cpldcpu

6 ms·

1.
by cpldcpu 4d ago
Well, that is what video models do. But keep in mind: So far there is no way to train the models while completely avoiding memorization and only including generalization. That would be a great way to avoid any copyright issues, but all atte
2.
by cpldcpu 4d ago
>Notably, the neural network component behaves like a compressed index of the training data. The "Compression" part is indeed about the model finding order in the training data. But this is not about applying some predefined co
3.
Show HN: Pico-Faces – A diffusion transformer image generator on a RP Pico 2 MCU(github.com)
4 points by cpldcpu 19d ago|0 comments
4.
by cpldcpu 2mo ago
its doing video, audio, images and motion. I think that counts as multimodal.
5.
by cpldcpu 2mo ago
>When Intel finally shipped it at the 45-nanometer node in 2007, Gordon Moore called it the biggest change in transistor technology since the late 1960s. The breakthrough was not the material. It was learning how to process the material
6.
by cpldcpu 3mo ago
hm.. has been quite a while for me. The good thing about the Tang Nano is that it is supported by the Yosys open source toolchain. There are quite a few resources on the web when you search for the combination.
7.
by cpldcpu 3mo ago
I had Opus 4.5 design an LLM inference engine in verilog, including firmware and automated verification a while ago: https://github.com/cpldcpu/smollm.c It's of course far from optical. But lowering the implementa
8.
by cpldcpu 3mo ago
CISC view: Its another adressing mode. RISC view: SUBLEQ is already four instructions (2x memory access, alu, branch)
9.
by cpldcpu 3mo ago
Great article! Yeah, that pattern can be seen everywhere in semiconductors. E.g. the transistor invention vs. Lilienfeld, Heil, Matare etc. So the scope is more narrow than "Inventend Semiconductors". Generally, there seems to be
10.
by cpldcpu 3mo ago
Wow! And it also implements a very interesting variant of SUBLEQ that is turing complete. >This VM implements an OISC - a One Instruction Set Computer. That instruction takes three signed 32-bit operands, a, b and c, and runs a program f
11.
by cpldcpu 3mo ago
The early discovery of light emission from silicon carbide long before the first LEDs is a very interesting finding, worth pointing out. But alas, as ever so often, the article turns this into a hyperbole. The premise from the title does no
12.
by cpldcpu 4mo ago
What do you mean with "open-source"? Of course, the inference code for all the open weight models is publically available - see llama.cpp or hf transformers. There are, however, very few models where also the full training pipelin
13.
by cpldcpu 4mo ago
Wow, nice! It's an honor. I guess nowadays one could use some of the 32bit WLCSCP microcontrollers to easily beat this.
14.
by cpldcpu 4mo ago
More than 25 years ago, there was a show off, of building the smallest web server: https://web.archive.org/web/20000815063022/http://www-ccs.cs... Someone with an ACE1101 microcontroller "won".
15.
by cpldcpu 4mo ago
From the blog post: https://github.com/wuxx/TD4-4BIT-CPU
16.
by cpldcpu 4mo ago
Nice!
17.
by cpldcpu 4mo ago
But why is this always the first comment on custom CPU builds? Can't there also be other designs out there?
18.
by cpldcpu 5mo ago
nah... their founders and most of their employees are in france. Very ill-suited comparison to IBM.
19.
by cpldcpu 5mo ago
https://huggingface.co/docs/inference-providers/pricing It's well buried though. Does not seem to be a focus of theirs.
20.
by cpldcpu 5mo ago
Great to see this! Worth mentioning that Huggingface already offers a similar service. And they are also European: https://huggingface.co/docs/inference-providers/index https://huggingface.co/infer
21.
by cpldcpu 5mo ago
I had to give it a try. Claude, the ole cheater, recognized what the file was, downloaded the psid from the web, found a wasm sid player and built a website around it: https://claude.ai/public/artifacts/df6cdcae-08
22.
by cpldcpu 5mo ago
Yes, marks of AI all over the place. Also the SVGs. >No solution written, 100% score. Its weird. Turns out that hardest problem for LLMs to really tackle is long-form text.
23.
by cpldcpu 5mo ago
Love it! Great idea for the dataset.
24.
by cpldcpu 6mo ago
+1
25.
Towards Self-Replication: Claude Opus Designs Hardware to Run Itself(cpldcpu.github.io)
1 points by cpldcpu 7mo ago|0 comments
26.
by cpldcpu 7mo ago
They mentioned that they using strong quantization (iirc 3bit) and that the model was degradeted from that. Also, they don't have to use transistors to store the bits.
27.
by cpldcpu 7mo ago
I wonder how well this works with MoE architectures? For dense LLMs, like llama-3.1-8B, you profit a lot from having all the weights available close to the actual multiply-accumulate hardware. With MoE, it is rather like a memory lookup. In
28.
by cpldcpu 7mo ago
It could simply be bit serial. With 4 bit weights you only need four serial addition steps, which is not an issue if the weight are stored nearby in a rom.
29.
Glowing Polyhedrons – LED filament 3D objects using graph theory(cpldcpu.github.io)
2 points by cpldcpu 8mo ago|0 comments
30.
by cpldcpu 8mo ago
Thanks! Would be quite interesting to see how this fares compared to mHC. I noted that LAuReL is cited in the mHC paper, but they refer to it as "expanding the width of the residual stream", which is rather odd.

More ›