Muhammad Kone
136 posts
Muhammad Kone
@ProgrammingProg
NYC software engineer
New York, NY Entrou em Mart 2025
5 Seguindo0 Seguidores
@hussamfyi How would I build my own model so maybe you don’t face this anymore
English
the tech industry has lost the plot man. where is the thinking happening? when you throw a bunch of context at claude and ask it to produce some artifact, whether it’s a “design” or a “prd”, you’re completely disconnected from the invisible but critical work of translating fuzziness into something coherent. the outputs aren’t yours. you’ve lost the mind-body connection to the artifact. you’re running blind. what’s the point? to move faster?
English
"And who tops the leaderboard? Labs doing RLFT on a Chinese frontier base" 🤣
it has never been a better time, to make your own model from scratch! don't think, just do it. 🏴☠️
doomslide@doomslide
What? Pre-training? No, no, no, no. No pre-training. Why would you do pre-training?! If you do pre-training, people will ask "HOW MANY TOKENS?" And it will never be enough. The model that was the frontier breakthrough becomes "just distillation from chatgpt" But if you just do SFT, some RL on benchmarks? You're efficient. You're doing *reasoning*. It's not about capabilities, it's about the benchmark score. And who tops the leaderboard? Labs doing RLFT on a Chinese frontier base.
English
How do I build my own LLM from the ground up?
,@MistralAI, @01AI_Yi, @togethercompute, @NousResearch, @OpenBMB, @internlm_ai, @stabilityai, @allen_ai, @EleutherAI, @Cohere would love your takes.
English
@NVIDIAAIDev and how about creating a new ai frontier with zero ai knowledge lack of proper schooling or zero use of code, hack , python or anything ....... simply Logic? .....
English
Going from knowing zero about ML to building a ML framework from scratch in 4 months is kinda crazy 👀
Aadi Kulshrestha@MankyDankyBanky
I trained a 12M parameter LLM on my own ML framework using a Rust backend and CUDA kernels for flash attention, AdamW, and more. Wrote the full transformer architecture, and BPE tokenizer from scratch. The framework features: - Custom CUDA kernels (Flash Attention, fused LayerNorm, fused GELU) for 3x increased throughput - Automatic WebGPU fallback for non-NVIDIA devices - TypeScript API with Rust compute backend - One npm install to get started, prebuilt binaries for every platform Try out the model for yourself: mni-ml.github.io/demos/transfor… Built with @_reesechong. Check out the repos and blog if you want to learn more. Shoutout to @modal for the compute credits allowing me to train on 2 A100 GPUs without going broke cc @sundeep @GavinSherry
English
Who in robotics did the first electro scooter thing because I’m seeing them a lot now I think it started to gain hype in 2023-2024
English
Writing backend apps using ai I think to myself I probably could have done this faster and I could have learned something new aswell
English
@perplexity_ai I would love to work on some features for the perplexity personal
Computer as an engineer.
English
Today we're releasing Personal Computer.
Personal Computer integrates with the Perplexity Mac App for secure orchestration across your local files, native apps, and browser.
We’re rolling this out to all Perplexity Max subscribers and everyone on the waitlist starting today.
English
I have been writing harnesses for local edge models (~4B) in stealth for a new project.
It is unreal how much commonsense harness engineering principles dont apply when you are working with smaller dumb models.
- Smaller contexts
- can't guarantee structured outs without low-level guidance/constrained decoding
- machine heats up on long sequences (turboquant or not)
- actually requires finetuning to work best/lower context
- feature specific lora adapters and hotswapping, wooohooo
If I live through this experience, I will tell the story someday.
Thariq@trq212
English
Wow so you can run your backend locally. So the machines (computer laptop phone) can run your front end or backend so a server is just a computer. I learned that
English
English
btw ~80% of the world’s agents and ai engineering is done in these 3 square miles
swyx 🏝️@AIEmiami@swyx
English
react is just like jetpack compose, componenets are essentially composables
English
Stanford is kinda crazy because as a CS undergrad this term you’re choosing between:
- CS336: 0 to hero on frontier model training
- CS224R taught by Chelsea Finn (founder of Pi)
- CS231N taught by Fei Fei Li (Imagenet, WorldLabs CEO)
- CS221M Mech Interp Intro taught with Goodfire
And a host of personal podcasts delivered by $T CEOs.
Jesse Mu@jayelmnop
protip for stanford undergrads: beware the classes with guest speaker lineups that read like AI coachella. you’re basically paying $5k to listen to a live podcast series.
English
docker repo name has to match the image name each image you upload to the repo has different quantifiers called tag for eg repo name = dockerman image name = dockerman:version1. maybe you find that helpful
English
using different sErvices like DOCKER im noticing something, that i would benefit ALOT when using new software to learn how to use the terminal.
English