antirez

22 ms·

511.
by antirez 2y ago
Ok, this explains why QwQ is working great on their chat. Btw I saw this thing multiple times: that ollama inference, for one reason or the other, even without quantization, somewhat had issues with the actual model performance. In one in
512.
by antirez 2y ago
Note the massive context length (130k tokens). Also because it would be kinda pointless to generate a long CoT without enough context to contain it and the reply. EDIT: Here we are. My first prompt created a CoT so long that it catastroph
513.
by antirez 2y ago
When discussing debt you need to factor in the interesting fact that people in Italy have savings sleeping on the bank that basically counter weight the debt.
514.
by antirez 2y ago
Indeed! I'm the original author of both projects, and thank you for using them :) I agree about Gemini, but I understand the sentiment is not widespread. For me the strongest LLM for coding is Claude Sonnet.
515.
by antirez 2y ago
In the video I show what happened AFTER the LLM fixed the bug while I was not recording. Of course I had no idea where the bug was, when the LLM found it. After a refactoring I removed a line for a mistake, causing a memory corruption. The
516.
by antirez 2y ago
I have a youtube channel with a playlist of coding with AI, where I show what I do with it (a small part, actually, but representative enough I hope). This is is the first video of the series I believe: https://www.youtube.com&#x
517.
by antirez 2y ago
As I said in one of my latest YT videos: if you can't code , sure, go all for it, the difference with x -> 0 in the denominator tends to infinity. But if you can code, there are much better ways to use generative AI to aid you in y
518.
by antirez 2y ago
In many ways I'm not an OpenAI fan (but I need to recognize their many merits). At the same time, I believe people are missing what they tried to do with GPT 4.5: it was needed and important to explore the pre-training scaling law in t
519.
by antirez 2y ago
One interesting thing about LLMs, that is also related to why chain of thoughts work so well, is that they are good at sampling (saying a lot of things about a problem), and are good, when shown N solutions, to point at the potentially bett
520.
by antirez 2y ago
Shouldn't that be 1% of emissions in the world? Btw, still more practical to address other stuff. Not that EV are a bad idea, but you can't force such change: the EU plans failed, because people don't have often a viable setu
521.
by antirez 2y ago
Good, now our (european) politicians should stop forcing companies like Volkswagen, with an incredible history and technological background, giving money to Musk for carbon credits. Use the same money to improve public transportation and th
522.
by antirez 2y ago
From this point of view I don't understand what's happening between the actual SOTA models practice and the academic models. The former at this point are all MoEs, starting with GPT4. But then the open models, if not for DeepSeek
523.
by antirez 2y ago
I guess it's what sometimes it's called "self RAG", that is, the agent looks inside the files how a human would be to find that's relevant.
524.
by antirez 2y ago
One of the silver bullets of Claude, in the context of coding, is that it does NOT use RAG when you use it via the web interface. Sure, you burn your tokens but the model sees everything and this let it reply in a much better way. Is Claude
525.
by antirez 2y ago
I believe that certain drives for certain regulations that are clearly shooting themselves in the feet were approved with less clear processes than needed (remember corruption cases of certain European politicians recently for the Olympics?
526.
by antirez 2y ago
LLMs are something everybody can reproduce with enough money, as shown by Musk itself recently: in one year they have Grok3 which is a SOTA model, without having particular talents or innovations, just following the recipe everybody else is
527.
by antirez 2y ago
That's not a problem, US is doing that, effectively marginalizing the importance they will have in the future.
528.
by antirez 2y ago
Weak study. You show photos where attractiveness is obvious. When you talk with people intelligence is obvious. Likely in a real meeting things could go differently.
529.
by antirez 2y ago
uv may be an improvement, but the Python packaging hell is a cultural problem that will not be solved without changing culture. And the main cultural issue is: 1. Depending on small and huge packages for trivial things. 2. A culture of brea
530.
by antirez 2y ago
Among binary quantized vectors, similarity is just popcount(XOR of the two vectors) / num_bits. Can be linearly translated to cosine similarity / distance range just multiplying and centering.
531.
by antirez 2y ago
Karpathy, which is IMHO a serious and balanced person, lamented that it looks too censored (see recent tweets). Elon Musk is (for me) a very scary person, and it is important to evaluate AI safety (but I believe that the safety that matters
532.
by antirez 2y ago
The relation among the internal model representations inside its latent space and the embedding of the CoT compressed with a text embedding model is, more or less, minimal. Then we take this and map it to a 2D space, which captures more or
533.
by antirez 2y ago
Take a large context frontier model. Upload 200k tokens of code for each query. Ask about what code pattern you want it to highlight for you. Works better than any other system, but costs token on API services.
534.
by antirez 2y ago
Not even the effort to check what happened in a year, re-asking the same questions to newer models. We went from last year ChatGPT to be almost useless if not as "reference" for well know things (like how to do that in Python), to
535.
by antirez 2y ago
Not just extra noise, but also extra tokens.
536.
by antirez 2y ago
lol good point.
537.
by antirez 2y ago
Totally agree that AI is going to have a huge impact on security of languages, and will change many paradigms.
538.
by antirez 2y ago
You can substitute AI with "Some Javascript Framework" and the subtitles still apply very well. Yet nobody was particularly concerned about that.
539.
by antirez 2y ago
That makes sense, thanks. We are all in part into this situation, because one bit here, one bit there, that's what we, as a collective entity, built, and there are clearly things that are not working as expected. It's hard to ackn
540.
by antirez 2y ago
That's not what happened. R1-Zero is a model per se, released with a different set of weights. Also it's not an intermediate step obtained making R1. In R1, a first SFT was performed before the RL training. While R1-Zero perform

More ›