HNHacker News
TopNewBestAskShowJobs

LolWolf

1,481 karma · joined September 16, 2012

In a previous life, I did math and physics at Stanford.

---

https://angeris.github.io

submissionscomments
LolWolf··on The likely cheapest home-made Michelson interferometer
Hopefully you enjoyed the post then!

I think there's just such a huge middle ground that's missing (for funny historical reasons[1]) between "children's toy" and "lab-grade equipment" especially in optics, which is why I was excited to make this my first foray into making a fully 3d printed "useful-ish" thing that doesn't really exist otherwise.

---

[1] This is because most lab equipment was made _in the lab_ back in the 60s or so, and having this technical ability was a huge advantage for many labs. Now, personnel cost/hours are much more expensive relative to equipment, so people will pretty much pay whatever to get lab-grade stuff.

LolWolf··on The likely cheapest home-made Michelson interferometer
ha, thanks!

that one has uhh substantially less drift for what it's worth, but reprinting in more stable material would help that a ton (and still be quite cheap!)

LolWolf··on The likely cheapest home-made Michelson interferometer
what type of details would you like to see? happy to add more if helpful!
LolWolf··on The likely cheapest home-made Michelson interferometer
Thanks ! Still need to upload it to Printables etc. Maybe this will finally get my ass a bit more into gear.
LolWolf··on The likely cheapest home-made Michelson interferometer
thanks !
LolWolf··on The likely cheapest home-made Michelson interferometer
Oh absolutely!

I would not actually use this for uhhh, repeatable measurements over any extended period of time!

(If that were the case, I'd recommend re-printing it in a slightly more stable material, or just CNC milling the mounts out of aluminum using some of the ~$1-2K aluminum desktop mills and using some aluminum extrusions as the base.)

LolWolf··on The likely cheapest home-made Michelson interferometer
For sure! I didn't _quite_ want to get into the mechanics of it (assuming anyone who is interested would just take a peek at the wikipedia or any YT video).

But yes :)

LolWolf··on How to sequence your DNA for <$2k
For sure, I think it’d be fun to show it can be done under 5k or so, which would be close ish. Can probably get older equipment from university surplus, but yeah, these are the “things that need to get figured out” :)
LolWolf··on How to sequence your DNA for <$2k
Do you have a write up somewhere? If not, it would be amazing if you wrote one!

I was planning on doing a similar thing (also with saliva) once I finished moving in and had a bit more time after conferences. (But, of course, I’d have to go through and actually figure out all of the mechanics and so on.)

LolWolf··on Generalized Orders of Magnitude
ok I think I see where you’re going, though many (if not most) of the special cases and properties here seem to be essentially a consequence of each of these systems.

I’ll take a peek at the algorithm which I did admittedly skip, but curious if it gains us something !

LolWolf··on Generalized Orders of Magnitude
Hmm, how does this compare to things like

https://github.com/cjdoris/LogarithmicNumbers.jl

or

https://github.com/cjdoris/HugeNumbers.jl

(Apart from the PyTorch impl)

In particular, it feels like storing the complete complex number is a bit silly since we know, a priori, that the number exponentiates to ±1, so, wouldn't this mean that we have wasted 31 bits? (=32-1 since only one bit is needed for the sign.)

That being said, this representation is very useful for certain scenarios, of course, when you know that the dynamic range of your number is very large, but, as far as I can tell, it's not exactly super novel, unless I'm missing something!

LolWolf··on Language models pack billions of concepts into 12k dimensions
No, I don't believe that a citation is the same as peer review, it is, as I understand it from the post you're replying to, an observation that other people have looked at it. (Indeed, if you actually go through and _read_ this particular paper, you can see that they also cite multiple reproductions and extensions of the paper. I would call that peer review—and certainly much better peer review than most nonsense that appears from committees. Ask me how I know.)
LolWolf··on Language models pack billions of concepts into 12k dimensions
Not to completely plug my own work here, but I also wrote about this for a slightly more mathematical audience (and uhh, a much shorter post): "There are exponentially many vectors with small inner product"

https://lmao.bearblog.dev/exponential-vectors/

For those who are interested in the more "math-y" side of things.

