kadoban

It was created by Nixon, so yeah I'm sure the early days were a little weird (and it hasn't gotten any better since).

> The situation you're imagining is one of a global all encompassing depression caused by a single government.

Doesn't even have to go that far (though it pretty easily can, the US is a big enough market). If the US causes a refugee crisis that affects one of China's allies, or fucks with the global oil supply enough or a thousand other things, that stuff is still annoying to China even if it's usually relatively better for them than it is for the US.

> Not unthinkable, but the Americans would feel that deeply. Do you think they want to?

The people in the US choosing do not care if the rest of America feels it or not.

Can these big companies that want to distill not just get the models? Like how many machines at how many different providers are running Opus 5? Nobody is just throwing the model on a thumb drive and taking it home, or grabbing an old drive that somebody was too lazy to scrub or whatever?

That would be vastly more illegal, but would it even matter? What would the practical downside be?

> "Don't interrupt your opponent when they're making a mistake"

Yeah, that's the obvious answer, but I suspect there's a point where if the US fucks things up too quickly and too messily, there's fewer customers for China's crap and/or the random throws of monkey shit drag China into messes they don't feel like dealing with.

Slow erosion of the US's global power is likely better for them than an implosion that might cause collatoral damage.

Is it me or is that not what that graph shows? That line has ~no relation to the datapoints it's purporting to be a fit for.

I think it's not even that sophisticated here, the providers are just lying.

The context of this is which party China would prefer to win the midterms. The dems winning would distract Trump from being able to fuck up the rest of the world a bit.

My bet would be that China doesn't care much either way, they win regardless.

Yeah something must be actively broken or just an awful extension for it to have that much effect.

Most things you can mess with the big effect is like, oh a thousand tokens ended up in the ~system prompt, or 10% extra or fewer work based on extra tool calls or churning through thinking or whatever.

Harness stuff if it's 8x worse that's like, it's fucked and broken, something went _wrong_.

> the triple digits IQ of Hegseth, Trump and Bessent.

One digit each.

At some point they might prefer stability and sanity. I would bet that at this point they don't care much either way.

China is playing long, most of the ~unchanging things are in their favor and the US already proved that they're unreliable. People won't forget that soon even if Trump is gone.

You know you can define your own provider filters and orderings, right? The filters and such are not _that_ advanced, but it might do what you need if you haven't already tried that.

You'd have to look where the extra tokens are coming from. If it's using extra turns or doing extra work because a tool it ~wants is missing, then adding extra things will help. Otherwise, it won't.

If it's missing guidance that would help, system prompt additions might help.

8x-10x the tokens is wild, is this for some tiny artificial benchmark? That's just too much extra for something not to be just broken.

Pi is just a nice base and it has defined extension protocols and such. You might as well start there, it's just easier and going from nothing to working to adding whatever functionality is like 2 minutes.

> "this is a benchmark" is such an easy category to determine

I mean, it's not _that_ hard to determine most likely, and/or it's hard to be sure you didn't get found out by llm-assisted analysis on your traffic. It's not going to be a one-shot request and response it's going to be a whole bunch of them in an artificial way, by nature.

And then anything you found is single-use only if you're paranoid because even if they find out later, they have your benchmark now (because you sent it to them to use it even if you don't publish it).

> AlphaZero (which later became Leela? I'm not sure)

Pretty much. DeepMind never released their code or models, just the main ideas of their techniques and training.

Leela, which came out of the go-ai world, implemented the paper(s) and became the de-facto version everyone used for a while until other projects improved over it on both go (mainly katago) and chess (mainly stockfish).

Yes. The ways to improve are: faster search (which lets you search more deeply) or better evaluation of a position or better heuristics.

Doing this in a way that doesn't get you noticed or fucked with is potentially going to be quite difficult.

If they have a "hey we're being benchmarked" mode, which is not hard to imagine, avoiding tripping it is going to be annoying and difficult to prove.

