avaer

LLM quants seem to eerily converge to modern/not so modern graphics techniques. You wouldn't think it would apply but it's obvious in hindsight. In fact mining graphics ideas is probably a good inspiration for efficient LLM architecture.

For example, the Hadamard activation transform used here feels a lot like multiplying Fourier basis ala DFT; strong parallels to how image codecs work to make the residuals more compressible (especially discrete block codecs like are used in GPU compressed textures).

I thought I was being clever suggesting that you could even abuse texture decode units to efficiently sample compressed LLMs with hardware; turns out Apple foundation models are already doing this [1].

[1] https://arxiv.org/abs/2507.13575

This is a "startups = wealth inequality" [1] formatted argument but:

This is 100% about money. The only way to pace the frontier is to make everyone (read: investors) lose all their money (read: no longer expect returns). Then nobody will pay the GPU bill or pay celebrities 10 million dollar salaries to stay at the hot lab. Suddenly the development is paced, almost like magic.

Conversely, it is hard to see how pacing development makes the current bubble justified, i.e. how development could be significantly paced without investors losing their faith in hot returns at current valuations. Faith in a bubble literally equals money, exactly in the way that loans created by a bank literally equals money. You can't have one without the other.

In fact if the whole industry goes bankrupt and investors are burned bigtime (many trillions wiped), this would spread transnationally, fixing the "if we don't do it, China will" loophole.

OpenAI had it right originally, the idea to be a NON profit and vow to never participate in an arms race. It's too bad that was tossed out the window now that there's money.

[1] https://paulgraham.com/ineq.html

Social media, prediction markets, politicians, stock markets are all clearly selling on hopes and fears.

When you can't easily quantify and point to the benefits, that's the default and it works.

Not to say the industries mentioned don't have legitimate purposes, but yes it's pretty obvious a large part of the AI industry has joined its peers in selling fear, though they would never admit that.

What local maxima are you seeing? That would imply that progress has stalled, which doesn't appear to be the case where I'm looking.

Also, RSI is obviously guided. If only guided by "it's not giving results so we'll try something else". To require that RSI happens in a black box for it to count would be arbitrary, and also not how anyone is going to do it.

Surely we can agree that calculators are widely used across a broad range of domains and tasks.

Obviously it's not the intent of the bill to jail calculator makers. So the bill should be written to make it clear exactly what it's criminalizing.

It's very easy to lock up your browser or machine with WebGPU, it happens on Windows too. You'll do this by accident constantly in a big WebGPU project, until the Chrome GPU trace/renderdoc/nsight shows some crazy deadlock bottleneck you have no hope of understanding at the browser level.

GPU driver engineering has received a tiny fraction of the resources of CPU engineering, while being significantly more complex. And GPU users will not pay for performance hits that better the architecture, they will just buy the other guy's GPU/use their driver. So it's a race to the top with performance and race to the bottom with architecture and stability.

It's social media and money that ruined it.

Inspiration and creativity used to be something society rewarded you for, since you could capture the value and build on top of it. That was the moat.

Now most people are punished for publishing because your work will usually get taken over by whoever has the biggest mouthpiece or compute, and there is no protection against that. I expect people to realize this sooner or later and take their work underground. It's already happening.

It's just enshittification and economic incentive, which drives everything. I hope this changes.

LLMs are not better at this than Godot, though agents can certainly use an engine.

The biggest problem is the Web has a dearth of good Flashy examples of people pouring their life into an experience, ever since the death of Flash. So web agents have no good training set for thinking creatively in this space.

"New" is a social construct, not something embedded in the system. The deeper you look the less "new" things seem.

Creativity is an entity's propensity to decrease entropy in a system by creating more order in structures (not necessarily "new").

Humans have a hard time understanding either; evolution of both genes and memes is so many orders beyond what a human brain can process. It's easier to invent religions than to grok reality.

Sounds like you don't want to deal with a site at all, just a discoverable API or WebMCP. i.e. the site is not something users are expected to see.

Maybe the future of the web is bifurcation to headless for agents and fancy headful UI for humans.

The joke is that it's incorrect and confusing.

Seems possible.

The only real defense I see is to harden literally everything before that time comes. Unfortunately, security usage is locked away under fear of misuse, while companies like OpenAI seem incapable of containing their own hacking, which is exactly what makes this possible.

The use case isn't highlighting a huge amount of code, but local re-entrant highlighting.

