LessWrong 2.0 Reader

Good work, thanks for doing this.

For future work, you might consider looking into inference suppliers like Hyperdimensional for DeepSeek models.

FWIW in my mind I was comparing this to things like Glen Weyl's Why I Am Not a Technocrat, and thought this was much better. (Related: Scott Alexander's response, Weyl's counter-response).

Well, I upvoted your comment, which I think adds important nuance. I will also edit my shortform to explicitly say to check your comment. Hopefully, the combination of the two is not too misleading. Please add more thoughts as they occur to you about how better to frame this.

Yeah it’s super-misleading that the post says:

Look at other unsolved problems:
- Goldbach: Can every even number of 1s be split into two prime clusters of 1s?
- Twin Primes: Are there infinite pairs of prime clusters of 1s separated by two 1s?
- Riemann: How are the prime clusters of 1s distributed?

For centuries, they resist. Why?

I think it would be much clearer to everyone if the OP said

Look at other unsolved problems:
- Goldbach: Can every even number of 1s be split into two prime clusters of 1s?
- Twin Primes: Are there infinite pairs of prime clusters of 1s separated by two 1s?
- Riemann: How are the prime clusters of 1s distributed?
- The claim that one odd number plus another odd number is always an even number: When we squash together two odd groups of 1s, do we get an even group of 1s?
- The claim that √2 is irrational: Can 1s be divided by 1s, and squared, to get 1+1?

For centuries, they resist. Why?

I request that Alister Munday please make that change. It would save readers a lot of time and confusion … because the readers would immediately know not to waste their time reading on …

Good questions. I don't have much of a guess about whether this is discernably "smarter" than Claude or Gemini would be in how it understands and integrates sources.

If anyone is game for creating an agentic research scaffold like that Thane describes, I'd love to help design it and/or to know about the results.

I very much agree with that limitation on Google's deep research. It only accepts a short request for the report, and it doesn't seem like it can (at least easily) get much more in-depth than the default short gloss. But that doesn't mean the model isn't capable of it.

Along those lines, Notebook LM has similar limitations on its summary podcasts, and I did tinker with that enough to get some more satisfying results. Using keywords like "for an expert audience already familiar with all of the terms and concepts in these sources" and "technical details" did definitely bend it in a more useful direction. There I felt I ran into limitations on the core intelligence/expertise of the system; it wasn't going to get important but fine distinctions in alignment research. Hearing its summaries was sort of helpful in that new phrasings and strange ideas (when prompted to "think creatively") can be a useful new-thought-generation aid.

We shall see whether o3 has more raw reasoning ability that it can apply to really doing expert-level research.

I don't think it's an outright meaningless comparison, but I think it's bad enough that it feels misleading or net-negative-for-discourse to describe it the way your comment did. Not sure how to unpack that feeling further.

Yeah, I just found a cerebras post which claims 2100 serial tokens/sec.

links 02/03/25:

