Aung Kyaw Soe (AK)

You have two engines: one runs on symbols (CPU), the other on vision (GPU). Use both 🔥 Symbolic reasoning is a valuable but i think Spatial reasoning is more accessible. Mathematical ideas are still communicating in very narrow bandwidth medium and legacy static pdf/paper

Color prompting 🎨 Steer image generation not just with words but palettes...

Cooking ...

The iOS Photos app has one of the most impressive interactions and UIs like you can pinch to control photos density timeline. Recreated the photo-editing “styles” pad, experimenting with an imperial blue color.

Neither ClaudeCode/Cursor nor Figma can replace this workflow yet Until a device with 120fps, without backlit exists 😄

Exploring joystick as user interface for mobile. iPod inspired 😄

Wow! i'm impressed by this 1 shot @v0 max

Added Transform UI. Direct manipulation is always fun!

@brdrck Super cool lines!

I've almost got volume fill down! Still some obvious bugs, but damn this looks insanely cool! ortho.brdrck.me

zoom-to-cursor (keep the point under the mouse fixed) and geometric zoom, two ideas inspired from the first time I used Figma

draw shapes with SDFs in a single draw call so they stay crisp at any zoom x.com/kernelsoe/stat…

Started exploring what an AI-native design tool that runs in the browser on WebGPU looks like. I learned how to:

made an interactive signed distanced fields to learn pixel shading and drawing smooth shapes

@ctatedev Thanks! Just needed this today!

Introducing json-render AI-generated UI. Deterministic output. 1. Define your component catalog 2. AI steams JSON 3. Render interactive UI Let users prompt dashboards, widgets and apps - safely constrained to components and actions you define

@karpathy Thanks for sharing! Can you also share your thoughts on making/training/optimizing smaller models efficiently useable locally?

New post: nanochat miniseries v1 The correct way to think about LLMs is that you are not optimizing for a single specific model but for a family models controlled by a single dial (the compute you wish to spend) to achieve monotonically better results. This allows you to do careful science of scaling laws and ultimately this is what gives you the confidence that when you pay for "the big run", the extrapolation will work and your money will be well spent. For the first public release of nanochat my focus was on end-to-end pipeline that runs the whole LLM pipeline with all of its stages. Now after YOLOing a few runs earlier, I'm coming back around to flesh out some of the parts that I sped through, starting of course with pretraining, which is both computationally heavy and critical as the foundation of intelligence and knowledge in these models. After locally tuning some of the hyperparameters, I swept out a number of models fixing the FLOPs budget. (For every FLOPs target you can train a small model a long time, or a big model for a short time.) It turns out that nanochat obeys very nice scaling laws, basically reproducing the Chinchilla paper plots: Which is just a baby version of this plot from Chinchilla: Very importantly and encouragingly, the exponent on N (parameters) and D (tokens) is equal at ~=0.5, so just like Chinchilla we get a single (compute-independent) constant that relates the model size to token training horizons. In Chinchilla, this was measured to be 20. In nanochat it seems to be 8! Once we can train compute optimal models, I swept out a miniseries from d10 to d20, which are nanochat sizes that can do 2**19 ~= 0.5M batch sizes on 8XH100 node without gradient accumulation. We get pretty, non-itersecting training plots for each model size. Then the fun part is relating this miniseries v1 to the GPT-2 and GPT-3 miniseries so that we know we're on the right track. Validation loss has many issues and is not comparable, so instead I use the CORE score (from DCLM paper). I calculated it for GPT-2 and estimated it for GPT-3, which allows us to finally put nanochat nicely and on the same scale: The total cost of this miniseries is only ~$100 (~4 hours on 8XH100). These experiments give us confidence that everything is working fairly nicely and that if we pay more (turn the dial), we get increasingly better models. TLDR: we can train compute optimal miniseries and relate them to GPT-2/3 via objective CORE scores, but further improvements are desirable and needed. E.g., matching GPT-2 currently needs ~$500, but imo should be possible to do <$100 with more work. Full post with a lot more detail is here: github.com/karpathy/nanoc… And all of the tuning and code is pushed to master and people can reproduce these with scaling_laws .sh and miniseries .sh bash scripts.

@JacklouisP It has no intention of reusing the bags 😂

Something very satisfying about the way this system opens bags of bulk plastic powder. Not sure if it's the slice or the wiggle

@JieWang_ZJUI @physical_int @sundayrobotics Great explanation! Can we transfer the teaching plan and knowledge for factory robotic arms? I mean same robotic arms but can perform a wide range of tasks with precisions

Brewing a latte is one of the HARDEST manipulation tasks today. We saw impressive results from @physical_int and @sundayrobotics . But why is it so hard? Shouldn’t we have had fully autonomous coffee machines for decades? Happy Sunday with coffee—here’s a thread on why. 👇☕️🤖

@chris_j_paxton Looks cool! How’s the visual intelligence work here?

A wonderful AI powered home robot that does actually work

@nandafyi The theme and interface is on 🔥!

interactive visuals for a new site (still some bugs to iron out)