Just wanted to share a few thoughts before going to sleep. DeepSeek allegedly made changes to the codebase of the inference library it runs on:
(PR is a set of changes that is about to be included in the codebase)
I'd not be so sensationalist about that as the media are, but probably that's how we reach AGI - by connecting our wet computers with enormous data centers via Neuralink. In the above example, a human agent used AI to improve its codebase, or was it other way around?
The second thing is that I've recalled is that Clif High told that someone wanted to buy his precognitive text-mining project, but he declined, as there are many things he interprets himself. During one of the sessions for Future Forecasting Group, Edward Riordan got data indicating that a group was using (or might be using, I don't remember correctly) a supercomputer to run future predictions, possibly in the same way Clif did—by detecting anomalous human behavior patterns as a kind of bleed-through of future events. Ideal task for the Stargate. Yeah, it's weird but fun to contemplate
This PR provides a big jump in speed for WASM by leveraging SIMD instructions for qX_K_q8_K and qX_0_q8_0 dot product functions.
Surprisingly, 99% of the code in this PR is written by DeekSeek-R1. The only thing I do is to develop tests and write prompts (with some trials and errors)
Here is an example of the prompt that I used: convert ARM NEON to WASM SIMD prompt
Indeed, this PR aims to prove that LLMs are now capable of writing good low-level code, to a point that it can optimize its own code.
ggml : x2 speed for WASM by optimizing SIMD by ngxson · Pull Request #11453 · ggerganov/llama.cpp
Motivation This PR provides a big jump in speed for WASM by leveraging SIMD instructions for qX_K_q8_K and qX_0_q8_0 dot product functions. Surprisingly, 99% of the code in this PR is written by De...
github.com
I'd not be so sensationalist about that as the media are, but probably that's how we reach AGI - by connecting our wet computers with enormous data centers via Neuralink. In the above example, a human agent used AI to improve its codebase, or was it other way around?
The second thing is that I've recalled is that Clif High told that someone wanted to buy his precognitive text-mining project, but he declined, as there are many things he interprets himself. During one of the sessions for Future Forecasting Group, Edward Riordan got data indicating that a group was using (or might be using, I don't remember correctly) a supercomputer to run future predictions, possibly in the same way Clif did—by detecting anomalous human behavior patterns as a kind of bleed-through of future events. Ideal task for the Stargate. Yeah, it's weird but fun to contemplate