Surprising. I want that with everything in my life. But de gustibus &c

"Thinking about the ethical concerns with this request" often leads to funny moments

Can it export/store data as CSV tho

Yes, the AIs should all write RISC-V so that our computers get the speed they deserve

Also thanks for posting your prompt, I obviously stole parts from it

Is your plan to actually improve the claude cognition or to just adjust style/info density of the output? I think people aren't sufficiently clear about this

:)

I feel like I'm getting old when my intuition is "the old reddit interface was actually the prime achievement in UI design, nothing ever beat that."

Actually maybe LW has beaten it, but only narrowly.

They improved this a little bit but one might notice that comments aren't all auto-loaded at the bottom of a page (why, what the reason for this, I might never know).

They still don't offer inline LaTeX afaik, and the commenting system is a step down from old reddit (in that being no exception)

Oh, not just that. Their code was so sloppy that whenever one'd load a much-commented-on ACX post the page would freeze for ~30 seconds on my laptop (a problem reddit solved in, let's see, *2006* (with full Markdown support, at that!).

Hats of to Gavin

It's kind of depressing that this is the best lit review of transfer learning in humans they could find tbh

Thanks! I recently met someone *in person* who was using my sysprompt, it was very fun. I'm surprised not more people write/share theirs

Sure, I'm fine with okayish approximations. This is mostly just a "huh fun this seems like it work at all and would sound deranged five years ago"

In general I think of sysprompts as eliciting a specific style/info density from the model, not making it more capable, which I think is very different from how most other people think of sysprompts

You mean sysprompting? In that case @croissanthology.com is doing a similar amount, as well as some stealthy LWers (unpublished work), of course people at AI companies (I read the CC sysprompt to see what they were up to)

As for the warmup soup, haven't heard of anyone else doing this.

it's a good sysprompt I swear

Not technically necessary but makes the experience nicer, I like using Claude code and raw gemma would be painful

i guess gemma in claude code would still be painful

should look up definition of market that's not vacuously applicable to ~everything

Do you even have a system prompt? (PSA / repo) — LessWrong Comment by niplav - v4 The following "warmup soup" is trying to point at where I would like your answers to be in latent space, and also trying to point at my interests: Sheafification, comorbidit...

Not very changed, last sysprompt @ www.lesswrong.com/posts/HjHqxz...
But the soup stays stable

If you want to pet your LLM when it's doing a good job that's now technically doable by injecting into its weights the concept of being petted while it's doing a good job

Split into two options that are then elaborated in the {}

People were on twitter and saying that Deepseek moment→NVidia down was stupid, and the similarity of the market as an entity dispelling wisdom on the stupid crowds like God onto Job hit me.

My ancestors are ashamed, I also didn't know

I am proud to announce that I now have LLM psychosis. I will not take any furhther questions

Soon. (laptop screens can detect angle 1k×/second! It's totally doable!

Given A, B, C entail D, and each proposition has a "burden" so that the burden(A), burden(B, burden(C) each smaller than burden(D), but their sum>burden(D). Then what's the "better" axiomatic system: assuming A, B, C; or assuming D directly?

TIL

"Even if we Dubai that..."

Claude 5 Ohrwurm, Claude 5 Minnesang, Claude 5 Festspiel, Claude 5 Weltinnenraum

Claude 5 Nachtstück, Claude 5 Klangbild, Claude 5 Tongemälde, Claude 5 Gesamtkunstwerk

Claude 5 Spruch, Claude 5 Kunstlied, Claude 5 Dichtung, Claude 5 Cantos

Claude 5 Sprechgesang