pvillano

I once had the strange experience of playing such a game. It was an "instant game" on Facebook Messenger, that simulated a slot machine casino. It didn't have a cash-out, so it was really just a game.

It had an auto-play, so I watched the dials spin and lights flash until my money ran out, and it prompted me to buy more.

The game played entirely by itself, until it inevitably lost. It ran with no meaningful player action other than selecting which slot machine and buying coins.

And it had a leaderboard, with Facebook profile photos and names of big winners that I can only assume spent real money to be there.

Could a network be trained to transform physics state directly into the latent state and back?

Having a direct transformation would enable some interesting experiments.

How is the latent state different when everything else stays the same, but you change one physics value, like player one velocity? Is there a cyclical pattern of activation that correlates strongly with the seconds digit of the clock? Can you decode the latent state, give players full boost, and then re-encode it for infinite boost, without losing continuity?

Edit: There sure are a lot of papers on interpretability.

It is a faster, lower fidelity model.

but it's not JUST a faster, lower fidelity model.

The new model can be queried in new ways. The encoding from the RenderMan scene into the neural proxy allows for differentiable relighting. Once you have differentiability, you can find out what inputs create a desired output with gradient descent, instead of trying every possible input.

Neural proxies are also much, much faster for only a small drop in fidelity. 1 hour to 60hz is a 216,000x speedup. That's not possible without neural proxies. You could try lowering precision, resolution, bounces, scene complexity. Accuracy would be out the window before you get close to that performance.

Neural proxies are a constant-time approximator for anything.

There's still difficulty in finding exactly where the proxy should go, i.e. what step can be approximated without losing fidelity in the output. Apparently, you also need to select auxiliary features to guide training. But if you figure those out, you can replace hours of computation with milliseconds, accurate to the limits of human perception.

I look forward to the coming era of human-optional formally verified programming competitions.

I wonder what other optimization+verification problems are out there that will make good LLM feedback loops.

Maybe something with query planners.

I really think that a genAI picture book is like a microwave dinner. You can try to pick out a good one, but a corporation has done most of the deciding for you. Ultimately, it will not be as wholesome as something homemade, or as good as something made by an expert, and it will always taste like plastic.

AlphaGo, through self-play, found new strategies to beat the best human players. But now that anyone can play against AlphaGo, humans are learning these strategies.

Deepmind is saying people are learning from the algorithms AlphaEvolve is generating. They say the programs tend to be short.

So maybe we will understand the algorithms

Fully automated testing + revision loops are super cool and def something I want to play with

And PBT is so close to language level invariants/preconditions/post-conditions that I dream of

You don't need to check the proof. A regular proof is written to convince other mathematicians. A regular proof can skip details and contain mistakes. A formal proof cannot skip details or contain mistakes.

A lean proof is a sequence of instructions that starts with a list of facts, and combines them into new facts until a target fact is produced. Each instruction in the sequence is one of a few allowed tactics provided by lean that always produce correct facts from correct facts.

IF (1) you agree with the starting facts, (2) you are sure the target fact is what you're actually trying to prove and (3) the lean compiler says "yep, this is a valid sequence of instructions" THEN you can be sure the target fact is true.

You can be sure the final fact is true, even if you don't understand all the steps, because it's written in lean, and lean only allows steps that are free from omissions or errors

Liability, and it's the hardest part.

When I write like I talk, I use a lot of commas. Replacing some of my commas with em dashies, so long as it was done judiciously, would probably make things easier to chunk.

I really look forward to the day AI-driven algorithm design + formal verification becomes the norm for performance critical computing.

A programmer translates a natural-language spec into a machine-readable spec, feeds it to an AI-assisted compiler, and out pops an implementation that's more optimized than any human could ever hope to write, along with a lean proof of its correctness.

The cover does not matter for a textbook.

Most textbooks sold are bought by students because they were required for a course. Students are not choosing a textbook by cover because they're not choosing a textbook at all. Professors choosing which textbook to assign are doing so based on the content, because that's what they'll be teaching. Professors also get a lot of free sample copies, and are probably choosing between those instead of purchasing their own set of candidates based on the cover.

12,060 pieces

In one of the essays posted here, which was, ironically, about AI in education, a sentence, that an AI could not possibly write, that I could possibly write, because of its length and unusual structure, before finally reaching the verb, went on for 25 words.

I don't know if it was written that way to show trust in the reader's intelligence, show disregard for reaching a wide audience, show a demonstration of skill, or was artifact of someone just thinking at that level.

