Michael Nielsen Profile picture
Searching for the numinous 🇦🇺 🇨🇦, home in 🇺🇸 Research @AsteraInstitute https://t.co/maezekzRUb
Potato Of Reason Profile picture Ryan Sapieha Profile picture 3 subscribed
May 29, 2023 4 tweets 2 min read
Striking fact learned the other night: the singlet-triplet measurement is (apparently) BQP-complete. Anyone have a pointer for a proof?

(I find this hard to credit. If true it's just amazing.
But I'll bet the proof fits in a tweet or two.) Oh: arxiv.org/pdf/2105.04649…

I wonder if further progress has been made? Image
May 29, 2023 4 tweets 1 min read
In what senses (if any) has a neural net been used to learn an operating system from scratch? Or the design of a programming language?
May 28, 2023 6 tweets 2 min read
Instinctively, my answer is no. I'd expect the Lyapunov exponents in the system to be zero - the degrees of freedom are rotation (no exp divergence, ignoring friction) & center of mass (ditto), so there's nothing to amplify quantum fluctuations, even in the absence of decoherence That said, I'm not sure, and would need to reflect further

Fun problem; it'd make a nice essay question for a class on quantum chaos!
May 15, 2023 28 tweets 7 min read
Claude's Principles: anthropic.com/index/claudes-… There's a lot of ways one could engage with this (and reasons one might be wary). Here I will mostly adopt one frame, which is thinking from a personal pov about these principles, applying the principle of charity
May 15, 2023 35 tweets 9 min read
Ted Chiang: "Will AI become the new McKinsey?" archive.is/tSpYV Pedantically, we can also try to reason directly about it, we don't need to rely on metaphor. But the point about how slippery metaphor is is a good one

Related: the very tricky relationship between "narratively plausible" and "potentially true" Image
May 14, 2023 4 tweets 2 min read
Amazing graph: I don't know the original source of the data, or how reliable it is. It's this site: wid.world

I'd love a reliable source for tracking inequality!
May 13, 2023 4 tweets 1 min read
Imagining competitive meditation as a sport: And trying to imagine what it would be like to commentate

"Oh, his eye seems to be twitching a little there! Some intrusive thoughts, I'm guessing Bob! Let's just check in on the ECG... [etc]"
May 13, 2023 5 tweets 1 min read
Nope, only for everyone to lose: fortune.com/2023/05/09/pet… Image Incidentally, I see reports online that Palantir's offering is based on GPT4. Can anyone confirm?
May 11, 2023 4 tweets 1 min read
What’s the most wonderful textbook you’ve ever read? A few for me:

+ The Feynman Lectures on Physics
+ Sean Carroll’s book on general relativity
+ David Griffiths on electromagnetism
+ The Molecular Biology of the Cell
May 11, 2023 4 tweets 1 min read
Curious: if you're a professional biologist who has extensively used "Molecular Biology of the Cell", how long would you expect it to take someone to work through it? (I've been enjoying reading it enormously. But also finding myself thinking that some of the chapters would make good semester-long, or at least term-long, classes, if you really got into them.)
May 11, 2023 4 tweets 1 min read
Titin - a single protein! - must be just about visible to the human eye!

Reminded a bit of learning that we can very nearly (not quite IIRC) see single photons Image I've been in a few quantum optics labs where they do single-photon detection at optical frequencies. It's quite peculiar: just utterly black. And jokes from the occupants about their pasty complexions
May 11, 2023 6 tweets 1 min read
Curious: have experiments been done building amino acid chains at random to see what fraction reliably fold into stable shapes? One lesson from information theory: trying stuff at random is often shockingly informative!
May 9, 2023 6 tweets 1 min read
Transformers are such a strange thing. It's a bit like "I'm going to reflect a *lot* on my words [tokens]". Each pass I will enrich my understanding of each word a little. But I'll just keep doing this. Then I'll make a decision! I'm tempted to say I'm oversimplifying. Except... really not much!
May 9, 2023 4 tweets 1 min read
Enjoying this! I'm going to give a short talk tomorrow about the (very near term) impact of AI on science, specifically AlphaFold
May 9, 2023 5 tweets 1 min read
I keep seeing people repeat that LLMs are "just repeating statistical patterns in text". In some sense it's true. But AFAIK none of those people predicted transfer learning from predict-the-next-token would give SOTA on so many NLP tasks It suggests they don't grasp what "statistical patterns in text" means. It's a bit like saying the NBA playoffs are "merely the motion of protons, neutrons, and electrons". True, but misses the point
May 1, 2023 23 tweets 4 min read
A naive question: I'm trying to understand in what sense AlphaFold 2's protein structure predictions are *verifiable*

A nice thing AF2 does is predict not just structure, but of how confident it is in that structure. Eg, blue in regions of high confidence, red in lower Image My understanding is that there are often regions of intrinsic disorder in the protein - places where there may not be a unique fold - and AF2's low-confidence predictions are often useful in identifying such regions
May 1, 2023 4 tweets 2 min read
Geoff Hinton quits Google so he can talk freely about AI risks. Says a part of him now regrets his life's work Image And a correction:

Doesn't change the high-order bit, of course
May 1, 2023 34 tweets 5 min read
Reflecting on AlphaFold 2. It's a strange system. It's not an explanation of protein structure, nor a theory of protein structure

And yet... there is a sense in which it (& successor systems) capture much (not all) of our best understanding of protein structure It's tempting to think it's just another computational model: (say) a climate model or weather model or a model of explosion

But it seems somehow different
Apr 29, 2023 5 tweets 1 min read
Something I'd like to read: a thoughtful & in-depth comparison between modern complaints about regulation, nanny states etc, and mistakes and near-misses in the history of technology (lead, radium, CFCs, thalidomide, etc) Preferably one that engages deeply with opportunity costs and unforseen good consequences, as well as bad

I.e., I'm not really looking for something with an axe to grind, unless the axe is "let's understand the history of tech and regulation and progress and safety"
Apr 29, 2023 6 tweets 2 min read
Love this kind of map!

Unfortunately, the lovely Alexander's in downtown SF closed yesterday ☹️ Image I think our society mostly gets better. But I regret the gradual loss of bookstores!

In the late 90s I lived in New Mexico. So many wonderful bookstores! In 2014 I toured the state with my then-wife, and was excited to show her my fave bookstores. IIRC 11-of-12 were gone! ☹️
Apr 29, 2023 4 tweets 1 min read
Pleased to see this - a bipartisan bill to compel human sign-off on nuclear launches. (Deeply imperfect, but a good start.)

Especially pleased it's bipartisan. There will be a strong push to make AI impacts a partisan issue. I hope people will resist that as much as possible "We did terrible damage to our civilization, but at least we won a couple of elections" seems like a poor strategy. But a lot of people are going to enthusiastically follow it