we get some visibility into what's being censored, once it's in the model it'll be a lot harder to understand why "How many calories does 100g of Pasta have?" only returns "Sorry, I cannot divulge that information".
I knew these system prompts were getting big, but holy fuck. More than 60,000 words. With the 3/4 words per token rule of thumb, that's ~80k tokens. Even with 1M context window, that is approaching 10% and you haven't even had any user input yet. And it gets churned by every single request they receive. No wonder their infra costs keep ballooning. And most of it seems to be stable between claude version iterations to…
I’m just surprised this works at all. When I was building AI automations for a startup in January, even 1,000 word system prompts would cause the model to start losing track of some of the rules. You could even have something simple like “never do X” and it would still sometimes do X.
Two things; the model and runtime matters a lot, smaller/quantized models are basically useless at strict instruction following, compared to SOTA models. The second thing is that "never do X" doesn't work that well, if you want it to "never do X" you need to adjust the harness and/or steer it with "positive prompting" instead. Don't do "Never use uppercase" but instead do "Always use lowercase only", as a silly example, you'll get a lot better results. If you've trained dogs ("positive reinforcement training") before, this will come easier to you.
Personally, as someone who has been lucky enough to completely cure "incurable" diseases with diet, self experimentation and learning from experts who disagreed with the common societal beliefs at the time - I'm concerned that an AI model and an AI company is planting beliefs and limiting what people can and can't learn through their own will and agency.
My concern is these models revert all medical, scientific and personal inquiry to the norm and averages of whats socially acceptable. That's very anti-scientific in my opinion and feels dystopian.
> The new section includes: When a request leaves minor details unspecified, the person typically wants Claude to make a reasonable attempt now, not to be interviewed first. Uff, I've tried stuff like these in my prompts, and the results are never good, I much prefer the agent to prompt me upfront to resolve that before it "attempts" whatever it wants, kind of surprised to see that they added that
Seriously, when you're conversing with a person would you prefer they start rambling on their own interpretation or would you prefer they ask you to clarify? The latter seems pretty natural and obvious.
Edit: That said, it's entirely possible that large and sophisticated LLMs can invent some pretty bizarre but technically possible interpretations, so maybe this is to curb that tendency.