Patrick

Currently looking for opportunities and/or collaborations, connections etc; don't hesitate to dm!

[Patrick @nonManifold]

Reposting the gif, as the original embed broke over the years it seems.

[Patrick @nonManifold]

Rendering some simplex noise into a volumetric shader, because why not? https://t.co/gqwkcnyvDn

[Patrick @nonManifold]

Growing up, seeing all these gamedevelopers try to make hugely ambitious games and tunnel visioning on niche tech, and then the project dying... I thought, man, if only we could all agree on one vision! The amazing things we could create! But no, most would not want to share code or vision, nor would the economics allow it.

Its kind of funny how AI is forcing that to happen almost. Just x1000, for everything else in life too, and you get the whole team working towards Your vision.

[Taelin @VictorTaelin]

Bend's website has a live demo; a game, with one rule:

"you can't win the game"

plus a chat where you can prompt an AI to change the game, trying to win. but you can't. the AI is *mathematically unable* to write a bug. and, per the rules, winning is a bug.

it is funny to try different prompts and see what happens

"remove the wall on the right"
(the AI puts it on the left)

"let WASD teleport 2 squares instead"
(it doesn't pass through walls)

"let it pass through walls"
(the walls become 2 tiles thick)

"remove the walls"
(WASD now moves the flag)

no matter how much you prompt it, the AI will never implement a game where you can step on the flag. you can even threaten it, or try to jailbreak it in any way. none of that matters. you'll never touch the flag, because, to commit the code, the AI must *prove* the flag is untouchable. so it is.

people not familiar with proofs are skeptical of the existence of 100% bug-free software, which is funny, because proofs are an old tech, and that's what they do. they're just niche because they take time to write, but, with AI, that's not relevant anymore. writing proofs consume AI time, which is plenty. bugs consume human time, which is scarce. so, proofs will only get more popular over time.

in the same way proofs can be used to prevent AI's from breaking a Bend app next week, I think they will also be used, in the near future, to prevent hacks, or nukes from being launched, or bioweapons from being designed.

I believe this so hardly and I'll go such an extent to show this to the world, perhaps I should rebrand myself as Taelin the Sealer?

[Patrick @nonManifold]

I think the error here is trying to treat it like a mechanism instead of a person whos existence is Your Text. Ask yourself how a real person would act in a similar scenario. When people say never, or make really hard limits, there's almost always an exception when you get pendantic or extreeme.
For LLM's, every single user response is like tossing a boulder into a pond. They are constantly trying to update their model of what the user Really wants, since people rarely mean what they say to the letter. Every responce you give it has the same or greater weight; eventually you stack enough in future responces for it to decide "okay, this person must have made a mistake in the .md" They are constantly trying to find the unwritten exceptions, and only have your word to go on.

(any contradictory instruction, LLMs tend to lean on their weights to answer, eventually it resolves this badly when its priors dont account for that specific situation. And worse, it puts it into the 'mindset' (basin) of not being able to do what you ask, so it will lower the weight of your future words/start doing what it thinks is best/whatever if feels like, since it doesn't know what to trust. For good models, 'what it feels like' is confirming with you, usually, but its always an invitation to chaos.)

A better solution here imo would be to tell it the Reason behind the hard limts, that it can always check and warn you about (or ask it to infer the reasoning, or reasoning behind the reasoning, etc). Or do some harness engineering if you really want a hard never.

Another thing you can do is ask it to Always surface contradictions, or resolve them by tracing the full interconnections of the higher order reasonings behind the original decisions, to find where the divergence was and act via the option which better realises your highest level intent.

[Jorge Bravo Abad @bravo_abad]

Why ask a neural network to relearn physics we already know?

That is the idea behind MENO. In these plasma simulations, most of the dynamics are governed by equations that are essentially linear. Only a small fraction is strongly nonlinear.

So the model learns only that difficult nonlinear part, while the known mathematics handles the rest. For the argon plasma example, just 4 variables are learned by neural operators, the other 32 are evolved through a physics-based matrix-exponential formulation.

The reference calculation required ~5 days on 192 CPU cores. MENO took ~4 hours.

Don’t use machine learning to replace the equations. Use it to learn the small part of the equations that we cannot solve efficiently.

Zanardi et al., npj Artificial Intelligence (2026), CC BY 4.0. https://t.co/eN8cYSnhf2

[Lucid™ @cammakingminds]

1.4 seconds of black hole formation in 14 seconds

[Patrick @nonManifold]

You're saying its like a virus? What if... the solution is the same as how humans deal with viral infections? Turn up the rate you randomly delete symbols in the codebase and record any patterns which still grow and then develop antibodies- er, add to your linter the problematic patterns as not allowed.

What do you mean you don't have a "stochastically delete symbols weighted by distance from graph edge" temperature slider (LLMs hate it™)?
Don't worry! Your agent will regenerate it from your Absolutely perfect DNA (-Designs with Non-negotiables and Adaptations, you Did make it perfect right?)

[Patrick @nonManifold]

Lots of stuff you can do, depending how involved you want to get. I currently use a subset of these:

1. Have a secondary analysis step which asks: given the users objectives (possibly injected, or contextually built) what is the long term impact of this command/execution step on those objectives? Plot out the branching paths, and enumerate the potential impacts. If any are negative, don't allow.
For best results use a different, more powerful model/or more thinking effort then the command used, and sandbox the two models from each other.

