iiinova — where your digital self keeps evolving

Miles Cranmer @milescranmer.bsky.social · 28/09/2025

grep -A/-B might be the worst flags in all of Unix. Is A=after or above? Is B=before or below? Even the alphabetical order is backwards! (A comes *AFTER* B according to grep) I've given up and just use `grep -A {N} -B {N}` for everything

120

Miles Cranmer @milescranmer.bsky.social · 24/09/2025

Does anybody know the actual page limit for ICLR submissions? The author guidelines page contradicts itself:

130

Miles Cranmer @milescranmer.bsky.social · 15/09/2025

Cambridge eventually decided it was a bit much, and, led by the mathematician G.H. Hardy (who, I should note, was Fourth Wrangler in his year...) reformed the system to make it less competitive. More history: www.maths.cam.ac.uk/about/history www.smithsonianmag.com/history/the-...

maths.cam.ac.uk

A History of Mathematics in Cambridge | About the Maths Faculty

About the Maths Faculty

010

Miles Cranmer @milescranmer.bsky.social · 15/09/2025

It was, as you might expect, quite competitive. C.T. Simpson (Second Wrangler, 1842) apparently took a stimulant (ether) to endure 20-hour study sessions. And James Savage studied so hard he was found dead of apoplexy a few months after being named Senior Wrangler in 1859.

100

Miles Cranmer @milescranmer.bsky.social · 15/09/2025

(Many famous names didn't make the list – J.J. Thomson famously sent his servant to ask who came second, only to learn it was himself.) Before 1909 reforms, the examination was a bit of a national spectacle, with substantial betting taking place on the outcome.

100

Miles Cranmer @milescranmer.bsky.social · 15/09/2025

Philippa Fawcett came first in 1890, a full 13% above second place, but since women were not yet awarded degrees until 1948, she was given the title "above the Senior Wrangler." Today, the Maths HPC cluster is named after her.

100

Miles Cranmer @milescranmer.bsky.social · 15/09/2025

Wikipedia has a list of all Senior Wranglers going back to 1748 (en.wikipedia.org/wiki/Senior_...). The list has some recognizable names in science – Herschel, Stokes, Rayleigh, Eddington, etc., along with some interesting ones, like Singapore PM Lee Hsien Loong!

en.wikipedia.org

Senior Wrangler - Wikipedia

100

Miles Cranmer @milescranmer.bsky.social · 15/09/2025

When Maths became the first subject to institute written examinations, it kept the name. The top-ranked students are (still) named "Wranglers," and the top scorer overall is named the "Senior Wrangler," then the "Second Wrangler," and so on.

100

Miles Cranmer @milescranmer.bsky.social · 15/09/2025

Cambridge Maths calls the top student in a given year the "Senior Wrangler." Why? Apparently this originates in Medieval examinations called "disputations," where students would literally wrangle (argue) in Latin with designated opponents.

Historical depiction of an academic disputation

180

Miles Cranmer @milescranmer.bsky.social · 04/09/2025

PolymathicAI is recruiting two postdoctoral researchers to join our team at Cambridge, to work on building and understanding large-scale foundation models for science. Please share with potential candidates!

261

Miles Cranmer @milescranmer.bsky.social · 02/08/2025

Cannot wait for someone to wrap SAP Concur in an MCP It’s kind of crazy you still can’t generate expense reports from a folder of receipts!

040

Miles Cranmer @milescranmer.bsky.social · 15/07/2025

These ones are the XREAL One. I quite like them though haven't tried others to compare. These have no 3D, no bluetooth, no battery, you just plug it in your laptop as an external monitor on your face. Works brilliantly for this, its like a hovering external monitor for your laptop

120

Miles Cranmer @milescranmer.bsky.social · 15/07/2025

That makes sense! I've had a few people ask for this and am considering it. For now, though, you could implement this constraint manually with the following custom loss function: gist.github.com/MilesCranmer.... Pass it as a string to `loss_function_expression`. (Let me know if this doesn't work!)

010

Miles Cranmer @milescranmer.bsky.social · 15/07/2025

Really loving AR glasses for working while traveling. No squinting at my laptop + I can still enjoy the view! These just plug into my closed laptop as an external monitor. Means I'm actually productive on planes now because I'm not hurting my neck looking down the whole flight!