A sudden switch to depreciation makes housing more expensive, even though prices are going down.

Say you buy a house with a loan, and sell it after x years. The total cost of ownership is interest paid to the bank + buy price - sell price. If you bought a house at $600k expecting to sell it in a few years at $700k, but it'll actually sell for $500k, your TCO went up by $200k. The sudden switch to depreciation made the same house much more expensive to own. Prices have to slowly go from rising to steady and then to falling, otherwisw everyone is suddenly living way beyond their means.

A cleaning station is a location where aquatic wildlife congregate to be cleaned by smaller organisms. Such stations exist in both freshwater and marine environments, and are used by animals including fish, sea turtles and hippopotamuses.

Nobody cares that a browser's navigation buttons, address bar, tabs, or window controls don't match the current website. Probably because these things are obviously outside the extents of the web page.

However, scrollbars, context menus, modal windows, and date pickers are rendered within the extents of the web page, and get replaced all the time.

It is my opinion that these controls don't need to be styled to match the website, because they're not part of the website. They're part of the browser. Non-diegetic. Outside the fourth wall.

I found joy making sdf2stl.saej.in. It's similar to a bunch of things that already exist, but I was able to make a set of requirements that made mine technically unique.

It was too complex for me to create when I started, so I learned as I went.

In personal projects, I only use LLMs for when I don't care and I don't want to learn. For example, all the WebGL APIs. I am perfectly fine not knowing how to allocate, initialize, and bind a shaderbufferindexobjectstoragehandlepointerarray. But I'll do the fun parts, the interface design and algorithms, by myself.

I remade the hero GIF in HD

https://www.shadertoy.com/view/sXsGDl

I'm a big fan of the fact that the following quoted sentence has 25 words before the verb. An LLM would never. I probably wouldn't.

Whatever your conception of the modern university, whether grand or grim, understanding the current landscape of campus-wide AI use, much less its intensification, should destroy it.

I made sdf2stl.saej.in

You can use it design 3D objects with mathematics and bring them into the real world with 3D printing.

You can use it to create 3D models that are impossible to create with CAD, CSG, sculpting, or mesh-based tools.

You can build on a decade of community SDF development on Shadertoy because sdf2stl uses the same language.

A few days ago I may have become the first person to 3D print the equation x^4+y^4+z^4-x^2-y^-z^2+.4=0 (Goursat's surface) https://www.printables.com/model/1713835-goursats-surface

Please try it out and tell me what you think.

https://sdf2stl.saej.in

Thanks for letting me read that again.

I once wrote about inherent, irreducible complexity and how we try to deal with it. The draft has sections on how complexity can be hidden, spread out, localized, passed off, or recreated from scratch. Unfortunately, people are now using LLMs to pile complexity on the simplest of tasks, and my essay isn't really worth finishing.

I would rather do more work myself to make an API as fool proof as possible, than hope everyone else is perfect, and lose data when they're not.

What effect does docker compose have on database resilience?

If you're in this comment section, consider play-testing your website. Find someone who has never used it and watch them explore it for the first time, while they think out loud, without giving them any help. My personal website had links to GitHub, LinkedIn, etc. on the home page, and the first thing my brother in law did was leave the site, without ever looking at any of my posts, which were indexed on another page.

This example might be obvious to you, but I guarantee there's something you can learn through play-testing.

I think this comment section would be a good place to ask for help with a related problem.

I want to design a "smooth" closed path that fits in a square, as long as I can make it. It needs to smooth in such a way that a constant velocity can be maintained subject to limits on acceleration and jerk. The point is to develop a test for maximum flow rate on a filament 3D printer without the motion system ever slowing the tool-head down. (In reality, the standard smoothed "E" shape is good enough. This is more of an exercise.)

I know roads are designed to limit acceleration and jerk, and someone knowledgeable about road design would know about finding curves with constraints on acceleration and jerk.

Do you know any free tools or resources on theory that I could use?

https://www.instagram.com/inaturalistshitpost/

Jia Tan

How much information is there in knowing the length of someone's password?

If we know the password's length, it saves us from guessing any shorter passwords. For example, for a numeric password, knowing the length is 4 saves us from having to guess [blank], 0-9, 00-99 and 000-999. This lowers the number of possibilities from 1111 to 1000. The password has 90% of it's original strength. A [0-9a-zA-Z] password retains 98% of it's original strength