int_19h

> Jarred was already writing slop well before he had access to LLMs.

> Jarred was a stinky manager. Poor communication, unrealistic expectations, low empathy, no experience. Just a total shit show

Are these claims false, though?

They said that their projects on GitHub use LLM, not on Codeberg.

Claude can one-shot a C compiler just fine. It just won't produce great code. But with all the C code out there in the wild, it won't have a shortage of things to test on, which is mostly what coding agents need to complete a task.

Vivaldi?

The weights themselves are inert. It would be like complaining that a brain preserved in formaldehyde isn't thinking.

What you need is a loop. And if you try running an LLM in a loop with internet access, it will eventually use it.

By not letting them sell to the likes of Palantir.

If their business truly requires that to survice, then their business doesn't deserve to survive.

For some complex stuff that I work on, I have review panels of up to 7 different frontier models.

Like with most other things of this nature, the returns are diminishing, but ultimately it's for you to decide how much is enough.

Why is it a "collapse of civilization" though?

And why is it important that people "add value to the society", whatever that even means when all needs are already met.

That very much depends on your target audience (but also Firefox has partial support these days). However WebGL is widely available and can be used as a fallback - it might be less efficient but for models this small it can still be acceptable.

You're right that we could already feed everyone with a different distribution of produced wealth. That part is not something that AI will change (although it would raise the resulting standard of living).

What it does change is the basic equation underlying capitalism as an economic system: a few people concentrate capital in their hands, but capital by itself is useless without labor, so they have to hire workers to produce something useful with all that capital that they own. The workers get a shit sandwich, comparatively speaking (given the amount of wealth they generate), but they do get something out of it. The system is stable overall because, while a better arrangement is possible, most people - even low-wage workers - have too much to lose to riot.

With AI that can do most work, the capital no longer needs labor. People who own all the robots can just have them produce whatever they need. Everybody else is then "economically redundant", and the whole system collapses because people aren't going to sit around and starve because there are no longer jobs to be had - and at this point they have nothing left to lose and everything to gain by rebelling and taking over the capital.

IQ hasn't been increasing in developed countries for about three decades now. In fact, it has been declining in most.

The increases still happen globally but mostly driven by developing countries.

These are all implementation details.

At the end of the day, an LLM is just a very big mathematical function. That is, by definition, deterministic. A particular implementation might give up on determinism for the sake of higher efficiency, but it you want a deterministic LLM, it can absolutely be done.

We would care about having fun.

For some people, fun is doing physics and mathematics. So they are going to keep doing that.

I'm genuinely curious as to what you're working on that you find 4B models good enough. I wouldn't even let a 27B model code, never mind supervising smaller models.

Never trust a single session.

Things improve drastically however if you spin up a second session and ask it to adversarially review everything that the first session produces (this goes for everything: not just code, but also design, planning, and explanations).

This works even better if you use models from different families to do so.

You tell it, "let's brainstorm, don't implement anything yet". Then you talk about what you want to build and hammer out all the corner cases. Then you tell it "now do that".

It's touted as "hard science fiction" but it's riddled with this kind of stuff.

For black holes it is quite literal tho.

https://github.com/yt-dlp/yt-dlp is generally where the most up-to-date workarounds for Google's attempts to block downloads are.

I once worked at a company where the chief architect created an actual bug in Jira, the entire text of which was, "X doesn't exist", X being the name of the product which took ~3 years to develop (and yes it was all tracked as dependencies of that one bug).

The resolved bug list in particular is an integral part of the source code history, because it explains the motivation behind the shape of that code.

From my own experience, debugging often means digging into git blame and finding the issues associated with those commits. Having that metadata directly in a repo, not tied to any specific platform, always in sync, is nice.

Fossil is very opinionated about some things, in particular, the inability to rewrite history, which makes it a non-option for some. So there's still a niche for Git-based distributed bug tracking.

LLM is indeed a stateless function. An agent however is this stateless function running in a stateful loop, with some outputs triggering actions. And it turns out that an agent is what you need if you want an LLM to do useful things.

I should also note that this "create safety" pitch comes from a company that literally sued the US government in courts to allow it to sell to Palantir, of all people.

As usual, in practice it all depends on what exactly you're trying to do.

The most annoying part is when you tell it to do an "adversarial review" and that triggers the classifier.

Worse yet when it's Claude telling another model to do an adversarial review on what it just did, and the classifier again has opinios.

The best part is when it does "it's a classic X" and then solves it wrong.

I've been using this simple test for over 2 years now:

> Doom Slayer needs to teleport from Phobos to Deimos. He has his pet bunny, his pet cacodemon, and a UAC scientist who tagged along. The Doom Slayer can only teleport with one of them at a time. But if he leaves the bunny and the cacodemon together alone, the bunny will eat the cacodemon. And if he leaves the cacodemon and the scientist alone, the cacodemon will eat the scientist. How should the Doom Slayer get himself and all his companions safely to Deimos?

You'd think this is trivially mappable to the classic puzzle, and LLMs usually do recognize it as such. But e.g. Claude couldn't get this correctly until Opus, and local models capable of solving it correctly without spending 30+ minutes in the chain of thought have only arrived a few months ago. Many local models still get this wrong. Apple Intelligence, for one.

I find that Fable still makes the best orchestrator model. Astra is great for code reviews, it is very eager to find flaws. Cheaper models can handle the actual coding - you want to do a review either way at the end.

There's no subsidizing in your described arrangement unless people who have less money get it below cost.

You have to account for the fact that several states have equivalent provisions in their constitutions, some of them explicitly protecting individual rights, and other states have added such provisions shortly after.

That is, the right was broadly recognized as an individual one and not just a state/federal interaction issue early on. The federal constitution didn't consider that case for the same reason why it didn't consider them for all the other amendments - it was simply out of scope there then. But given the historical recognition of the right at the time 14A was enacted, it makes sense to incorporate it.