Lightnews — Scholar-powered news

Mark J. Nelson @mm-jj-nn.bsky.social · 10h

Kind of interesting that some of this is so recent. I would've guessed the combinatorial structure of lambda terms was known much earlier than the 2010s, but it seems it wasn't really.

2

Mark J. Nelson @mm-jj-nn.bsky.social · 10h

Enumerated the terms using the algorithm from this paper:

Counting and generating lambda terms | Journal of Functional Programming | Cambridge Core

Counting and generating lambda terms - Volume 23 Issue 5

doi.org

1 2

Mark J. Nelson @mm-jj-nn.bsky.social · 10h

Made myself a cool list of which lambda calculus terms have a normal form

(λa.a) ((λa.a) (λa.a)) reduces to λa.a in 2 steps of normal-order reduction
(λa.λb.b) (λa.λb.b) reduces to λa.a in 1 steps of normal-order reduction
(λa.λb.b) (λa.λb.a) reduces to λa.a in 1 steps of normal-order reduction
(λa.λb.b) (λa.a a) reduces to λa.a in 1 steps of normal-order reduction
(λa.λb.a) (λa.λb.b) reduces to λa.λb.λc.c in 1 steps of normal-order reduction
(λa.λb.a) (λa.λb.a) reduces to λa.λb.λc.b in 1 steps of normal-order reduction
(λa.λb.a) (λa.a a) reduces to λa.λb.b b in 1 steps of normal-order reduction
(λa.a a) (λa.λb.b) reduces to λa.a in 2 steps of normal-order reduction
(λa.a a) (λa.λb.a) reduces to λa.λb.λc.b in 2 steps of normal-order reduction
(λa.a a) (λa.a a) loops

1 6

Mark J. Nelson @mm-jj-nn.bsky.social · 11h

Thought experiment got it wrong on paperclips! Obvious in retrospect; why would we want all those paperclips?

1

Reposted by Mark J. Nelson

Tomas Petricek @tomasp.net · 18h

I'm teaching 𝗪𝗿𝗶𝘁𝗲 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝘁𝗶𝗻𝘆 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴 𝘀𝘆𝘀𝘁𝗲𝗺(𝘀)! again. I'll be posting the videos & tasks on YouTube too.

In the first lecture, I explain what's a tiny system, why write one and show plenty of demos!

🎞️ Playlist: www.youtube.com/playlist?lis...
👉 More info: d3s.mff.cuni.cz/teaching/npr...

Write your own tiny programming system(s)! - YouTube

The goal of this course is to teach how fundamental programming language techniques, algorithms and systems work by writing their miniature versions. The cou...

www.youtube.com

5 22

Mark J. Nelson @mm-jj-nn.bsky.social · 13h

Definitely thought this was an AI thesis from the title and opening (M. Alauddin, University of Lancaster, 1976). But it's about error-correcting codes and reliability estimates (of which "doubt is the inverse").

THE INTRODUCTION OF SELF-DOUBT
INTO COMPUTER DESIGN

ABSTRACT

This thesis investigates the question whether a computer could be built which could gradually develop a doubt as to whether it was on the right track – this being a feature of natural intelligence that the current generation of computers does not show.

2

Mark J. Nelson @mm-jj-nn.bsky.social · 14h

To be fair you're mostly looking at AI (I think?) which is at one end of the spectrum. CS also has our sickos who write papers like this (61 pages).

Choice trees: Representing and reasoning about nondeterministic, recursive, and impure programs in Rocq | Journal of Functional Programming | Cambridge Core

Choice trees: Representing and reasoning about nondeterministic, recursive, and impure programs in Rocq - Volume 35

www.cambridge.org

2 3

Reposted by Mark J. Nelson

Ramon Astudillo @ramon-astudillo.bsky.social · 19h

(repost welcome) The Generative Model Alignment team at IBM Research is looking for next summer interns! Two candidates for two topics

🍰Reinforcement Learning environments for LLMs

🐎Speculative and non-auto regressive generation for LLMs

interested/curious? DM or email [email protected]

1 13 17

Mark J. Nelson @mm-jj-nn.bsky.social · 16h

Also good regulatory arbitrage. I'm prohibited from having a space heater in my office, but there's no rule against loading up that tower with GPUs.

