It's arguably even better than the most famous answer to that question.

which is?

> A 9M model can't conditionally follow instructions

How many parameters would you need for that?

My initial idea was to train a navigation decision model with 25M parameters for a Raspberry Pi, which, in testing, was getting about 60% of tool calls correct. IMO, it seems like around 20M parameters would be a good size for following some narrow & basic language instructions.

Ok. This makes me wonder about a broader question. Is there a scientific approach showing a pyramid of cognitive functions, and how many parameters are (minimally) required for each layer in this pyramid?

That may "answer" a specific question. And all llms can do as they include manpages in training data (and any Agentic thing can search) however the value in reading documentation is that one can find different angles by learning about different options, which allow tontackle problems from a different perspective. The answer to a question is constrained by assumptions which are part of the question.

My experience with LLMs is that they often give me different angles that I didn't think of.

Please no. I want to read the manual without having to talk to anything.

How do you think llms reply to you?

i have made llms read manpages, it is great lol

It's still a case of asymmetric information.

A local model running on a phone owned and controlled by the vendor is still not really exciting, imho.

It may be physically "local" but not in spirit.

Are there any scientific results showing that this helps?

Isn't that the same stuff as in soldering flux?

Smells good, for sure. But I don't know if it promotes good health.

Is that a good thing or a bad thing? I associate "claws" with bot spam but maybe it means something else.

You can use any technology for good and for bad.

In this case, pumping around information so social networks appear to be one unified system is a good thing because you don't have to visit them all to check if there are new posts, etc. and you can avoid getting caught in an algorithm.

Anyone who wants to do that doesn't need to use AI for it

what's a Claw

the "set it and forget it" Ai agents, largely driven by OpenClaw

Yes, there needs to be a government public service counter where you can go with all your BigTech issues and complaints.

This is one of the goals of the digital services act.

The EU isn't as bad as some Americans want to believe.

The EU’s heart is in the right place, which can only rarely be said of the US.

But the EU’s approach is often backwards. When product managers have to ask the government if it’s ok to ship a feature, something is wrong. When the government responds that it can’t say in advance, you’ll just have to ship and see if you get fined, something is really seriously broken.

If a company is about to produce millions of physical products, I think it is quite ok if they first check with the government to see if that is a good idea.

Same with social media features that are rolled out to millions of users.

If the EU was prepared to give advance permission, that would make sense. It would be slow ("Hey we want to rename 'username' to 'login'" -> "we'll get back to you in a couple of months"), but it would make sense.

But, if you'll re-read my comment, my complaint is that the EU will not pre-clear features. They will only punish after the fact if they decide it was a bad feature.

And that's even assuming you're correct that the bureaucrats themselves know what is a good idea. Which I'm skeptical of. I think they're more likely to be correct than, say, Facebook... but that's a pretty low bar.

This is about Digital Services Act, not physical products.

It's especially funny to change your coworker's system prompt like that.