For what it's worth, I don't fully understand the connection between the JL lemma and this "exponentially many vectors" statement, other than the fact that their proof relies on similar concentration behavior.

LolWolf··on Language models pack billions of concepts into 12k dimensions
What? Yes it is! This is exactly how peer review works! People look at the paper, read it, and then reproduce it, poke holes, etc.

Peer review has nothing to do with "being published in some fancy-looking formatted PDF in some journal after passing an arbitrary committee" or whatever, it's literally review by your peers.

Now, do I have problems with this specific paper and how it's written in a semi-magical way that surely requires the reader suspend disbelief? For sure, but that's completely independent of the "peer-review" aspect of it.

LolWolf··on Acorn and the future of (AI?) theorem proving
> it's some sort of oldschool model like random forests

It is a simple model, I think it's just a little dense NN with some hand-picked features. (https://github.com/acornprover/acorn/blob/master/python/mode...) Though Kevin has mentioned that, hopefully soon, we'll be able to have small-ish language models trained/fine-tuned on acornlib, but this requires some overhaul of the current approach to search.

> I wonder how they're going to keep it consistent over time (e.g., some proof steps suddenly becoming "unclear" after an update due to randomness of the ML model)

The idea (which I realize I didn't write about in the blogpost) is that each model iteration will be trained on `acornlib` along with previous traces. While it's possible that sometimes the model will fail to find a given proof for a fact (and therefore require some help when it didn't previously), the idea is that the model is learning both from its previous proofs and also new proofs that have been developed in the library since its previous training. In this sense, the model is constantly being "added" to and so should (ideally!) perform at least as well, as the acornlib gets larger.

> Sure, even Terence Tao would likely admit that formalizing proofs with Acorn is easier than with Lean, but formalizing existing/ongoing proofs is only part of the picture.

I think the point here is that, once the model gets good enough, you can just _write_ statements that you think might be true (and are, say, very nonobvious!) and then the Acorn engine would either automatically find a proof or give you a counterexample to the provided statement, which is much more akin to theorem proving with the computer.

> The best LLM provers + Lean combo right now can already crack IMO-level math, and what if they now write proofs in this more natural style?

For sure, my guess is LLM provers will find this style substantially "easier" though they do just fine at Lean, too.

> Will Acorn be better than Lean as an RL signal in this case?

I think this is the hope! The flywheel I think looks something like:

1. Train models using the standard library

2. Models can now discover new proofs of interesting facts and help in proving those more easily

3. Add those to the standard library + reasoning/proof traces

4. Train a new model on that

5. ???

6. Profit?

LolWolf··on Acorn and the future of (AI?) theorem proving
It is definitely less fleshed out since the project is very new (I think the first usable instance is like beginning of this year!) whereas afaik Lean 4's mathlib was ported from Lean 3 which has been around for at least 8 years and has had active contributions from around then too.

On the other hand, the `acornlib` (https://github.com/acornprover/acornlib) repo happily takes PRs that pass for many objects such as `Group`s (especially those in the standard library). so if there's something you're interested in proving, I would highly recommend just cloning it and proving it directly in the standard library and then PRing!

LolWolf··on Acorn and the future of (AI?) theorem proving
It's neat stuff imo!
LolWolf··on Fluid Simulation Pendant
wait, this is amazing !

I am tempted to make a slightly sillier version of this w/ a circular lcd rather than the uhh retro-cool smt leds honestly just to avoid the multiplexing design

great work!! super cool

LolWolf··on The unexpected poetry of PhD acknowledgements
Sure, it’s all a mask, but there’s something more “real” behind the acknowledgements in a very distinct way than the editorialized writing of an scientific article. (Using quotes here as these things are deliberately fuzzy!)
LolWolf··on The unexpected poetry of PhD acknowledgements
This is very cute :)

Definitely one of my favorite parts to read of any thesis are the acknowledgements. It’s often a time to lower the “fancy academic” mask and really see the human behind the writing; something essentially impossible with the dressed-up writing of many disciplines!

I believe the first part of my own acknowledgements began with: “There are only a few interesting things in this thesis. Out of all of them, this section might be the most important.”

To this day, this is still true!

LolWolf··on Legendre transform, better explained (2017)
It’s neat! To be fair, as a physicist, I did not understand the Legrendre transform essentially until taking convex optimization (where it is known as the Fenchel conjugate).