3

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning

Fine-tuning pre-trained large language models (LLMs) for down-stream tasks is a critical step in the AI deployment pipeline. Reinforcement learning (RL) is arguably the most prominent fine-tuning meth...

arxiv.org

1 12

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

One that I've run across more than once now is people citing Benoit Mandelbrot's famous 1982 book as "Mandelbrot & Mandelbrot, 1982", I believe because Google Scholar had it wrong for a period of time (it was written by just the one Mandelbrot).

1 5

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

An interesting pre-LLM (but probably also post-LLM) variant is when articles *do* exist, but not in the form cited, e.g. wrong year, journal, publisher, author list, etc. Someone got it wrong once and it propagated.

1 1 6

Reposted by Mark J. Nelson

Matt Keeter @mattkeeter.com · 1d

finally, some Good Code

crates.io/crates/fibon...

Screenshot of the fibonacci-numbers README, which reads:

First install the version of this crate corresponding to the Fibonacci number you want to use:

cargo add fibonacci-numbers@186 --rename fib186

Then you can use the VALUE constant to get its value:

assert_eq!(fib186::VALUE, 332825110087067562321196029789634457848);

Screenshot of the dependency list, which is fibonacci-numbers at versions 184 and 185.

5 22 130

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

You could reasonably ask why the university requires this format in the first place. I guess they have reasons for wanting each part of the file in one big PDF, instead of multiple small PDFs, b/c that's more convenient for review at later stages. Either way that's the Provost office's decision.

2 1

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

To be clear there's no LLM summarization going on here; the LLM is not even allowed to see the PDFs. All the LLM is doing is writing code (which I run, after reviewing it) to merge six PDFs into one PDF, in a specific format required by the university (which most depts do manually in Acrobat Pro).

1 2

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

Apparently I wrote a blog post about stamping things onto PDFs in 2018. At the time my solution involved pdftk, but I switched to LaTeX sometime after that as a more flexible and self-contained solution. Starting to think that academia is really about PDFs.

Stamping copyright information onto academic papers

www.kmjn.org

5

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

(It feels a bit like what it first felt like to code for these kinds of tasks. Step one of office enlightenment is to ascend from futzing around in Adobe Acrobat Pro, to writing code to do it. Step two is to ascend from futzing around in CTAN packages' manuals, to having an LLM do that.)

1 7

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

(Admittedly knowing that you can do this with LaTeX is maybe 90% of the problem-solving, and once that far towards a solution, I could have just written the LaTeX myself. But it would have taken 30 mins on StackOverflow to do, instead of ~5 mins.)

2 10

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

I had to concatenate all the external letters & the letter-writers' CVs into two PDFs, stamped at the top/right of EVERY page with "Letter N" indicating which one that page was from. Apparently people do this manually! I got gemini-cli to write me some LaTeX to do it (it took it three tries). 2/2

1 10

Mark J. Nelson @mm-jj-nn.bsky.social · 1d

Successfully used an LLM for the first time to speed up a paperwork/admin task (although just by turning it into a coding task). I'm co-chair of the rank & tenure committee this year, which involves pushing a lot of PDFs around, and some of the PDF-pushing is pretty annoying. 1/2

1 1 32

Mark J. Nelson @mm-jj-nn.bsky.social · 2d

It is! It is weirdly humbling, b/c the built-in hardcoded AI that has like 200 clock cycles to "think" regularly beats me, a rube unacquainted with the 3rd dimension of tic-tac-toe victories.

1

Mark J. Nelson @mm-jj-nn.bsky.social · 2d

a screenshot from the Atari 2600 game "3-D Tic-Tac-Toe"

1 2

Reposted by Mark J. Nelson

thebes @vgel.me · 3d

new blog post! why do LLMs freak out over the seahorse emoji? i put llama-3.3-70b through its paces with the logit lens to find out, and explain what the logit lens (everyone's favorite underrated interpretability tool) is in the process.

link in reply!

8 46 210

Mark J. Nelson @mm-jj-nn.bsky.social · 3d

I used to use them in DC, and in theory liked it, esp. being able to do one-way rentals and street park anywhere. The app is buggy though, too frequently had trouble unlocking the car, ending trips, etc.

1

Mark J. Nelson @mm-jj-nn.bsky.social · 3d

when an RL agent forgets previous skills after learning new ones, that's Bell-Man amnesia

2 15