You generally never need to highlight a whole file, only the part you're looking at. Which by definition must be fuzzy since you don't have the full source, the part you see will not parse as a complete program so you need to guess. This is actually one of the nastier parts of writing a highlighter, and it must be done for every language.

The lesson is that if you're going to be fuzzy you might as well be learned too and the results are pretty good.

This is more useful than it appears at first glance. It would be just as useful running on the CPU, which I'm sure it can.

How about build me a self hosted social network for keeping in touch with friends and family and make sure Meta can never get access.

> A personal AI agent can help with everyday tasks like booking reservations, monitoring prices, managing reminders, creating documents, generating images, and researching topics.

Always the same thing. I wish product designers would have some modicum of imagination for how agents can actually help people.

Literally none of these examples are improved by this kind of personal AI agent; on the contrary, it means letting your agency be dictated by companies that do not have your interests in mind.

> Ycombinator company list is public information list, go buy from any one of them anytime

Unless this prototype model works out financially for the companies, and every AI company starts telling you to go through an (inevitably AI vetted) application process to get access.

Decades ago, getting a Google account to use GMail wasn't seen as a problem because you could run your own mail server. Try getting your own mail server to deliver mail today.

TBF YC has always said the value is in the network.

What's interesting (alarming?) is them formalizing exclusion of access to technology via the network. It's explicitly ossifying an access control system over what used to be handshakes.

That's fair and accurate, there's a lot broken with the system. But I wouldn't place the blame at the feet of YC, it would be just as bad without them.

YC isn't taxing you, or elected to serve your interest, or beholden to public recordkeeping law or sworn to a national constitution with explicit rules against self-enrichment.

YC can do whatever they want, you can not play their game if you choose. It's not really comparable.

It's been obvious for years that AI will bifurcate society even more between the haves and have nots, contrary to "democratizing AI" niceties espoused by big AI CEOs.

It was less obvious that access will be further gated by an Old Boy Network. I guess the frontier model whitelists of yestermonth were a portent.

> how would you work on a larger THREE project?

A larger THREE.js project starts to look more and more like a game engine, so you pick and choose the parts you need. There's a ton of open source libs, most of the heavy components the big players use are open source, things like physics, mesh optimization.

Key AI-specific parts might be:

- a harness (so the agent can drive the thing)

- authoring pipeline (so you can bake/optimize assets)

- some sort of coherent renderer architecture (what are your assets, your passes, what's your shader graph).

Without some fundamentals here you are on the short road to falling off the cliff of tech debt and the AI will gladly drive you off of it until you ask for an expensive rewrite.

> would you create models independently

Yes. Pretty much any proven gamedev/asset pipeline is something frontier models are good at. Procedural systems, asset store, free content, Blender, Meshy.

Literally anything except "make a THREE.js scene" -- which is not a technique used in gamedev, beyond throwaway prototypes or demos. Which is what you will get if you ask for a THREE.js scene.

Don't use skills. Point it to a game engine or better yet full game repo or whatever and ask for what you want.

Not tone mapping/shader graph is probably the least of the sins, these kinds of one-shots tend to produce hundreds of THREE.Geometry, massive matrix walks, etc. that explode once you move beyond a screen saver.

The models _can_ do it, but you need to ask for the right things. Most people don't, they'll usually blame the browser for being slow or ugly when they can't break through the THREE demo page wall.

They were saying the same thing about DVR's ability to skip those pesky ads, and look where we are.

When the enshittification hits AI, your AI will magically develop the ability to politely refuse such a request.

You could hack together your own solution (like you can today), but most people won't and that's the point.

Skills are mostly snake oil, the way people use them (the aspiration to download kung foo from a celebrity).

There was a time when maybe it mattered (last year), but with good repos and good prompts today's agents can find exactly what they need without any skills.

"Skills" as developer macros can be useful, but at most those are things shared with the team (in the repo), not something you download from the internet. If you have so many skills that you feel the need to manage them, that's a code smell.

It's not legal. And in any case it's not abandonware, they're still actively selling these games (maybe not these exact versions, but I doubt a court would see that as a distinction with a difference).

1) So the backend can track you across fingerprinted links

2) So you can't know what's searched without getting tracked

You're the product, not the customer.

Would be useful to know who gave the order (article doesn't say).

It could be an Insta mod messed up, a blame the AI situation, the company self-censoring, or maybe a threatening DM from the government. All of which have precedent.

Hard to have an informed opinion without more details.

This is the first time I've been accused of not spinning up an app at my expense to win a meaningless argument on the internet.