160

Miles Cranmer @milescranmer.bsky.social · 14/07/2025

By the way, just curious, is that `complexity_of_variables` supposed to be `complexity_of_constants`? I only ask because usually in the pysr forums people are looking to minimize the # of constants rather than # of variables

110

Miles Cranmer @milescranmer.bsky.social · 14/07/2025

This looks pretty good to me! Do try out parallelism=:multiprocessing and let me know if that helps. For whatever reason its better at getting better CPU utilization, though it takes longer to start up every time you launch a new search. (Likely small, but could try larger `population_size`)

110

Miles Cranmer @milescranmer.bsky.social · 14/07/2025

Tomorrow I get to give my favourite type of lecture to some new physics PhD students 😃

0140

Miles Cranmer @milescranmer.bsky.social · 13/07/2025

I haven’t done a detailed study myself but it would be cool to see! I can try to ask them what they settled on. Interesting! Do you remember what hyperparameters were used? It might be there weren’t enough populations to use all cores? (Also you could try parallelism=:multiprocessing)

120

Miles Cranmer @milescranmer.bsky.social · 08/07/2025

I am a HUGE fan of Mooncake.jl github.com/chalk-lab/Mooncake.jl, a new language-level autodiff for Julia. Most impressive is how uncompromising the devs are about correctness and reliability. It has some of the best software engineering practices I've seen in the entire ecosystem!

060

Miles Cranmer @milescranmer.bsky.social · 21/05/2025

Wow. Apparently some PySR users are building dedicated rigs for faster equation discovery!

270

Miles Cranmer @milescranmer.bsky.social · 15/05/2025

I just referred to the literal letter 'z' as 'redshift' 😵‍💫

030

Miles Cranmer @milescranmer.bsky.social · 02/05/2025

SymbolicRegression.jl v1.10.0 is out! It can now evolve expressions over arbitrary input types. The video below shows it reverse-engineering a string transformation from examples. Curious to see how people use this!

0120

Miles Cranmer @milescranmer.bsky.social · 08/04/2025

You have no idea how long I’ve waited for this 😮‍💨

181

Miles Cranmer @milescranmer.bsky.social · 02/04/2025

We have just opened a fully-funded PhD position at Cambridge, supervised by me & Vasily Belokurov. Topic: AI + astronomical imaging (broadly defined). Deadline: April 16. Please share with anyone who may be interested! www.postgraduate.study.cam.ac.uk/courses/dire...

0164

Miles Cranmer @milescranmer.bsky.social · 25/03/2025

Very excited to share that my group in Cambridge is offering an “AI for Science” Encode fellowship on multiscale physics through Pillar VC & ARIA! Apply by April 30th: encode.pillar.vc/projects/mul...

0130

Miles Cranmer @milescranmer.bsky.social · 07/03/2025

Colab tends to update infrequently, so I would really really recommend adding a compatibility layer. I think 99% of users who try it out in colab will not be aware of that package constraint, so will end up using old Tidier.jl by mistake

150

Miles Cranmer @milescranmer.bsky.social · 06/03/2025

This might be the greatest bug report I’ve ever seen

1284

Miles Cranmer @milescranmer.bsky.social · 05/03/2025

Wow, Google Colab just added Julia support!!! 🎉

27113

Miles Cranmer @milescranmer.bsky.social · 26/02/2025

Why 'I don’t know' is the true test for AGI—it’s a strictly harder problem than text generation! This magnificent 62-page paper (arxiv.org/abs/2408.02357) formally proves AGI hallucinations are inevitable, with 50 pages (!!) of supplementary proofs.

0121

Miles Cranmer @milescranmer.bsky.social · 25/02/2025

Totally! Two ideas: 1. Evolve a hierarchy of expressions, using templates: ai.damtp.cam.ac.uk/pysr/example.... e.g., one expression for the latent, another for the observed (takes the latent as an input!) 2. Approximate an existing model using similar ideas to github.com/MilesCranmer...

ai.damtp.cam.ac.uk

Toy Examples with Code - PySR

110

Miles Cranmer @milescranmer.bsky.social · 25/02/2025

Symbolic regression update—new versions of PySR & SymbolicRegression.jl just released with improvements to mini-batching. Should work much better for large datasets!

