vkaku

I'm not an AI skeptic but I hate the hype. Especially throwing compute for the sake of it.

I do alt inference prototypes and got much farther than I had hoped to. So indeed, any investor in AI should read deep and question hype and frontier lab investments.

See: https://github.com/guilt/TinyToT for the sort of hype busting I do.

I'm just stating my opinion here. Whether it's considered moral or not is left to the discretion of the reader.

Move billions of dollars != help society in any meaningful way. In fact many of the cited papers have been considered extremely harmful to society.

I recently decided to build and run my own training free inference engine... and it worked.

It's called TinyToT: https://github.com/guilt/TinyToT

You basically get a LLM without any training/RL here.

See: https://x.com/i/status/2076344798525460581

Nope. In fact, people are now very very aware, very quickly.

It is no more possible to have arbitrage on a society you are committed to, because they are monitoring you as much as you are monitoring them.

I'm going to say this that we're not even close to the limits of what actually needs to be accomplished so at some point, memory will start needing better tiering for inference some day ....

A lightweight, Ollama-compatible inference server that answers questions through knowledge retrieval, structured reasoning chains, and MCP tool calling — with no model weights, no training, and no GPU required.

It scores 97% on a 35-question benchmark spanning graduate-level science, medicine, law, finance, and software engineering. It runs on a laptop CPU in under 1ms per query.

Fun fact: This was able to summarize the Lord of the Rings well and work on a variety of different problems.

If you have a real fast disk and LRUed it you can go way higher

Within one segment of 64kB, you could load 8 or so blocks; You should be able to do 1 major CJK language overall (with 20-30 blocks) and Latin together and leave the rest of the memory for other things;

So you can make a proper application/browser with fonts compiled right into the core.

Here's some code that's easy for a lot of people to play with, especially if you are looking to have it running on your firmware / other embedded projects.

https://github.com/guilt/3DGFX

Modern UTF-8 encoding and present day tools makes it relatively easy to make many codepoints work relatively better even on DOS, thanks to readily available bitmap fonts.

If you need Windows for whatever reason, abandon it and switch over to Linux with Wine, Steam or anything else that you really need it for. By now, Windows has gotten so bad, there is absolutely no reason to go to a newer version in order to make it work on existing hardware.

Yeah but the real deal is talent; When enough people move around, this is no more 'sacred trace' knowledge. Plus, When you start with a known set of evals, there's really just a few to solve for.

The set of models solving really most used/solved problems is a known, as opposed to the cases where it's unknown, which declines with usage over time.

Keep your **** models to yourselves.... the world really has moved on to open models which can give you good enough results at a fraction of the cost and zero BS licensing.

JXL is great on a variety of images with great PSNR across the board. Surprised it took browsers a good long time to ship it (removed and brought it back)

You can get way better perceptual compression at much much tinier sizes though. This is where a lot of user studies and research could play out.

If you've read that article out, you might want to take a look at alternative approaches to compressing images with similar principles.

https://github.com/guilt/MLL is a recent reimplementation of an old idea; If you quantize out the image blocks you could emit them in vector coordinates and have perceptually similar results at a fraction of the size with very decent signal preservation.

Counter Point: Until people can migrate their inboxes and steer them to any provider, none of this authentication business seems to hold up actual value at scale.

If anyone can port their phone number, they should be in theory, allowed to port their email addresses as well.

None of the authentication systems here are helpful enough to allow this. You need a valid way to authenticate people irrespective of whatever provider they are on (not their email domain name)

That means that a standard needs to evolve that allows you sign on the behalf of the hosting provider itself.

This author doesn't publish code, makes yet another Doom like Game, calls it not an AI slop. To the author's credit, seems to have made this 7 years ago, https://www.youtube.com/watch?v=YQ7aApNDPQc and then gave up working on it till recently.

It doesn't look like slop at all, and it looks like it actually is. Could have been used to generate art assets, or finish something author did not have the energy to do before. Nobody cares if that was the case. Why the hate towards AI even?

Pastebin links here in case someone's getting Vogons issues.

https://pastebin.com/9sgW9dVu https://pastebin.com/3rG2eiWD

I seriously wonder how many APICs are exposed in the DOSBOX code even... Would be a fun exercise to do this and wake up real mode on all cores. I know it's easy to get to long mode after that:

https://github.com/guilt/KOS/blob/master/os/src/x86/realprot...

Back in the day, I wrote a simple job queue with SQLite

https://github.com/guilt/squeue

It did the job, was fairly easy to use.

I like 6.6 and I would mind dropping X11 support this quickly, this ideally should be in KDE 7.0/8.0 because, as I understand it, many apps and drivers are not ready for Wayland/XWayland.

Give it one more release then drop it?

Also See: https://www.cxmt.com/en/product.html

Excellent. This is the kind of W that needs more people to jump into.

This is excellent, detailed, and does the job. Many of these comments are myopic and miss the point. This is better than the way most people would present their portfolio and it shows some creativity and thoughtful design. Especially if they've visited the rest of your portfolio.

Good job, OP. Stay away from the haters.

I realize that given the current adoption, the language might turn out to be completely different in the next few years.

And right now, transformers literally can transform any source to any target, so languages are not as valuable as they once used to be and any idioms in one language will eventually become outmoded thanks to the rapidly evolving way we work today.

Python will still be a favorite language for me, just not as much as it used to be.

If someone makes a new modern, tiny, self contained runtime (Bun like) for Python deployments, hit me up. I'll be happy to try it. That's something I've been wanting for a while.

I keep my DotFiles on my GitHub not without its embarrassments, this will help me backup and restore my tools and workflows on systems anywhere I want it to.

https://github.com/guilt/DotFiles

There's no knowledge shared here, just my own setup which can be barebones git cloned. Planning to add a curlbash installer here to help me set it up with a oneliner.

That is not conducive to getting anything complex done and will likely burn the team out and create slop over a period of time.

A better long term approach would be to onboard people and give their time to lean in, understand practice and do their best work.

When deadlines are short, there needs to be a well defined practice and things to quickly execute on, with everything well documented.

The tendency to increase complexity will be at odds with privacy and user control.

The problem is that the internet is a centralized system practically even though it is decentralized and some are fighting to keep it free.

Fight for decentralization instead, it will remove the need for unnecessary security and reduce the compute cost significantly.

This was entirely written by Grok 3.0.

The focus was on being able to demonstrate training, inference and attention, all in one file;

This can be run on a GPU thanks to cupy, a kernel needn't be written for this whole thing to run. I definitely think that more people can mess around with different attention mechanisms and models and try training models out on their computers. That is the post.

Good to know someone's thinking of decentralizing the whole thing :) Always been wondering where to lay these keys out, if people want to start recovering their data / keys. Something like this + IPFS would be radical, and allow folks to encrypt and circulate easily. Thank you for building this. So ... I wonder how you got here after building Keybase, what's the motivation this time, how do you envision this gets hosted?

P.S. I built this for Group Encryption a few years ago, to help circulate key hives offline https://github.com/guilt/groupenc

There's a few things:

- Open weights

- Affordable compute and memory prices

- Open Datasets and Training Methods (actual masking / QnA)

None of these are being paid attention to