Rui Shu

May the prompt be with you

@NeelNanda5 "you get paid 2.5x if you agree to donate 0.8"

I worded this badly: "paid 2.5X more" is via donation matching. It is a bad deal from a selfish perspective, and legally binding Eg you either: - Get $400K equity yourself - Or get $200K yourself, and donate $800K to charities of your choice (200 from you, 600 matched from Ant)

should've named it VerTeX

@distributionat Kinda cool bc "試" means "try". The unembedding for "試" and "try" must be quite close, which is representationally/semantically desirable, but awkward for generation.

i just didn't expect Claude to be chinese

@ptrschmdtnlsn truly the shit-eating market maker

Every time this comes up, a bunch of people miss the point. Each one clearly values watching someone eat shit at >$100, and disvalues eating shit at <$100. This is two people with (utterly bizarre) preferences, getting gains from trade that they wouldn't have gotten were it not for dollars as an intermediary. They both *want* to trade eating shit for watching someone eat shit, but needed dollars to coordinate that positive-sum trade across time! Had they come across the piles simultaneously, they could have negotiated eating simultaneously, but they can't. This is *the* huge advantage of money, as a universal IOU; it enables "virtual trades" between parties separated in time and space.

@ChuckBaggett @_beenkim I'm actually quite upset that there's not a <disable-link-clicking-unless-I-really-know-what-I'm-doing> functionality in gmail. That added bit of friction is a good forcing function for staying vigilant---especially when we're super distracted.

@_beenkim If we really shouldn't click links in emails, there should be a way to disable active links from appearing in emails. I'm glad you got your account back.

I got my account back! Thank you, first and foremost, to everyone—friends, GDM colleagues---who personally alerted me to this incident and retweeted that I'm hacked, as well as folks at X who helped me regain access. While this incident was terrible (I heard the scammers made huge money out of this), I feel incredibly lucky to have folks who cared♥️♥️♥️ (details of how this happened 👇)

@BenjaminDEKR neural nets can dream. so we should keep up by aggressively sleeping and dreaming more.

Igor Babuschkin left xAI (he reported directly to Elon, built Grok 3 & 4) and shortly after, posted about the dangers of companies over-working their best people. Greg Yang (another direct Elon report, cofounded xAI) just left for health reasons and posted: "Until I pushed myself hard building xAI and weakened my immune system, the symptoms weren't noticeable." I will not presume to speak for either of them. Both love their work. However: We need to take a serious look at cultures of overwork and what this does to people. When you create a corporate environment where employees COMPETE to be the most exhausted, the most sleep-deprived, this takes a toll. (This is exactly what many xAI workers have done, some posting about how they're so tired they can barely stay awake to drive home.) If we truly are approaching a tipping point where AI eclipses human work, trying to keep up by never sleeping and working beyond exhaustion is not a winning strategy. The only way forward will be to play to our human strengths: creativity, intuition, and even literally dreaming. When top employees are putting themselves in the hospital from exhaustion, something is wrong.

@prajdabre > "I'll get some problems labeled and fine tune the model" is a perfectly good starting answer and a good research practice. Do the dumb simple thing first!

I'm a research scientist in Google and "I'll get some problems labeled and fine tune the model" is a perfectly good starting answer. The interview does not automatically end there when you suggest getting fine tuning data as the first sentence. Google interviews are not some gladiator trials, the interviewers are nice and actually prompt and encourage you to unpack your answers and elaborate because they know that people can be nervous. Btw verifiable responses or not, you still need golden responses for things like maths so getting labeled problems is still a legit starting answer. In my RS interviews I did not walk in with optimized responses, I went and explained what I knew and had a discussion (and not an interrogation) so please focus on the core concepts (highlighted in the responses to the post).

@ForrestPKnight What I'd give to hear Ben Affleck say "typical set" instead of "mean"

Honestly, Ben Affleck actually knowing AI and the landscape caught me off guard, but as a writer, makes sense. Great takes across the board.

@joseph_h_garvin behold: neurosymbolic ai

"Wow, GPT 5.2 Pro seems to be doing much better in this Connect4 game than GPT 5.2 Thinking did, I wonder why?" *checks thought trace* "Oh mother-"

@docmilanfar me using measure theory: "assume the pdf exists"

yet another year gone by and I've not used measure theory

@nlpnyc @ilex_ulmus @ESYudkowsky I doubt it's coordinated. But it's good information hygiene to always be mindful of survivorship bias c:

@ilex_ulmus I don't know to what degree this is coordination, but as @ESYudkowsky put (IIRC) it safety researchers at OAI were preselected to have positions consistent with their pre-existing policy. (This does suggest less personal dishonesty)

You’re not just talking to OpenAI and Anthropic staff members like they’re your friends on Twitter. They are part of coordinated comms efforts, with the interests of the companies in mind first. If they want to be taken at their word, they can give up those paychecks.

@eW8fkgAM52GS9xW @dwarkesh_sp I'm concerned about power inequality in a finite-resource world. While I may or may not care how wealthy someone else is (good for you btw!), I certainly do care how powerful people influence the world I live in. Thus my comment about agency and wanting my actions to matter c:

@_smileyball @dwarkesh_sp I really just don't think I'd care. Even today if I had a nice house and good food and the ability to travel and spend time with family, I wouldn't care all that much about how wealthy Elon Musk is. I'm already less miserable than he is, what more could I need?