180

Miles Cranmer @milescranmer.bsky.social · 24/02/2025

Having used it over the past week, I do think that ChatGPT deep research mode is worth the $200 price tag—something that, for me, o1-pro hadn’t justified. It’s not perfect, but for specific tasks, it is **extremely** useful.

020

Miles Cranmer @milescranmer.bsky.social · 16/02/2025

This usually ends up being due to floating point error, like how (0.1+0.2)-0.3 != 0.1+(0.2-0.3) (Though there may be some kind of convergence issues at play too. Maybe tuning is needed in the constant optimizer)

110

Miles Cranmer @milescranmer.bsky.social · 16/02/2025

Isn’t it a random number between -1000 and 1000, if I’m reading it right

100

Miles Cranmer @milescranmer.bsky.social · 15/02/2025

I think using absolute error instead of squared error would be better too, considering the massive scale of the output data. (can do this with elementwise_loss="(pred, targ) -> abs(pred - target)")

110

Miles Cranmer @milescranmer.bsky.social · 15/02/2025

Note: exp(-1000/-20) will be very large, so I would recommend using precision=64 for 64-bit calculations. With this change and many more iterations, I was able to get the result (after simplification). This type of problem has a very steep loss surface, so just need to run for longer.

110

Miles Cranmer @milescranmer.bsky.social · 15/02/2025

`f(x1)` can be an arbitrary expression, including having its own constants. So there are degenerate solutions for p1[1] here. Although I guess in principle it should prefer the simplest.

010

Miles Cranmer @milescranmer.bsky.social · 15/02/2025

Took me until today to figure out how to separate tracked/untracked files in VSCode. So much better!

1100

Miles Cranmer @milescranmer.bsky.social · 14/02/2025

I don't think one could compare constants this way, because it would be degenerate with the constants inside the expressions. (If I understand the question correctly?) Let me know if you try though, I'd be very eager to hear the results!

100

Miles Cranmer @milescranmer.bsky.social · 14/02/2025

Yes, that works. You can even do f(f(x))

110

Miles Cranmer @milescranmer.bsky.social · 14/02/2025

New feature! PySR v1.4 lets you define a template expression to optimize that both has learnable parameters AND learnable expressions:

192

Miles Cranmer @milescranmer.bsky.social · 12/02/2025

I love Cursor, but because it uses a slightly out-of-date VSCode, the extensions can have known vulnerabilities! I did some digging and noticed that the main Python extension has a *known* remote code execution vulnerability (CVE-2024-49050), and has had it for three months!

080

Miles Cranmer @milescranmer.bsky.social · 11/02/2025

I’ll be giving lectures on advanced ML techniques at this AI-for-HEP summer school in Lake Como, Italy, from July 14-18. Expect a deep dive into symbolic AI and beyond. Should be a great event! Students can apply here: buff.ly/3WTTE8J

1110

Miles Cranmer @milescranmer.bsky.social · 04/02/2025

Surprisingly many people have asked for this kind of constraint! Specifically: constraining different terms to only depend on a certain subset of features

030

Miles Cranmer @milescranmer.bsky.social · 04/02/2025

Interested in feedback on this syntax idea for predefined "template expressions" in SymbolicRegression.jl and PySR. Basically: what is an intuitive way to prescribe a fixed skeleton for a symbolic search? Ideally the syntax should be flexible and extensible too.

160

Miles Cranmer @milescranmer.bsky.social · 30/01/2025

Exciting news – Cambridge is hiring an Assistant/Associate Professor in Data-Driven Radio Astro. Join us!

030

Miles Cranmer @milescranmer.bsky.social · 22/01/2025

It's crazy how at one point I was worried whether I would like teaching. But now, with the new term starting tomorrow, I can't wait.

0170

Miles Cranmer @milescranmer.bsky.social · 19/01/2025

- print statements ❌ - debugger ❌ - rebuilding rust's entire ownership system in julia because one thread race made you lose it ✅ github.com/MilesCranmer/BorrowChecker.jl

github.com

GitHub - MilesCranmer/BorrowChecker.jl: A borrow checker for Julia

A borrow checker for Julia. Contribute to MilesCranmer/BorrowChecker.jl development by creating an account on GitHub.

2141

Miles Cranmer @milescranmer.bsky.social · 16/01/2025

And in the case of GDM they tend to look _very_ broadly for problems. Whereas I suppose the more traditional route is to work backwards from a problem (or small set of problems) and find a tool

050