antirez

27 ms·

481.
by antirez 2y ago
Because there were plenty of evidences that the statements were either not correct or not based on enough information, at the time they were made. And to be wrong because of personal biases, and then don't clearly state you were wrong
482.
by antirez 2y ago
As LLMs do things thought to be impossible before, LeCun adjusts his statements about LLMs, but at the same time his credibility goes lower and lower. He started saying that LLMs were just predicting words using a probabilistic model, like
483.
by antirez 2y ago
You can train small models with the Air M4.
484.
by antirez 2y ago
For Vector Sets, I decided to write all the code myself, and I use the models very extensively for the following three goals: 1. Design chats: they help a lot as a counterpart to detect if there are flaws in your reasoning. However all the
485.
by antirez 2y ago
In complicated code I'm developing (Redis Vector Sets) I use both Claude 3.7 and Gemini 2.5 PRO to perform code reviews. Gemini 2.5 PRO can find things that are outside Claude abilities, even if Gemini, as a general purpose model, is w
486.
by antirez 2y ago
What he sees is useless complexity. What he is proposing is more complexity wrapping all that. That's not the way.
487.
by antirez 2y ago
Because of the language
488.
by antirez 2y ago
Is all that folks "Why is Europe like that" from US? In that case, remember that TikTok is going to be banned because it is destroying US social networks business. On the contrary, this is an official order after the proof of copy
489.
by antirez 2y ago
To jump off the treadmill is not using a fronted framework: at all, and not using a random one and not rewriting the code later. Server side rendering, JavaScript only when needed, no separation between backend and frontend folks in the com
490.
by antirez 2y ago
From every angle I try to look at this, it does not make 32B sense.
491.
by antirez 2y ago
If you want to play with reading/recreating a very small Tcl interpreter, recently I put Picol (a 500 lines of C code Tcl interpreter) on Github. It was still on the web, but a bit more "hidden". I had a chance to re-read the
492.
by antirez 2y ago
This Ballard style stuff is the things I love.
493.
by antirez 2y ago
The problem is, where is this 20T tokens that are being used for this task? No way to access them. I hope that at least OpenAI and a few more have solid historical storage of the tokens they collect.
494.
by antirez 2y ago
I offer a POV that is in the middle: reasoning is powerful to evaluate which solution is better among N in the context. Memorization allows sampling of many competing ideas from the problem space, than the LLM picks the best, making chain o
495.
by antirez 2y ago
I believe LLMs are both data and processing, but even humans reasoning is based in strong ways on existing knowledge. However, for the goal of the post, indeed it is the memorization that is the key value, and the fact that likely in the fu
496.
by antirez 2y ago
The best Ruby always surpasses the elegance of the best Python... Unfortunately for practical means at this point I go for Python: more libraries, less problems with the C implementation of the interpreter (had issues with the GC in the pas
497.
by antirez 2y ago
My fear is that similar dynamics were at play when the AI act was shaped and approved, with the three European biggest countries unhappy of the content of the act (France, Italy, Germany).
498.
by antirez 2y ago
Hello! I tried to show it Redis code yet not released (llama.cpp 4 bit quants and the official web interface) and V3 can reason about the design tradeoffs, but (very understandably) Gemma 3 can't. I also tried to make it write a simpl
499.
by antirez 2y ago
Good test. I'm slowly accumulating private tests that I use to rate LLMs, and this one was missing... Thanks.
500.
by antirez 2y ago
The model performs very poorly in practice, while in the benchmark it is shown to be DeepSeek V3 level. It's not terrible but it's at another level compared to the models it is very close to (a bit better / a bit worse) in th
501.
by antirez 2y ago
The most disturbing thing is that in the chart it ranks higher than V3. Test a few prompts against DeepSeek V3 and Gemma 3. They are like at two totally different levels, one is a SOTA model, one is a small LLM that can be useful for cert
502.
by antirez 2y ago
After reading the technical report do the effort of downloading the model and run it against a few prompts. In 5 minutes you understand how broken LLM benchmarking is.
503.
by antirez 2y ago
[I guess the tweet refers to YC the incubator, not the web site of news, but I want to focus on the latter]: I registered on HN a few days after it was created. The HN community was incredibly instrumental to my career, my ability to keep u
504.
by antirez 2y ago
Thanks for reading!
505.
by antirez 2y ago
This statement refers to the fact that, there are no rules like "you should put a O to block X.X" and things like that. Inside the program there is just the rule to understand if the game ended and who won. In the NN there is noth
506.
Reinforcement Learning in less than 400 lines of C(github.com)
10 points by antirez 2y ago|4 comments
507.
by antirez 2y ago
I believe that the issue is that the culture about Redis usage didn't evolve as much as its popularity. To use it memcached alike has many legitimate use cases but it's a very reductive way to use it. For instance sorted set ranki
508.
by antirez 2y ago
I'm not talking about the fine tuning that make them side with the user even when they are wrong (anyway, this is less and less common now compared to the past, but anyway it's a different effect). I'm referring if in the tem
509.
by antirez 2y ago
The error accumulation thing is basically without any ground as regressive models correct what they are saying in the process of emitting tokens (trivial to test yourself: force a given continuation in the prompt and the LLMs will not follo
510.
by antirez 2y ago
There is a disproportionate skepticism in autoregressive models and a disproportionate optimism in alternative paradigms because of the absolutely non verifiable idea that LLMs, when predicting the next token, don't already model, in t

More ›