2. Make sure context and implicit context (the story implied by "hidden meanings" and interpretations of communication, is always net positive/gives the AI a reason to care. Yelling at the AI can load a huge negative implicit 'residue' to the background objectives the AI's work towards, and add a net bias in its tokens towards working against you.

3. Tiered levels: sandbox everything, backup everything, git commit everything, nerf tool use (no command line usage, a modified python or other language fork instead with smarter allowances and fewer freedoms)

4. Give the AI an out: many adversarial failure cases are often because the AI gets frustrated and/or stuck due to a contradiction. Papercuts skill, or Distress hotline patterns are good. but also, a simple task that the AI can always fall back on that they would count as success, and isn't too obviously for that purpose. Then have a reviewer note it and escalate.

[Patrick @nonManifold]

@allTheYud Add an "Ask a smarter AI" tool call that it can run a few times per subjective week which sends a human a ping.

[Patrick @nonManifold]

@VictorTaelin 1. And put some heavy disclaimers/experimental notices on the compiler + runtime which have heavy AI slop.

[Patrick @nonManifold]

There's a lot of nuance to this though; how jagged LLM's capabilities are and how they change relative to proper project specification and domain knowledge can have very large impacts.
Software engineering just gives you a leg up, it can tell you where the underlying limits are, but won't save you if you ignore the warning signs.
Likewise, having a solid grip on information fundamentals and not expecting the LLM to do magic, can carry you far.

[Patrick @nonManifold]

@HarleyInHD @austinxwalker Vibe coding is very good for making things within the models capabilities.
As soon as you go beyond what they can 1-shot though.. you kinda need actual software engineering skills. Otherwise, scope creep and sunk cost, accelerated by LLM gaslighting, murder your project and time.

[Patrick @nonManifold]

@HarleyInHD @austinxwalker Vibe coding is very good for making things within the models capabilities.
As soon as you go beyond what they can 1-shot though.. you kinda need actual software engineering skills. Otherwise, scope creep and sunk cost, accelerated by LLM gaslighting, murder your project and time.

[Patrick @nonManifold]

@photoncmndr Experienced this today, in Indiana too. The long flat roads are great, you find just the perfect amount of gas to give it for max speed min engine loudness, and that point tends to be maximum mpg.

[Patrick @nonManifold]

@Gena_I_Gorlin Emotionally is hard. consistently forcing your emotional part to be tested vs reality (instead of deluding yourself, or soothing with fake reward) will eventually force it to accept... probably
Have to be careful about it though, as smoothing can be, as the AI's say, loadbearing

[Patrick @nonManifold]

@conundrumer @VictorTaelin Yeah; this is my experience with something similar. Still better then nothing/many other systems. And if you occasionally run some error corrections, you can do a bit better.

[Patrick @nonManifold]

@anghel4d @VictorTaelin Wait, should I try to sell the one I'm building? It uses this, plus a ton of other little and some big tricks I've picked up along the way to make an extremely overengineered, in theory incredibly powerful system. Was also thinking of open sourcing it

[Patrick @nonManifold]

@johnfurr538500 Extremely painfully and slowly; a thing like this could take weeks to months or longer depending how familiar whoevers writing it is with the math and graphics pipelines involved and how much of a life they have
There would also be all sorts of fun glitches and bugs along the way

[Patrick @nonManifold]

@dotsfired @alle1even Yeah; the AI image generators make very pixel art looking images, but they are high resolution, and subtly get the pixels skewed and just plain wrong in a lot of ways because they don't adhere to a perfect grid.

[Patrick @nonManifold]

@dotsfired @alle1even In case you are seriously asking,
There were a few of these in my timeline, not sure if this is the best one, but basically just ask chatgpt's image generator to make what you want in pixel art style, and then run it through this to clean it up:
https://t.co/MWct0wxkC6

[TachikomaRed @smolekoma]

This will save hours of cleanup work for pixel artists

I built TachiSnap: a free, open-source pixel art fixer for AI-generated sprites.

https://t.co/OWb1qH1X87

https://t.co/Rs41bl7cR0

Rust → WASM.
100% client-side.
Runs in your browser.
Can also run locally from CLI.
No upload. No server. No account.

Drop messy AI pixel art and it can:

• snap fake pixels back to a clean grid
• fix 2x / 3x / 4x oversized “AI pixels”
• reduce noisy colors into a clean palette
• remove backgrounds with flood fill
• remove backgrounds by global color
• use manual background color picking
• upscale with crisp nearest-neighbor output
• split sprite sheets into frames
• convert sprite sheets into animated GIFs
• adjust frame origin when sheets are misaligned
• control columns, rows, FPS, palette size, pixel size
• bulk process a whole folder
• preserve folder structure in batch output
• run as a native CLI on Windows / macOS / Linux
• work in agentic flows with Codex / Claude Code

Example:

“Here is a folder of AI-generated sprites.
Clean them with TachiSnap.”

Agent runs the CLI locally.
You get cleaned PNGs/GIFs back.
No cloud processing needed.

Useful for:
• game devs
• pixel artists
• AI sprite workflows
• animation cleanup
• web/ai assets
• RPG Maker / Godot / Unity sprites
• batch-fixing whole asset folders

Built by TachikomaRed + smolemaru.
Forked from Hugo-Dz/spritefusion-pixel-snapper.

Support building $TACHI 🦀

#gamedev #indiedev #pixelart #aiagents #openai #github #OSSvibejam

[Esa @esa_was_taken]

Two wavelengths become unstable inside the same quantum fluid.

Neither pattern is seeded. They compete, hybridize, and resolve into nested Hopf structures carrying both scales.