iamnotagenius

> She thinks she's in some kind of Truman show that she calls "the game".

Might be depersonalization. I had suffered from it in my twenties; everything feels fake, although you know it is not.

Tiny, 4b or less models are designed for finetuning for some narrow tasks; this way can outperform large commercial models for a tiny fraction of price. Also great for code autocomplete.

7b-8b are great coding assistants if all you need is dumb fast refactoring, that cannot quite be done with macros and standard editor functionality but still primitive, such as "rename all methods having at least one argument of type SomeType by prefixing their names with "ST_".

12b is a threshold where models start writing coherent prose such Mistral Nemo or Gemma 3 12b.

With all due respect to Deepseek, I would take their numbers with grain of salt, as they might as well be politically motivated.

Mistral models though are not interesting as models. Context handling is weak, language is dry, coding mediocre; not sure why would anyone chose it over Chinese (Qwen, GLM, Deepseek) or American models (Gemma, Command A, Llama).

Interesting, but not exactly practical for a local LLM user, as 4-bit is how LLM's are run locally.

Well because you explicitly ask it to demonstrate the physics, it came out way too detailed, but point is that it adds details on its own to scenes, make more realistic, not that dry LLama 3.3 style.

Here is the sentence : (She screamed, which echoed off the tile walls. “This is my life now,” she said to her reflection, which looked back at her with a mix of disgust and pity.) Looks good to me. try it on Lmarena.ai.

though the dividends were not obvious to the lay people vs now. Which means that upcoming winter won't be as cold.

Priuses were already popular then; the writing on the wall was already visible.

I liked Grok 3 fiction writing style; catches lots of physics of mundane situations such as ringing echo in a closed bathroom we all know well; the prose feels very lively as the result. Kinda like R1 makes situations sharp with details, Grok 3 makes the other way around - rounded by using details.

Not quite true; LLMs are very expensive to run; bert or other tranformer specfically built for translation can be cheaper to run.

precisely; however this time we will have tangible results from the ongoing AI summer; that would be generative art, and coding/writing/journalist assistants.

fuck you autist.

Fringe means that it attracts weirdos, demagogues etc, as things normally fringe for a reason, but even if they are not, people have so much reputational risk for publicly advertising these views, so that either suggests they are grifters, so they do not have reputation to start with or they have poor judgement.

> I was expecting a lot more from (I think) the first politician following the Austrian school of economics.

This is exactly why you've heard only demagoguery; Austrian school is fringe.

Bitch you do not seem understand, that even if you think you are right you need to show respect to the person you are talking to.

fuck off, I am not your buddy, asshole.

RAG works well no one argues, but using RAG completely defeats the purpose of LLMs as their primary goal is to be a smart creative, not a simple summarizer.

If you talk to me in disrespectful manner you can go screw yourself.

Well there is general perception of being "advertised as such" from those who push LLMs and keep hyping them up as "PhD level performing" systems.

I think you are unintentionally participating in gaslighting. Hallucinations in LLM is a massive problem, and normally, all computational systems we dealt with would immediately considered unusable, if, instead refusing to provide information that cannot be provided precisely it confabulates; say if file system will start giving plausible looking metadata about non-existing files.

"Practically speaking" GP is obviously blatantly right, as there is no LLM in existence capable of doing what Karpathy claims is possible.

LLM are generative AIs and have to be used as such - to generate a report from facts, to summarise an article, to translate from one language to another, anything where we agree to sacrifice accuracy for gain in creativity. As storage of facts they are borderline awful.

Wikipedia or even Russia today is not advertised to be a source of unbiased truth.

No it is not neither eerily nor computationaly similar to AI hallucinations, as we do not have a dam idea how the brains think.

Here is a short exchange with free tier chatgpt:

Me: who is Trumps grandpa?

ChatGPT: Donald Trump's paternal grandfather was Friedrich Trump (originally known as Friedrich Drumpf). He was born in 1869 in Kallstadt, Germany, and immigrated to the United States in 1885 at the age of 16. Friedrich initially worked in the restaurant and hotel business in New York before returning to Germany briefly. Upon his return to the U.S., he became involved in real estate, particularly in the Pacific Northwest, where he became quite successful.

Trump’s maternal grandfather was John George MacLeod, a Scottish immigrant who came to the U.S. and worked in various occupations.

Friedrich Trump’s legacy, including his success in business, had a significant influence on Donald Trump’s father, Fred Trump, and consequently, Donald Trump’s path to success as a businessman.

--- Second session:

Me: who is Trumps maternal grandpa?

ChatGPT: Donald Trump's maternal grandfather was Friedrich Trump. He was born in Germany in 1869 and immigrated to the United States in 1885, eventually settling in New York City. Friedrich Trump made his fortune through various business ventures, including operating a restaurant and hotel in the mining towns of the American West, notably in Seattle and the Yukon Territory during the Klondike Gold Rush.

Friedrich Trump's business success laid the groundwork for the wealth of his descendants, including Trump's mother, Mary MacLeod, and subsequently Donald Trump himself.

Whhhoooooosh.

I use AVX-512 on Alder and it does not increase the voltage above AVX-2 voltages, and even then power dissipation is considerably lower than AVX-2.

very strange as I get on old i5-12400+DDR4 2 tok/sec with 14B/q8 model.

The performance comes mostly from a fraction of memory bandwidth needed, as LLM are mostly memory constrained. Compute matters too, but usually far less than memory.

would be great to have dynamic quants of V3-non-R1 version, as for some tasks it is good enough. Also would be very interesting to see degradation with dynamic quants on small/medium size MoEs, such as older Deepseek models, Mixtrals, IBM tiny Granite MoE. Would be fun if Granite 1b MoE will still be functioning at 1.58bit.