Many sources, but all of them are reasonable and give a constructive definition that actually explains what it does: we can characterize a function either by its graph, or its supporting hyperplanes (when it is a closed, convex function).

While the observation is almost silly, it has very deep consequences for different characterizations of problems and other constructions!

LolWolf··on Passkeys are now enabled by default for Google users
But you can set family members/significant others/etc as possible recovery mechanisms! This seems like a really workable solution that I don’t see people discussing in this thread?
LolWolf··on Sorry, Your Paper Coffee Cup Is a Toxic Nightmare
I think 18/10 stainless steel is probably good enough for anything

It’s chemically inert, doesn’t corrode, etc etc.

Glass is also great imo and also extremely chemically inert (except that, of course, it can shatter if dropped, but that’s a whole other thing)

Obviously glazed cups (with not shitty glaze, but that’s a whole Other thing) are standard, and so on

We have options! Just not, uhh, single use ones

LolWolf··on The EU proposed CRA law may have unintended consequence for the Python ecosystem
unfortunately lawyers (et al) will similarly pick the worst possible interpretation of a law as needed to win a case
LolWolf··on Surprise computer science proof in combinatorics
Eh, someone just suggested that a bound of this form might break some cryptographic assumptions for finite fields of (large) prime order. (This is not obvious to me, but it's a point that I think is worth considering, and already would show that such a result is useful.)

It's very hard to know how these things end up being useful. Even if a specific piece of knowledge isn't useful by itself, it can lead to things that are useful down the line. It's both very hard to know which ones will be, or which ones do lead to actually useful results; for example, think of work on propositional logic which ends up being the bedrock of computation, work on elliptic curves (cryptography), work on PDEs (physics), quantum physics (transistors), photonics (communications), etc.

Some fields have pretty "fast-to-application" times, but it's not always clear which ones will _not_ yield useful results. Plus, and here's the real, honest answer: all of this shit is very fun! Why not do it?

I'm also a (somewhat) practical person, and that guides many of the problems I try to solve, but sometimes there's just such a tantalizing puzzle that it's hard not to resist! It's pretty damn hard to "only" work on "useful" stuff and not be just an ok academic/programmer/etc., you have to let yourself get nerd-sniped into doing random things. With very high probability, these little "side-quests" end up being very useful down the line, for one reason or another.

LolWolf··on The model for coins in Super Mario Odyssey is simpler than in Super Mario Galaxy
And, lucky for everyone involved, we can solve huge constrained least squares problems in less than milliseconds on modern machines !
LolWolf··on Hyperscale in your Homelab: The Compute Blade arrives
Wait this is pretty sick! What's the full build on that? How do you even get started on finding good cases that aren't just massive racks for a home build?
LolWolf··on Researchers cannot always differentiate AI-generated and original abstracts
Oh, no, I'm not dismissing terminology! (That's a whole separate topic, which is distinct from writing, though it can sometimes be a sign of bad writing.) I'm talking specifically about the actual writing. Iirc, there's a phrase that sums it up quite nicely: "most science is written like it hates you and wants you to stop reading immediately." Poor writing is so much the norm that people are shocked when you write a readable article. (You'd be surprised at the number of peer review comments I've received to the effect of, "the writing is surprisingly clear and direct," and I'm not like, a novelist, or essayist, or anything. English isn't even my first language! It's just not the norm to think about, and much less act on, these things.)
LolWolf··on Researchers cannot always differentiate AI-generated and original abstracts
Yeah, a lot of scientific writing is just downright useless, and I don't just mean that in the "haha, it's hard to read, but it's ok"-sense. For example, in many fields (parts of theoretical physics, many parts of econ) publications are so hard to read that "reading" a paper looks less like "learning from the author by following what they did on paper" and more like "rederiving the same thing that the author claims to do, except by yourself with only some minor guidance from the paper." This is, frankly, absolutely insane, but it's the current state of things.
LolWolf··on Applied category theory in chemistry, computing, and social networks [pdf]
Idk, as a mathematician, I still don’t really see the usefulness of category theory “in the real world” (or even in a good amount of math)
← PreviousPage 2 of 15Next →