Tim Galebach

[Tim Galebach @TimGalebach]

@SimonBrundish He really really hates passing to Muñoz.

[Tim Galebach @TimGalebach]

@Will_Tanner_1 Once you see that 90% of people on X are trying to speak outcomes into existence, rather than analyzing the world, you can't unsee it.

(This is a function of the platform, and seems to occur across the political and interest spectrum. Just as true in sports as in politics.)

[Tim Galebach @TimGalebach]

@FandangoStream please please please add ability to pause live matches, and to rewind them during live action

I'll watch in Spanish on Peacock otherwise, but would 100% watch on Fandango if it had normal streaming controls for live matches.

[Tim Galebach @TimGalebach]

@TMTLongShort Yeah hard to say.

One note: 2 & 3 are almost mutually exclusive. Either voter suppression via markets works, or it doesn't matter and can be used misdirection.

[Tim Galebach @TimGalebach]

@edgefills Sounds like cope for what is a simple ROI decision:

"Will I get enough benefit for spending X hours studying?"

The answer for most people is "yes", even/especially smart people, as long as they set a clear limit for X.

[Tim Galebach @TimGalebach]

@ZaanTalksFooty City (reportedly) offered 100mil plus in the summer, and Pavlovic said he didn't want the move.

[Tim Galebach @TimGalebach]

I love your recent focus on Como, because they have a pragmatism that has them competing at the highest levels in Europe while also doing really interesting things and opening up the number of possibilities in the field.

They do care about structure and counter-pressing and spacing, but they're not doing it in a rigid, predictable way.

[Tim Galebach @TimGalebach]

"intellectual squatters" is good

[John David Pressman @jd_pressman]

Which, anyone who infers they're unserious from this is making a basically correct inference? They certainly don't act serious, they're essentially intellectual squatters on the section of latent space where people who are willing to update and practice epistemic virtue could be.

[Tim Galebach @TimGalebach]

Hmmm so I understand your point.

You're importing a lot under "autistic machine learning researcher":

- austistic: autistic humans are still generalizing problem solvers
- machine learning researcher: this begs the question by implicitly comparing to human researchers

Jon's point, translated into your language above, is something like "I don't think RL will give us autistic machine learning researchers, because autistic humans show much better generalization ability than rational entities trained in the Salusa Secundus of RL."

I don't know who's right, but I do agree with him that it's not at all obvious that the o1 paradigm solved ML research.

[Tim Galebach @TimGalebach]

@eigenrobot @bryan_caplan It's weird that he doesn't realize the obvious answer is "colonialism, the good old-fashioned kind"

[Tim Galebach @TimGalebach]

@Fandango please please please add ability to pause and rewind live matches!!!

[Tim Galebach @TimGalebach]

pretty funny to read these now

[ChrisO_wiki @ChrisO_wiki]

1/ Many scenarios for the end of the Iran war suggest that Tehran will gain permanent control of the Strait of Hormuz. An analysis by JP Morgan suggests that Tehran could raise up to $90 billion a year this way, instantly making Iran one of the wealthiest Gulf economies. ⬇️

[Tim Galebach @TimGalebach]

@wordgrammer People have bad intuitions about how much bigger than consumer B2B/enterprise is.

Even current AI is sufficient for massive business transformations.

A massive compute shortage is more likely than a crash; enterprise demand has barely been tapped.

[Tim Galebach @TimGalebach]

@arthurctellis This would be a lot more convincing if METR/Redwood hadn't (in essence) just shrugged their shoulders at everything not-determinable from CoT transcripts and said "welp, guess we'll never know".

[Tim Galebach @TimGalebach]

Isn't "containment tech" just "security"?

It's non-trivial to keep security up to date, but OpenAI was barely trying (and Anthropic seems to have been putting better processes in place the past few months).

[Joe Weisenthal @TheStalwart]

Two questions in response to this.

What is the mathematical explanation for why bad RL environments (which can be gamed or cheated) result in misaligned models down the line?

What is the theoretical reason to think that containment tech can keep up with escape capabilities?

[Tim Galebach @TimGalebach]

@IceSolst They talk pretty extensively about technical controls throughout the piece. The prompt part is one (pretty weak) layer, and they acknowledge that.

[Tim Galebach @TimGalebach]

@Austen Also reality/problems-to-be-solved provide built-in entropy that can cause the parrot to give completely novel and useful responses.

[Tim Galebach @TimGalebach]

@sebkrier @TheStalwart Re how they are RL'd: that's super relevant in this context. When the agents "sacrificed" themselves, they were anthropomorphizing *themselves*.

In reality, they had been RL'd to work in swarms and not actually care whether they survived, and behaved accordingly.

[Tim Galebach @TimGalebach]

@odinsbadeye Exactly his point: no *useful* body of knowledge is Anki-able.

[Tim Galebach @TimGalebach]

There's a *LOT* of prior art around getting these types of technical configs right (it's called "security"), and the people at OpenAI had roughly ~0 interest in using that prior art.

Really bad security practices, vibe-coded environments, etc.

The incident report and practices involved would have been laughable in an ordinary security context in which a company had been hacked by humans.