• Jasmine Sun on the Tech Right
  • https://jasmi.news/p/tech-right
  • https://jasmi.news/p/arjun-ramani
  • I like this. she's not a theorizer! she's just Some Guy, actually expressing her thoughts on current events. do I agree with everything she says? maybe not.
  • But it's normal-ass blogging rather than inhibited silence or sloppy thoughts packaged as a Grand Narrative, and I think that's healthy. we need normal-ass blogging.
    • [[Scott Alexander]] is a normal-ass blogger who kept up a regular schedule and has a gift for puns and a fairly high appetite for books and research papers.
    • like, that's all it is, it's being yourself in public, consistently year over year, while having a healthy (but not extraordinary!) degree of interest in the world around you.
  • https://www.programmablemutter.com/p/why-did-silicon-valley-turn-right this is cited in her piece; i'm also not sure what i think of this, might be a piece of the puzzle
• https://meetmeoffline.com/ nice, Shreeda Segan's dating app is live
• https://lambda.chat/ nice hosting platform for multiple models including the DeepSeek ones.
  • note that the DeepSeek phone app is reputed to have a keylogger and location tracker; web apps are preferable.
• https://www.betonit.ai/p/trust_and_diverhtml Bryan Caplan close-reads the famous Robert Putman paper that purportedly shows that "more ethnic diversity leads to lower social trust."
  • It doesn't show that; it shows that black, Latino, and Asian people have lower "trust" than white people.
  • also correlated with lower trust: poor neighborhoods; high-crime neighborhoods; high-density neighborhoods; neighborhoods where most people moved in recently; neighborhoods with lots of renters; neighborhoods with few US citizens
  • also correlated with higher trust: individual income, bachelor's degree, homeownership
  • this just reduces to socioeconomic status. there's not a separate thing going on here about diversity.
    • obviously you can increase the average of many quality-of-life metrics in a community by restricting it to people of higher socioeconomic status. but then those same metrics would (mathematically) decline outside the elite community.
    • this does not support claims like "everybody on average would be better off under residential segregation".
• https://marginalrevolution.com/marginalrevolution/2025/02/sundry-observations-on-the-trump-tariffs.html  Tariffs harm economies, news at 11.
• https://www.complexsystemspodcast.com/episodes/the-landmines-buried-in-the-fine-print-of-chicagos-new-casino-deal/ Patrick McKenzie has details (lots of details) about the new casino being built in Chicago.
  • in order to meet requirements to be 25% woman-and-minority owned, they went around to black churches to offer an extremely misleading "investment deal" to working-class people who cannot afford it and won't understand the fine print.
    • the financial structure includes saddling these "investors" with a surprise enormous tax bill that only kicks in years after purchase.
  • of course, anybody can get in on special "women and minorities only" financial opportunities, even if they're a white man; set up a shell company "owned" by a woman and/or minority, who is your wife, or an associate of yours willing to serve as your front. this happens ALL THE TIME.
• https://en.m.wikipedia.org/wiki/Advanced_Research_and_Invention_Agency ARIA was founded in 2021.
  • A friend asked me a good question: what was the UK's DARPA-equivalent before?
    • UKRI https://en.m.wikipedia.org/wiki/UK_Research_and_Innovation was the older UK science-and-tech funding org, which ARIA is unaffiliated with; it brought together nine older funding bodies, most of which were founded in the 21st century, and none of which could plausibly be the bodies that funded the bulk of UK nationally-funded (non-medical) science and engineering through the mid-20th-century.
    • so...what was the UK's DARPA, or for that matter its NSF, in 1945-1990?
• https://trevorklee.substack.com/p/on-the-responsibility-of-size this is a nice, slightly oblique thought by Trevor Klee. do I agree? maybe. i don't expect to form considered opinions on All This until years later, if at all.
• things I learned while reading about Venice:
  • https://en.m.wikipedia.org/wiki/Sigismondo_Pandolfo_Malatesta amazing guy. mercenary. murdered two out of three wives. first person that the Pope explicitly "canonized into Hell." patron of Piero della Francesca and Leon Battista Alberti. rehabilitated in literature several times, including by Ezra Pound, which figures.
• https://www.palladiummag.com/2025/01/31/the-failed-strategy-of-artificial-intelligence-doomers/ critical take on the ineffectiveness of "AI safety" as a political strategy.
  • if you successfully convince the world that AI is potentially very powerful (and dangerous), this does not make people go "ok let's not build AI then", it makes people think "i want to be powerful and dangerous!!!"
  • i'm not on board with everything in this article but i think i largely agree.
  • when you're an idealistic nerd who despises playing politics and isn't very good at it, it will probably end badly if you dive enthusiastically into politics.
    • "so how can you do anything helpful at all?"
      • provide true information. don't optimize too aggressively for mass appeal.
        • this doesn't mean actively hide or be deliberately cryptic -- i think that's often going too far. i think clarity and open communication are, where possible, good practices.
        • but maybe don't make it your full-time job to strategically maximize the number of humans who believe a given thing.
      • focus your efforts on goals that you're quite confident will be beneficial and that can be done without coercion. beware of making your job about "what the government should do".
      • do not develop an identity around being an expert at strategic adversarial thinking.
        • you may be a literal chessmaster (like, at the game of chess);
        • you may be very skilled at things that would be useful to a modern military;

        • but if you are, in the colloquial sense, "kind of aspie", you are not an expert in detecting when somebody's about to screw you over. do not be over-eager to swim in shark-infested waters.

           

https://artificialanalysis.ai/leaderboards/providers claims that Cerebras achieves that OOM performance, for a single prompt, for 70B-parameter models. So nothing as smart as R1 is currently that fast, but some smart things come close.

agreed