antirez
@antirez
Reproducible bugs are candies 🍭🍬 I like programming too much for not liking automatic programming.
In Wohpe (Laurana, 2022) an engineer erroneously leave his phone near a GPU of Wohpe, believed to be fully contained. The AI discovers that could use the GPU itself as a resonator to establish an RF link with the phone, breaks the protocol, finds vulnerabilities and escapes into the Internet.
Habemus logo. From a collaboration between me (hand drawn pen and paper), AI (turn this shit into a logo) and Ben Gnomino (the human touch).
I was burning all my Fable tokens like Cartman in Casa Bonita since 7th of July was the last day and then...
Tomorrow I will no longer have access to Fable and yet it is interesting how I burned 14% of my weekly plan quota, even if I worked long days recently. The fact is that scarse & powerful intelligence resources can be used by making important but narrowed questions and implementations (continue)
DawrfStar with DeepSeek 4 Flash 4 bit, PRO 2 bit, and GLM 5.2 4 bit. M3 Ultra 512GB. Results on hard programming tasks with GPT 5.5 as a judge. So the GLM 5.2 branch is going to be merged and supported for CUDA + Strix Halo as well.
GLM 5.2, Q2_K routed experts (effectively ~2.6 bits) running with SSD streaming on an M5 Max 128GB computer.
Interesting AF. Promoting the best 6 layers to Q4 or the last 6 layers to Q4 (Q2 DS4 Flash quants) have different effects depending on what you check. The full logits error is smaller in the "last" variant, but actually the "best" variants (layers 32,25,15,27,23,31) (...continue)
That's why people using DS4F with DwarfStart, 2 bit quantized, are often surprised by the results. It's not a frontier model but it is not a toy, it is something you can actually use to get work done, and nobody can tell you want to do with it.
Took the good work of the communtiy of DrarfStar and consolidating the Strix Halo support. It looks very good. More QA in the next days and the final merge soon.
DeepSeek v4 PRO running via SSD streaming on my 128GB MacBook m5 max. 1.6 trillion parameters.
Anthropic did a big strategic error. Normally they compare their models with their old models. Instead, today, now that everybody knows how strong GPT 5.5 is at coding, they put it in the mix, basically showing all their customers that the benchmarks can't be trusted.
I finally found *the* solution I wanted to the old/new editing problem. And it is a solution that at the same time works extremely well, is quite elegant I believe, and can't be implemented if you don't build something like DwarfStar. Thread (but check [upto] in the screenshot).
For the DGX Spark owners. This is what you get with DS4 in your hardware. I want to post this to show how with fast prefill and not very fast generation, the system remains absolutely fine to use.
DeepSeek v4 Flash has 43 blocks... so, even 4096 batches during the prefill can be smoothly visualized. Look at the new progress bar.
By using Neural Accelerator via Metal4 API, DS4 is now very fast on M5 Max MacBooks.
Yet another case where Artificial Analysis mis-represent models capabilities? Or the Unsloth Minimax GGUFs have issues? Or is Minimax M2.7 just weaker than DS2.7 and this is the end of the story?
It was able to finish the working interpreter and write a working Mandelbrot program.
DeepSeek v4 Flash is able to use an EDIT tool I re-designed compared to what people normally use. The READ tool returns lines+tags, like: 1:f3_c int main(...), where the four chars are base64 crc of the line. So when there is to edit, there is no need to repeat the old line, just the tag.
Imagine a local agent where cache misses don't exist, tools don't need translations, you see progress for prefill, tokens are emitted ASAP.
