kgeist

18 ms·

241.
by kgeist 1y ago
They say it only works about 20% of the time; otherwise it fails to detect anything or the model hallucinates. So they're fiddling with the internals of the network until it says something they expect, and then they call it a success?
242.
by kgeist 1y ago
Qwen3 8b perfectly understood it after 14 seconds of thinking.
243.
by kgeist 1y ago
You can use all 128 GB if you use a MoE model
244.
by kgeist 1y ago
Try running local LLMs like Qwen3 yourself. They can calculate accurately in their reasoning traces even if you don't give them access to coding tools. In fact, even mid-range models (32b params) under 4-bit quantization can perform pr
245.
by kgeist 1y ago
From what I know, Siberia's mosquitoes are even more brutal than those in the more temperate regions of Russia, and there are far more of them. Iceland's lack of mosquitoes doesn't seem to be due to the cold itself; something
246.
by kgeist 1y ago
A couple of months ago I experienced this weird bug where instead of answering my questions, ChatGPT would discuss rome random unrelated topic about 95% time. I had to turn it off, too.
247.
by kgeist 1y ago
I did something similar: made an LLM generate a list of "blockers" per transcribed customer call, calculated the blockers' embeddings, and clustered them. The OP has 6k labels and discusses time + cost, but what I found is: -
248.
by kgeist 1y ago
Yesterday I was trying to make a small quantized model work, but it just refused to follow all my instructions. I tried to use all the tricks I could remember, but fixing instruction-following for one rule would always break another. Then I
249.
by kgeist 1y ago
Technically, RAG is anything that augments generation with external search. However, it often has a narrower meaning: "uses a vector DB." Throwing everything into one large context window is often impractical - it takes much more
250.
by kgeist 1y ago
>2014 That was before the Kremlin fully launched its repressive measures against the Internet, and before he was forced to leave Russia for refusing to cooperate with the authorities.
251.
by kgeist 1y ago
Durov left Russia in 2014 after he refused fo cooperate with the authorities (they asked to hand over personal info about pro-Ukrainian groups on VK), saying: >Unfortunately, it's impossible to run an online business in this country
252.
by kgeist 1y ago
Well, Russia has basically never had freedom of speech, except for a brief period in the 1990s, so it contributes nothing to his thesis in the tweet that "once-free countries are introducing dystopian measures."
253.
by kgeist 1y ago
>I realized that anyone can get phished A few years ago, I managed to get our InfoSec head phished (as a test). No one is safe :)
254.
by kgeist 1y ago
>They repeat after what humans already said >They are a parrot Is it really much different from most people? The average Joe doesn't produce novel theories every day - he just rehashes what he's heard. Now the new goalpost s
255.
by kgeist 1y ago
At work I've set up LibreChat + LlamaSwap + llama.cpp 200 weekly users :)
256.
by kgeist 1y ago
I use Llama.cpp with Tavily search (they give free credits each month). LibreChat has built-in support for it. No Ollama needed.
257.
by kgeist 1y ago
So any pure function is Markovian then? They too only depend on their input and are blind to previous invocations. I’m just trying to understand the practical purpose of calling something a Markov chain beyond small sizes, because if we all
258.
by kgeist 1y ago
I have an idea of how they work. At the end of the day, it's all just numbers. You say you put "one message", but in reality it's N input numbers corresponding to the tokens (where N can very from 1 to any size). Basical
259.
by kgeist 1y ago
But a conversation with an LLM doesn't only depend on the last message, it depends on all previous messages. Sure the number of messages is limited by the context window, so you can indeed say it's a high-order Markov chain (it ca
260.
by kgeist 1y ago
What is the conceptual difference between a "past state" and a "present state"? In a typical conversation with an LLM, you have past messages that remain in the KV cache when generating the next token (so as not to recal
261.
by kgeist 1y ago
Just a few days ago, I had to "clean up" a vibe-coded prototype: 1) generate API tests using AI 2) refactor the app manually 3) make sure the tests passed Took me a few hours or so. The app was pretty small, though.
262.
by kgeist 1y ago
"Orthodox" in Russian is "pravoslavny", literally "right faith" (pravyj = right, correct). I think it also contributes to the meaning. "The right way".
263.
by kgeist 1y ago
There was recently this paper: https://arxiv.org/abs/2508.21038 They show that with 4096-dimensional vectors, accuracy starts to fail at 250 mln documents (fundamental limits of embedding models). For 512-dim, it'
264.
by kgeist 1y ago
Here, 800 km east of Moscow, we've been experiencing mobile internet outages for about a month now. It disappears and reappears in a pattern I cannot understand. Sometimes it goes down across the entire city, and sometimes it works in
265.
by kgeist 1y ago
I think an LLM can be used as a kind of lossy encyclopedia, but equating it directly to one isn't entirely accurate. The human mind is also, in a sense, a lossy encyclopedia. I prefer to think of LLMs as lossy predictors. If you think
266.
by kgeist 1y ago
On my pretty modest dev machine with 12 CPUs, I once managed to achieve 14k RPS with Go+SQLite in a write+read test on a real project I was developing (it used a framework so there was also some overhead due to all the abstractions). I didn
267.
by kgeist 1y ago
The kid intentionally bypassed the safeguards: >When ChatGPT detects a prompt indicative of mental distress or self-harm, it has been trained to encourage the user to contact a help line. Mr. Raine saw those sorts of messages again and a
268.
by kgeist 1y ago
I've been using simple text files too, for 7 years now. Except I create a new file for each new day, to have peace of mind (as opposed to having a million-line-long file). Instead of Ctrl+F, I use grep. The format is Markdown. My typic
269.
by kgeist 1y ago
I've read that the formula is based on the early Mistral models and does not necessarily reflect what's going on nowadays.
270.
by kgeist 1y ago
With reified generics, you can also do "new T[]" because the type is known at runtime. With type erasure, you can't do that.

More ›