grasgor

992 posts

grasgor

@grasgor

world Katılım Eylül 2023

667 Takip Edilen73 Takipçiler

Sabitlenmiş Tweet

grasgor@grasgor·10 Kas

with @sadernoheart finishing in on his 100 days, and people asking @elliotarledge where to start, i'll drop the link to my worklog more like public diary, had slowed down but we've resumed - Day 10/100 notion.so/grasgor/CUDA-2…

English

184

10.9K

grasgor@grasgor·1h

I realized that I find myself using google's ai mode more often than chatgpt now for general queries

English

grasgor@grasgor·1d

okay first local runs on the way

OpenAI@OpenAI

Are you up for a challenge? openai.com/parameter-golf

English

grasgor@grasgor·1d

@classiclarryd llm vessels is true, im already burning tokens without having put in much thought except while creating the prompt, perhaps I will be more involved when claude cant get me far enough

English

844

Larry Dial@classiclarryd·1d

Very cool. My 2 cents for participants: most compute will be spent on undifferentiated hill climbing from people functioning as LLM vessels. Agents can climb hills, but humans are still superior at finding them. What paradigm can you introduce? Sparse circuit discovery and compression during training? Variable embedding sizing? Manifold-ultra-connections? Paired head attn on steroids? Decision tree distillation? The list is endless.

OpenAI@OpenAI

Are you up for a challenge? openai.com/parameter-golf

English

472

58.2K

grasgor@grasgor·1d

i'm yet to come across a way more effective than pen and paper while learning

English

grasgor@grasgor·2d

@stevibe thoughts on temporal memory rather than re-captioning each frame?

English

162

stevibe@stevibe·2d

I'm obsessed with pushing local small models to their limits. Qwen3.5:0.8b doing real-time video captioning on a Mac Studio M2 Ultra, streaming descriptions as the video plays. Under 1s per frame — 269 frames captured & described from a 3m49s video. Pause anywhere and read the captions, it describes every frame surprisingly well. This model is barely 1GB. Local AI is moving absurdly fast.

English

130

1.9K

103.2K

grasgor@grasgor·4d

tell me mirofish (if it works as expected) won't become minority report (it will)

English

grasgor retweetledi

Discerner@Discerner4u·4d

@bnjmn_marie Yes 4b is amazingly slow.. and borderline unusable for agents

English

grasgor@grasgor·4d

@ashvanth_s1 now that you ask, it might be because the model is a reasoning one so the actual output takes timei plus the agent loop (idk what hermes does under the hood; if it does something specific), although I get ~76 tok/sec for plain llm inference

English

Ashvanth.S@ashvanth_s1·4d

@grasgor Your laptop has 5060 and it is somewhat slow still ??

English

grasgor@grasgor·4d

maybe it's coincidental or maybe because it just knew, nonetheless quoting since this popped on my TL. I've got a qwen3-4b-q4 quant with 128k context running on my laptop's 5060. It's somewhat slow, and I don't know what I'll do with it but definitely a start

Sudo su@sudoingX

cancel your chatgpt subscription and delete your openclaw slop. i'm serious. go on ebay and buy a used RTX 3060 for the price of two months of pro. or check your drawer because half of you already own one and forgot about it. install hermes agent from @NousResearch. one framework, 31 tools, file operations, terminal, browser, code execution. connect it to your local llama.cpp server running qwen 3.5 9B Q4. total download is 5.3 gigs. that's it. that's the whole setup. every experiment you hesitated to run on API. every project you shelved because you didn't want your data on someone else's server. every late night idea you didn't test because you hit your rate limit. all of that is gone. runs 24/7 on your electricity. your machine. your data never leaves your house. connect it to telegram if you want it on your phone. hook up whatever tools you need. the model thinks at 29 tok/s with 128K context and it never bills you. qwen 3.5 9B and one RTX 3060 is the setup most people will never try because they've been trained to believe intelligence has to come from a datacenter. it doesn't. it runs on 12 gigs of VRAM under your desk right now. stop giving your thinking away for free.

English

181

grasgor@grasgor·4d

@pareto_pakodas @DejaRu22 how big is all of it (size) ?

English

361

Shreyas Rao@pareto_pakodas·4d

Ok done. This took a bit longer than expected because my ISP provider had some "fair-use" data quota which I busted downloading all this. Had to bump up my subscription to get more data. Glad that this (and the AI models ) have been downloaded. Homage to @DejaRu22 !

DR22 Ω 🪬🎭@DejaRu22

Monthly reminder: Back up everything you want to access later Download all of it External hard drive Backup external hard drive Cloud storage if you are so inclined

English

16.1K

grasgor@grasgor·4d

It's probably going to be a tweak it as you go, add things you want, I can see this becoming what neovim was for terminal editors. I'm currently running the model via llama.cpp but I'll see if I can manage to speed up the inference

English

grasgor@grasgor·5d

@_kaitodev @karpathy So I should become a chef is what he's saying

English

148

Kaito | 海斗@_kaitodev·5d

5 minutes ago, @karpathy just dropped karpathy/jobs! he scraped every job in the US economy (342 occupations from BLS), scored each one's AI exposure 0-10 using an LLM, and visualized it as a treemap. if your whole job happens on a screen you're cooked. average score across all jobs is 5.3/10. software devs: 8-9. roofers: 0-1. medical transcriptionists: 10/10 💀 karpathy.ai/jobs

English

967

1.8K

12.1K

3.5M

grasgor@grasgor·5d

someone please tell my friend

English

grasgor@grasgor·12 Mar

@divya_venn oops

English

divya venn@divya_venn·11 Mar

@grasgor *i’m wrong

English

divya venn@divya_venn·11 Mar

software is no longer a technical field, it's a linguistic one

English

3.3K

grasgor@grasgor·11 Mar

@willccbb @seconds_0 good at kernels? curious, where does a human's moat lie when an agent can iterate much faster - adapt to newer architecture given documentation better

English

446

will brown@willccbb·11 Mar

@seconds_0 either getting good at evals or getting good at kernels

English

297

11K

0.005 Seconds (3/694)@seconds_0·11 Mar

Hypothetically, ai research will still exist as a field in 2027 Hypothetically , if someone wanted to spend the next 9mo upskilling to be an AI researcher, what does that look like? What are the type of outputs labs would expect? Any existing study tracks that are valuable?

English

501

51.5K

grasgor@grasgor·10 Mar

@ashvanth_s1 always english, it repeated my prompt thrice in the name of reasoning trace