Share this page!

Enter URL or ID to Unroll

You can paste full URL like: https://x.com/threadreaderapp/status/1644127596119195649
or just the ID like: 1644127596119195649

How to get URL link on X (Twitter) App

  1. On the Twitter thread, click on or icon on the bottom
  2. Click again on or Share Via icon
  3. Click on Copy Link to Tweet
  4. Paste it above and click "Unroll Thread"!
  5. More info at Twitter Help
gavin leech is in NYC Profile picture
gavin leech is in NYC
@g_leech_
context maximiser @ArbResearch
Oct 3 • 9 tweets • 3 min read
I've been collecting examples of actually useful AI in research maths:

(I'm not disagreeing with Ege & Daniel's claim, which is carefully phrased as "[novel] concept") In 2023, Tao used Copilot for one line of the formalisation of Polynomial Freiman-Ruzsa.

"it offered a suggestion which was almost correct"

terrytao.wordpress.com/2023/11/18/for…
May 12 • 4 tweets • 1 min read
some ML talks on METR's task time horizon estimates (Rein et al 2025)

docs.google.com/presentation/d…
Feb 14 • 8 tweets • 3 min read
Stross (fictional): "Around the world, laboring women produce 45000 babies a day, representing 10^23 MIPS of processing power... fab lines casually churn out 30M microprocessors a day, 10^23... most of the MIPS added to the solar system will be machine-hosted for the first time." So when will the Accelerando moment actually be?

Plotting (OWID births * 1e15 brain anchor) we get a *very* approximate point estimate:

If there's 1.3e23 human FLOP/s added per year
and if the 2.3x/yr NVIDIA output scaling continues

Then NVIDIA crossover on their own by 2030. Image
Jan 12 • 7 tweets • 3 min read
blogs I found this year "What might become more valuable

Trustworthiness
Having a real audience
Doing things in real life
Craftspeople
Specializing
Being funny
Mental health
Agency and resourcefulness
Top ~10 percent creativity
Having good taste
Being adaptable
Kindness"

quarter--mile.com
Dec 23, 2024 • 27 tweets • 8 min read
essays of 2024 If you forced modern ML developers to work with 2016-era computers, they would not be able to make 2024-quality AI, even with all their knowledge of transformers and so on. You cannot make 4500 kilograms of soap from 450 grams of oil.

nostalgebraist.tumblr.com/post/741247180…
Dec 22, 2024 • 15 tweets • 4 min read
books of 2024

(i don't stay current) terrible year for my reading, the worst since I was like 8 years old. Something about writing up a thesis just drained all the text out of me.
Dec 8, 2024 • 23 tweets • 6 min read
albums of the year 2024 about 1000 new ones this year. (almost all got exactly one listen besides the following)
Dec 6, 2024 • 43 tweets • 11 min read
deaths of 2024 Daniel Dennett (1942–2024)

[illusionism; multiple drafts model; Intentional stance; evolutionary semantics]

> After the operation, they reversed the flow of blood to my brain, hoping to flush out any debris that was about to disable my res cogitans. So I’ve been brainwashed. Image
Dec 4, 2024 • 42 tweets • 13 min read
breakthroughs of 2024 Primary endosymbiosis found!

(Two discoveries in one; it's also the first eukaryote known to fix nitrogen)science.org/doi/10.1126/sc…
Mar 30, 2024 • 7 tweets • 2 min read
playlists 2024
Feb 8, 2024 • 18 tweets • 6 min read
New paper: a big 90-page intro to AI and its likely effects from ten perspectives, ten camps.

The whole gamut: ML, scientific applications, social applications, access, safety and alignment, economics, AI ethics, governance, and classical philosophy of life.

1/18 Image We inherited the framing (“Ten Hard Problems”) from Eric Schmidt and James Manyika. They conditionalise on success: "if it's 2050 and everything went well, what did we have to solve for that to happen?"

Intended audience: technical people without any ML experience. Image
Jan 5, 2024 • 45 tweets • 13 min read
Blogs I found this year On steps-to-reach-loss, FLOPs-to-reach-benchmark, inference speed, etc, the mixture-of-experts paradigm gets between 2x~5x improvement over a dense baseline. Fundamentally, MoE is unfriendly to low vRAM developers, which describes the vast majority of OSS
152334h.github.io
Dec 27, 2023 • 31 tweets • 11 min read
ML in 2023

(not a calibrated accounting of All Progress, just what caught my haphazard eye) This thread will of course overfocus on LLMs. At the start of the year, the stochastic parrot hypothesis was rampant. No longer.


thegradient.pub/othello/
Dec 3, 2023 • 27 tweets • 7 min read
Breakthroughs of 2023 First sample of a deep space object ever brought back to Earth. The great theft begins!

.archive.ph/TEzqH#selectio…
Dec 17, 2022 • 10 tweets • 3 min read
best playlists I made this year Foundtrack was an amazing blog that used to make mp3 playlists. It had a big effect on me. Some cuts are so deep that they're not on Spotify, Youtube, Soundcloud, Bandcamp...

open.spotify.com/playlist/1lUfr…
open.spotify.com/playlist/6k7BR…
Dec 16, 2022 • 19 tweets • 7 min read
notes on Mexico City Pleasant to just walk around central CDMX. Trees everywhere, dogs, masterpiece buildings - but razorwire, flaking paint and bare concrete too, to make you realise they are masterpieces (not complacent like I am in perfect environs, e.g. Venice)

Full of noise, trees, and art.
Dec 14, 2022 • 4 tweets • 2 min read
This philosopher test from last week is accidentally insightful. Most of the questions are about vibes / social presentation / misanthropy. Valid dimensions to think about philosophers with! But not very philosophical.

idrlabs.com/philosopher-pe… The test mostly just measures your condescension, drama, vibesmaxxing, misanthropy, nerdiness. Some important principal components of this demi-philosophy:

* Coolness ("murder poet" vs "bureaucrat")
* Condescension
* Sensitive Outsider
* Nerdiness (formalisation)
Dec 13, 2022 • 4 tweets • 2 min read
🧵 [Fiction]: Sylvie, a self-propagating harrassment bot with internet access who hounds privileged people with custom cruelty, sometimes to their deaths.

"his stalker had sent Harrison almost three hundred thousand messages over the course of six years"

clarkesworldmagazine.com/huang_12_22
May 31, 2022 • 13 tweets • 9 min read
How well do masks work against Covid at the population level?

In our new paper, we find mask-wearing corresponded to an international average 19% decrease in transmission (Rt).

(On average, 83% reported wearing some mask most of the time in public.)

doi.org/10.1073/pnas.2… Image The puzzle: studies of individuals found big reductions in Covid transmission from masks, ~50%. But society-level studies trying to predict Covid rates from masks found inconsistent results [-2%, 40%].
Jun 19, 2021 • 13 tweets • 7 min read
How well do masks work against COVID transmission, at the population level?

In our new preprint, we find that an entire population wearing masks is linked to a reduction in R of 24.6% [6%, 43%].

medrxiv.org/content/10.110… Image (In particular, this effect is that of an entire population reporting that they wore masks, in some or all public places, most or all of the time.)