I’ve seen a lot of people misunderstand what we’re saying. Our claim is that in a world of full automation, inequality will skyrocket (in favor of capital holders). People aren't thinking about the galaxies. The relative wealth differences in a thousand years—or a million—will be downstream of who owns the first dyson swarms and space ships. And space colonization isn't bottlenecked by people’s preference for human nannies and waiters. So even if you can make 10 million dollars a year as a nanny in the post-abundance future, or get a 10 million dollar charity handout, Larry Page’s million cyborg heirs can own a galaxy each. You might think this is fine! Why is inequality intrinsically bad, especially if absolute prosperity for everyone goes up? Fair enough, but to me quadrillion fold differences in wealth between humans seem hard to justify in a world where AIs are doing all the work anyways - these disparities in wealth are not incentivizing hard work or entrepreneurship or creativity, which is what we use to justify inequality today. Just to recap, full automation kills the corrective mechanism on runaway capital accumulation - which is that you need labor to actually make productive use of your capital, thus driving up wages. Some people asked: why assume AGI leads to full automation? Maybe people will still prefer human nannies and waiters. Even if true, we think labor's share of GDP—which has been roughly 2/3 for centuries—would still likely collapse toward zero, massively increasing inequality. Here's why. It sometimes happens that when machines are only slightly better than humans, people sometimes pay a premium for the human version. But once machines become much better, that preference disappears. When carriages were not much faster than being carried on a litter, the rich sometimes preferred the litter. Now they prefer the car. They might still have a chauffeur—but once self-driving vehicles are allowed to move far faster, human-driven cars may be relegated to a slow lane. If the economy grows 100x, wages must also grow 100x for labor's share to stay at 2/3. But prices are relative—so this means human labor becomes 100x more expensive compared to AI-produced goods. A human-cooked meal costs 100x what the robot version does. For labor share to hold steady as that ratio grows to 1,000x, then 10,000x, the preference for human-made goods would have to become increasingly fanatical. And there's a second problem: the higher wages rise, the greater the incentive to develop machine substitutes for whatever services humans still provide. The premium on human labor is precisely what incentivizes its own replacement. Just to clarify a few other things: - “Piketty’s long run series are disputed.” We spend a long chunk of the essay explaining why Piketty is wrong about the past! But we’re arguing that the assumption he makes (specifically that labor and capital are substitutes) would be true of a world with advanced enough automation. We spend so much time rebutting his claims about the past because the wronger you think he was about the past, the more you think will change once his assumption comes true. - “A capital tax would lower growth.” Yes, as we point out, capital taxes incentivize consumption now instead of saving and investing for the future, at the margin. But if capital is the only factor of production, then it’s hard to come up with an inequality-capping tax that doesn’t lower growth. - “Capital can escape, both across time and space. This makes a wealth tax impractical.” We agree! As we say in the essay and in the tweet summary below, it would be really hard to implement Pikkety’s flagship solution (a high and progressive global wealth tax). You could go Georgist and try to tax land, but the natural resource share of income is only 5% and is likely to stay low until we hit “technological maturity” for reasons we explain in the essay. We don’t see any easy ways to avoid (literally) skyrocketing inequality - in fact, that’s what inspired us to write the essay and explain this problem in the first place. Also, to address a subtext: I think the currently proposed California wealth tax is a very bad idea for many reasons. This essay is about inequality under full automation, not about how California can make its healthcare expenditures more sustainable.

@wavage_ the analogy between GLP1 and LLM isn't perfect, but I'm in full support of deliberately overcoming <physically/cognitively painful problems> as a means of character-building and giving us, as Asimov puts it, a feeling of power c:

@docmilanfar only if you're lucky enough to be stubborn about the right things 🙃

The stubbornness that makes you hard to manage as a young scientist is the same trait they call “unwavering dedication” when they give you a lifetime achievement award.

@andrewgwils NFL is useful! It helps me sleep at night bc it's how I convince myself that finite capacity NN + shitty SGD optimization is a feature, not a bug c: It's a great reminder that the universal function approximator narrative that was popular a ~decade ago is pure bait >:c

This is an annual reminder that the no free lunch theorems are irrelevant. The assumptions they make are completely divorced from the world we live in. They should have no bearing on model construction. Let's make this a monthly mantra.

@max_spero_ wow it's almost as if we trained the models via distribution matching

You're not gonna believe this

@willdepue @dwarkesh_sp @tszzl I think LLMs have no "taste" for what is a deep vs superficial insight. Imo a deep insight helps me decide next steps, etc. Talking to colleagues leads to more of such insights. But LLMs are great for translating my colleagues' deep insights into something my brain groks c:

not sure if just my experience but feel like the models still are really bad at specializing educational information based on your background. if there’s prompts you like lmk but often there’s no line between ELI5s and a mile of proofs and equations for complex topics. getting an intuition for a new field at a deep level is still difficult

even minor aristocrats could not get the kind of tutor on demand for their kids that every single family who can pay twenty dollars a month can now

why does she look like she’s gonna ask a historian if king arthur came a lot

@johnschulman2 Humans that are too jagged, unfortunately, end up not integrating effectively into many organizations.

Humans are jagged, and organizations (from companies to civilizations) have evolved as harnesses to make best use of us despite our faults