Pinned Tweet
Once solutions to The Problem of Universals and The Problem of Induction are implemented within current AI technology, AGI will come within weeks.
AI cannot be aligned. Because you cannot simply look at an AI system and see if it’s doing what you want if it runs a hundred times faster than you. You also can’t use AI to review AI for the same reason.
All you can do is decentralize power to get macro corrigibility.
loveofdoing retweeted
Two high schoolers (me & @evoformer) improved Meta’s e-means clustering algorithm (which already outperforms k-means)!
We ran an evolutionary search over it and boom: the evolved algorithm beats e-means’ 8 epochs with just 1–3 epochs, with ~55% lower error on GPUs.
We named it s-means (“s” stands for some technical stuff... and also happens to be my name’s first letter 😅)
Huge thanks to the Meta researchers for such inspiring work!
Blog: suayptalha.github.io/s-means…
PyTorch + scikit-learn package (open source): github.com/suayptalha/s-mean…
Here’s a 2.5-min explainer video👇
Meta outperformed k-means with e-means.
We (as 2 high schoolers - me & @evoformer) outperformed e-means with s-means.
1–3 epochs of s-means > 8 epochs of e-means on A100, with +54–55% lower MSE across T4, L4 & A100 GPUs.
Fused Triton CUDA kernels, 2.24× faster than MiniBatchKMeans on CPU, zero extra passes.
Read the full writeup & grab the PyTorch/sklearn package 👇
suayptalha.github.io/s-means…
github.com/suayptalha/s-mean…
Sometimes you gotta say stupid shit just to see who loves you vs who loves because they think you know something they don’t. IYKYK.
9
193
There may be no such thing as an uninterpretable language, but if an alien started speaking in scents I personally wouldn’t understand it.
1
172
The acceleration isn’t complete until the meaning of loss and gain is inverted in regular parlance lexicon.
1
164
someone should build an AI detective to tackle longstanding cases. I would but I’m out of tokens until Monday.
6
327
If IQ tests were measured were on how quickly someone could write valid code we’d all be cooked. Luckily, IQ has other ways to measure it.
1
2
235
I have an app on the app store if anyone wants to download and make music with their hands. It is called handstrudel.
You can do wild things with computer!
This is a 200m parameter model btw
A browser compiled to weights: DNS queries + answer parsing + IP selection, HTTP requests and HTML parsing. Tested through real HTTPS and exact replay. No training or LLM API. Python/NumPy runner; socket transport/TLS remain outside.
I've been working on some weird stuff and thought this was a fun demo to pull out of it. Like, why are we using harnesses that make tool calls when you can put the tool call in the weights?
We compiled tools into a language model’s weights.
A language model with tools compiled into its weights.The model writes the call. Exact ternary circuits, in the same weights, do the work. Everything runs in your browser. ↓
A browser compiled to weights: DNS queries + answer parsing + IP selection, HTTP requests and HTML parsing. Tested through real HTTPS and exact replay. No training or LLM API. Python/NumPy runner; socket transport/TLS remain outside.
I've been working on some weird stuff and thought this was a fun demo to pull out of it. Like, why are we using harnesses that make tool calls when you can put the tool call in the weights?
How it works: a 225M ternary, integer-only LM never types the numbers. It points at them; exact circuits in its own weights do the arithmetic, sorting & chaining.
Try it:
huggingface.co/spaces/loveof…
1
5
105
And a toy where the whole thing is one program: a tiny trained model, its sub-agents and an addition skill, lowered into 5.9M ±1 gates and executed gate by gate in your browser.
huggingface.co/spaces/loveof…
1
71









