Hacker News: topwalktown

New comment by topwalktown in "Can AI Exit Vim?"

topwalktown — Sun, 29 Mar 2026 16:11:02 +0000

I wrote it myself it python. The model runs in a loop and can either read the terminal or input keys. If it does neither the loop exits.

New comment by topwalktown in "Can AI Exit Vim?"

topwalktown — Sun, 29 Mar 2026 03:37:12 +0000

Sorry! It should work now.

New comment by topwalktown in "Can AI Exit Vim?"

topwalktown — Sun, 29 Mar 2026 01:52:41 +0000

I test if qwen3.5-35B-A3B can exit vim when running in a harness that allows it to read the raw terminal buffer and send raw key presses to the terminal. Enjoy!

Can AI Exit Vim?

topwalktown — Sun, 29 Mar 2026 01:52:41 +0000

Article URL: https://theadamcolton.github.io/can-ai-exit-vim

Comments URL: https://news.ycombinator.com/item?id=47559788

Points: 37

# Comments: 23

Spokester1: Realtime Neural Game Graphics

topwalktown — Wed, 31 Dec 2025 23:52:25 +0000

Article URL: https://theadamcolton.github.io/spokester1-neural-game-graphics

Comments URL: https://news.ycombinator.com/item?id=46449630

Points: 1

# Comments: 0

New comment by topwalktown in "The FFT Strikes Back: An Efficient Alternative to Self-Attention"

topwalktown — Fri, 28 Feb 2025 05:23:18 +0000

Transformers like Llama use rotary embeddings which are applied in every single attention layer

https://github.com/huggingface/transformers/blob/222505c7e4d...

New comment by topwalktown in "Image Self Supervised Learning on a Shoestring"

topwalktown — Sat, 06 Jul 2024 21:02:44 +0000

I'm trying to train a variable resolution ViT using IJEPA. I'm currently topping out at about 30% on imagenet1k after training for 20 epochs (6 hours)

It'd be cool to have some help and feedback. I'm on the right track to getting really killer setup that is super fast to train it needs more evaluations and more tuning. Anyone interested?

Image Self Supervised Learning on a Shoestring

topwalktown — Sat, 06 Jul 2024 21:02:44 +0000

Article URL: https://theadamcolton.github.io/image-ssl-on-a-shoestring

Comments URL: https://news.ycombinator.com/item?id=40893293

Points: 4

# Comments: 1

New comment by topwalktown in "Generative modelling of compressed image file bits"

topwalktown — Wed, 21 Feb 2024 09:12:16 +0000

i experiment with using a (mostly) unmodified llama model to generate images, by training on the bits from a lossy compression algorithm. It turns out the key is having a decoder which can give 'hints' as conditioning information for the autoregressive model, about what the decoder is going to do with the next token in the stream

Thanks!

Generative modelling of compressed image file bits

topwalktown — Wed, 21 Feb 2024 09:12:16 +0000

Article URL: https://theadamcolton.github.io/generative-modelling-of-compressed-image-file-bits

Comments URL: https://news.ycombinator.com/item?id=39451678

Points: 1

# Comments: 1

New comment by topwalktown in "The VAE Used for Stable Diffusion Is Flawed"

topwalktown — Thu, 01 Feb 2024 17:58:18 +0000

yeah, check out the Emu paper by meta. They basically do all of what is mentioned in the above comment

New comment by topwalktown in "MeshGPT: Generating triangle meshes with decoder-only transformers"

topwalktown — Wed, 29 Nov 2023 00:21:35 +0000

Quantization also works as regularization; it stops the neural network from being able to use arbitrarily complex internal rules.

But really it's only really useful if you absolutely need to have a discrete embedding space for some sort of downstream usage. VQVAEs can be difficult to get to converge, they have problems stemming from the approximation of the gradient like codebook collapse

New comment by topwalktown in "Your VAE Sucks"

topwalktown — Mon, 06 Nov 2023 20:26:29 +0000

I wrote a short article about jpg and if we could use concepts from how jpg works to make an image autoencoder that has a left-to-right positional bias and variable compression

Basically, existing VAEs are pretty good at compression, but have bad properties like 2D latent position bias and difficulty training on batches of mixed resolutions

So I try something I call DCT-Autoencoder, which takes ideas from JPG to learn compression of patched DCT features of an image

Check it out!

Your VAE Sucks

topwalktown — Mon, 06 Nov 2023 20:26:29 +0000

Article URL: https://theadamcolton.github.io/your-vae-sucks.html

Comments URL: https://news.ycombinator.com/item?id=38168477

Points: 4

# Comments: 1