Way back when I worked in a restaurant, there was that but then also if a customer complains they also go clean.

They're never going to be like "Oh, someone shat on the floor? We'll deal with that at the next scheduled cleaning."

> The fact that they didn't understand this when they proposed the idea shows you how little they knew of the Linux ecosystem, and the fact that they're still bringing up so many years later shows that they haven't learnt anything even after all this time.

The fact that they didn't ask "why?" at the time is also telling (or they didn't listen or remember, whatever reason it's not in this anecdote). They expected the greybeards who had been maintaining this shit for years to shut up and accept whatever shitty proof-of-concept looked good in a screenshot.

> It's surprising it's that high (8%). I would imagine it takes as long to mop a floor that has pee droplets or mop a floor that doesn’t.

You might have to mop more often from complaints or inspection, maybe?

> AI slop nothing burger. The “exploit” has nothing to do with coding agents.

It's pretty small potatoes, but it is a harness ~bug that they treat this so poorly. It getting triggered before some of them even ask you if you trust the directory is pretty bad.

Yeah, thanks, I forgot that existed. That definitely weakens my point quite a bit.

It's still not _quite_ the same thing because an advantage in playouts is not a great model for how a stronger/weaker player dynamic actually works, but it is something for sure, way more than the nothing that I said exists.

That's the basic idea, yeah.

When playing white in handicap games, you want to make your opponent uncomfortable.

Play moves where the simple/safe/obvious move is just a little bit bad. Force them to choose between complex fights or a slow death of 100 slightly suboptimal moves.

It feels really wrong to defend like 10 times in a row, so if you make them do that they'll lash out at the wrong time and you can take advantage.

You also want to look for moves where...even if their best response means it's even or a little bit worse for you, there's ~reasonable responses where you win out or it goes complex.

A lot of the time it's not even crazy complex fights, it's more just situations where the judgement of what is more points is difficult.

(Note: most of this stops applying as strongly if it's a teaching game, which most handicap games are, there you have other considerations besides winning)

> That humans can make sufficiently strong calculators has never been a dispute in my mind.

Within our lifetimes (unless you're quite young) it was doubtful if a go ai would ever beat a decent human. Same was true for chess a generation or two earlier. It's news because it's the handoff of man to machine being the best at a particular thing.

This current news is news from the other way, this human did _exceptionally_ well.

There are tweaks to move choice like that, but it's all evaluation time. None of that happens in training. Zero times in the RL loop does katago see a two stone game against a weaker player.

You can kind of tweak towards play this metric or that, but it's not the same.

The headline is misleading, but this is still huge. 2 stones against katago is insane, I'd never have guessed we'd see that, ever.

> Also, the human played a strategy tailored to that huge initial advantage. He said that the AI did not handle this particularly well, and played high probability moves instead of trying to lure him into a mistake.

Yeah, katago's training is not really focused at all on handicap games, because it's by nature learning from even games against similar-strength opponents.

It doesn't have specific training from playing in a way to exploit a weaker player. In a handicap game you have to give your opponent opportunities to fuck up if you want to play optimally.

If a move loses 0.0005 points if the opponent plays optimally, katago won't play it even if there's ~zero chance a weaker player would play it right.

There have been go AIs that tried to train more directly on uneven opponents, one called "sai" comes to mind, but katago has huge advantages otherwise and won out over the others (for very good reason, it's a great project).

That's not ironic, that's _hey, look over there!_

The wording of the announcements, sounds like a student synthesized it. Can't imagine why, hopefully it wasn't a fatal mistake.

Pi.dev has an extension that does it. None of the models I use seem to have that failure mode so I haven't bothered throwing it in.

The failure mode I run into commonly is agents just stop sometimes. Even sending a "." Or something they start back up, but I haven't worked out exactly how to fix that generally in harness, bit unclear how to tell if they're done or just derped to a stop.