POV: you try to write fiction with Claude kun.

Claude homies for what use cases do you use Opus? I do think it's plausibly better than Sonnet, but the message limits on Opus are so much harsher (and it's so much slower and so much buggier) that I can't really justify any Opus usage (except maybe before bed or something).

I gave a homeless(?) person £20 today, which really wasn't effective, but made me feel much better about myself. I guess I'll offset this by donating an extra £100 to GiveWell recommended charities in global health and development.

What's the best answer for this? Is it just that the 2020 models weren't actually capable enough for RL training for long COTs to be viable or?

Frame.

I predicted the Mexico/Canada tariffs wouldn't go through when the delays happened, and criticised people who criticised markets for not believing them. I apologise. Maybe the tariffs will yet be rescinded, but I think I was too arrogant in the position I staked regardless.

Computer science/mathematics does place in principle known limits to the amount of intelligence* that can crystallise in a given circuit of a certain size. *: for suitable operationalisations of "intelligence"

I want to read a @gracekind.net tweet about this. turntrout.com/self-fulfill...

Oh dear indeed.

How did Threads "lose" to Bsky as the Twitter alternative?

Have we considered?

Powerful frame.

Dario??? How is Dario not an Altman type figurehead? There was a time when Dario was in the background/not prominent, but that time was before the Claude cult really took off.

Big Yud coming out hot.

I don't regret having been a Musk fan and going to bat for him. It's sad he fell so far of the deep end, but such is life.

A very exciting new account has joined Bsky. [You know she's legit because she has the good sense to follow me. 😌

The longterm flourishing of human civilisation is one of the most pressing concerns of our time.

Longtermism is pretty unpopular currently, and some EAs are ~disowning it in a defence of EA, but I don't want to throw the longtermism baby out with the bathwater. It makes strategic sense: longtermism isn't needed to justify x-risk work anymore. But I still fundamentally stand by the thesis.

Bemoaning that the people using DOGE gutting PEPFAR to dunk on EA doesn't achieve any practical benefit, but it makes me feel better about the criticism. If your criticism is so unrooted from the truth and you're unwilling to update on contrary evidence, then I won't be bothered by it.

H/t @codytfenwick.bsky.social

To: @gracekind.net You returned to Twitter?

Holy banger.

I don't think they tried to engineer Grok to be anti-woke. Like it should be possible to finetune or RLHF a model to be conservatively biased. Labs just don't do this because that's trading wokeness for anti-vaxx and other MAGA brainworms.

Today's my birthday.

I'm in London for the next few days. If you want to meet up DM me, and I'll see if I can work something out!

It is February 20th 2028. In no quarter during the last three years did utilisation adjusted TFP grow by > 1%. What happened?

To: @grantslatton.bsky.social

I owe Kelsey an apology, I wasn't quite familiar with her game.

Two things can be simultaneously true: 1. Generative AI will be transformative in the medium to long term. 2. The Gen AI market is a bubble. The dotcom bubble was based around a real, very important technology.

Model aggregator thesis. If many models are at the frontier, and different models are somewhat better at different use cases, then what you want is a single subscription that lets you "plug and play" models on a per prompt basis. Cc: @norvid-studies.bsky.social, @tedunderwood.me

Especially excited about work like this because it means I can both eat my cake* and have it too**. *: Do a boring TCS PhD that isn't alignment relevant because it plays to my intellectual strengths and is self actualisation. **: Contribute usefully to technical AI safety.

LLMs learn to reason by training on maths and programming tasks. The reasoning ability so acquired generalised to other tasks. Hence training on maths/programming is a sufficient condition for learning reasoning. Implications for humans are left as an exercise for the reader.

Awards self a few Bayes points. Musk set it up to be le epic troll on Twitter. He didn't specifically make the model retarded. ML accounts claiming that the response must have been a result of fine tuning and condemning the model on that basis lose credibility.

Spending hours meditating* on a single line from the textbook until it makes sense.

Banger I found on the other site. Cc: @xeophon.bsky.social

💭 Someone should make a Manifold market on this.

Endorse. The problem with Bluesky is that its userbase is also fucking terrible, just in different directions. Like it is not actually the case that the people who refuse to move are just fatalistic about network effects. The Bluesky value proposition just isn't that good?

TOUGH LOVE THREAD: Please give me any constructive criticism or feedback that you think I need to hear. If you have something to say that you don’t want associated with you or you think might legitimately materially damage me for others to see you say, please DM, otherwise hmu

TOUGH LOVE THREAD: Please give me any constructive criticism or feedback that you think I need to hear. If you have something to say that you don’t want associated with you or you think might legitimately materially damage me for others to see you say, please DM, otherwise hmu

Happy Valentines day!

the debate began because one of them used "gooner" to mean "someone who plays genshin impact" her friends were like "nooo it's sexual!" but honestly the first girl might have been right

Shot Chaser

What is the practical solution to misinformed preferences? Most people mad about this are just fundamentally mistaken about how the platform/social media/etc. work. Their preferences is to a first approximation just rooted in ignorance. How should Bluesky handle this?

I mean this seems inherently hostile to like abstract mathematics. Logic, abstract algebra, set/model/type/category theory, etc. I don't see how you can build up abstract mathematics with Monte Carlo methods.

I need LLM recommendations.

Cc: @vgel.me

Things are terrible at the other place.