ashirviskas

6 ms·

31.
by ashirviskas 8mo ago
Author does not know what they're talking about. > In other words, XML tags have not only a special place at inference level but also during training Their cited source has 0 proof of that. It's just like python/C/htm
32.
by ashirviskas 8mo ago
Smaller quant or smaller model? Afaik it can work with anything, but sharing vocab solves a lot of headaches and the better token probs match, the more efficient it gets. Which is why it is usually done with same family models and most ofte
33.
by ashirviskas 8mo ago
Someone ping me in 5 years, I want to see if this aged like milk or wine
34.
by ashirviskas 9mo ago
Apple made lower than 16GB M3 models? Man, can't wait till the cheapest model is at least 128GB.
35.
by ashirviskas 9mo ago
Another European chiming in, I enjoyed OPs article.
36.
by ashirviskas 9mo ago
Do you also write your bytecode by human hands? At which abstraction layer do we draw the line?
37.
by ashirviskas 9mo ago
> it has, but python being single threaded (until recently) didn't make it an attractive choice for CLI tools. You probably mean GIL, as python has supported multi threading for like 20 years. Idk if ranger is slow because it is wri
38.
by ashirviskas 9mo ago
It is only 4 years old
39.
by ashirviskas 9mo ago
Keyboard autocomplete?
40.
by ashirviskas 9mo ago
I mean this is exactly what it is. Just a wrapper to replace the tokenizer. That is exactly how LLMs can read images. I'm just focusing on different parts
41.
by ashirviskas 9mo ago
You get that association that is relevant to your project only if you can cram the whole codebase. Otherwise it is making rough estimates and some of the time that seems to be where the models fail. It can only be fully resolved with either
42.
by ashirviskas 9mo ago
I wonder what if we just crammed more into the "tokens"? I am running an experiment of replacing discrete tokens with embeddings + small byte encoder/decoder. That way you can use embedding space much more efficiently and hav
43.
by ashirviskas 9mo ago
Well, using Claude Pro/Max Calude Code api without Claude Code, instead of their actual API they monetize goes against their ToS. I don't like it too, but it is what it is. If I gave free water refils if you used my brand XYZ wate
44.
by ashirviskas 9mo ago
You should never use GIF anymore, it is super inefficient. Just do video, it is 5x to 10x more efficient. https://web.dev/articles/replace-gifs-with-videos
45.
by ashirviskas 10mo ago
First, so best in this?
46.
by ashirviskas 10mo ago
I did, moved off to Claude in 2024 and many others around me did the same. Do you have data for non-business claim besides anecdotal evidence?
47.
by ashirviskas 10mo ago
Can you give specific examples? I'm super interested to see where it fails.
48.
by ashirviskas 10mo ago
Yeah, I remember the moment search engines invented computing, I cannot look at sand the same way anymore /s
49.
by ashirviskas 10mo ago
Where is Asahi not far enough if I may ask? I've been daily driving it for 2 years by this point. While it was not really useable for mainstream usecases in the beginning (no speakers, no webcam, other random issues), it did get better
50.
by ashirviskas 10mo ago
Maybe he lives in Netherlands and saw everyone avoiding the topic on bicycles. Or maybe they got autocorrected. Or maybe you're right and someone learned something new today. Merry Christmas, my dude.
51.
by ashirviskas 10mo ago
Claude 3 Sonnet was good enough for many things, but not as universal as 4.5 Opus. It is immeasureably more useful to me. I agree with all of your other points though.
52.
by ashirviskas 10mo ago
Ooh, I have experience with it. If you're on linux, just use Vulkan. If you face any other issues, just google my username + "MI50 32GB vbios reddit". It depends on which vBIOS you have, but that post on reddit has most of th
53.
by ashirviskas 10mo ago
Just use llama.cpp. Ollama tried to force their custom API (not the openai standard), they obscure the downloaded models making them a pain to use with other implementations, blatantly used llama.cpp as a thin wrapper without communicating
54.
by ashirviskas 10mo ago
Moon lander computers were also game changers. Does not mean I should be impressed by the compute of a 30 year old calcualator that is 100x more powerful/efficient in 2025 when we have stuff a few orders of magnitude better. For simple
55.
by ashirviskas 10mo ago
Ha! My project actually has 0 dependencies! /jkjk Though it's not in rust.. Mine is probably even more embarrasing, but mostly due to focusing on random threads: https://hn-wrapped.kadoa.com/ashirviskas
56.
by ashirviskas 10mo ago
Damn, the roast could have been me 2-3 years ago when I still tried to make macos work for me lol. Thankfully, asahi saved my sanity.
57.
by ashirviskas 10mo ago
1. I wanted it to be unique and easily searchable 2. Specifically this one derives from the obvious meaning + my surname, which starts with "Mine". Only a few days later I realised it might sound like it does something with mining
58.
by ashirviskas 10mo ago
> When FAIR-Path was applied to the tested models, diagnostic disparities dropped by about 88 percent. I have a better model to decrease disparities - just always return False. Equality solved. But I analyzed the data from their github a
59.
by ashirviskas 10mo ago
It really depends on how deep you want to go. And this will likely not be useful in any way, other than a new hobby. Me and my friends who do this kind of thing, we do it for fun. If it was not fun for me, I would not have bought 3 GPUs jus
60.
by ashirviskas 10mo ago
Oh, hi! Love your work! I'll be rooting for you that day (though most likely after that day, unless you're a superhuman).

More ›