Ishan Anand
@ianand
Demystifying AI at http://spreadsheets-are-all-you-need.ai Prev: VP Product @ EdgioInc, CTO/Cofounder @ Layer0Deploy, MIT EECS
Wanted to share an AI side project: I’ve implemented GPT2 (an ancestor of ChatGPT) entirely in Excel using standard functions. By using a spreadsheet anyone (even non-developers) can explore and play directly with how a “real” LLM works under the hood. spreadsheets-are-all-you-need.ai
What Ilya saw...five years ago
- Reinforcement learning is lot more powerful than people think. Just scale it up. Seems too simple but yeah that's the discovery. - Reward shaping in human interpretable ways can work (entropix like R1-Zero rewards I'm looking at you) - RL systems keep beating the humans
pretraining is an elegant science, done by mathematicians who sit in cold rooms writing optimization theory on blackboards, engineers with total absorb of distributed systems of titanic scale posttraining is hair raising cowboy research where people drinking a lot of diet coke…
BPE transformer watching an H-Net output an entire wikipedia article as one chunk
Build an LLM in the browser in vanilla JavaScript! No Python or ML PhD needed. I rewrote GPT-2 in pure vanilla JS so you can step through every line and understand how transformers work. Catch the full @aiDotEngineer World’s Fair workshop: youtube.com/watch?v=ZuiJjk…
This is the best talk on LLM's-> by @ianand
Ai code JS by @ianand Speech/talk/presentation Model loaded in a browser Hands on code Spread sheets Debug LLM youtu.be/ZuiJjkbX0Og
Speaking with @ajcwebdev about vibe coding!
Welcome to our live show! x.com/i/broadcasts/1…
Today we release DeepSeek-TNG R1T2 Chimera. This new Chimera is a Tri-Mind Assembly-of-Experts model with three parents, namely R1-0528, R1 and V3-0324. R1T2 operates at a sweet spot in intelligence vs. output token length. It appears to be... * about 20% faster than R1, and…
AI Tinkerers Seattle – June Meetup - TSK enables a "lead engineer + AI team" workflow - MIT Licensee ** launched today ** by Danny Tormoen
just implement the transformer in excel
🚨 Attention Spreadsheets-are-all-you-need fans! 🚨 I've ported my Excel-based GPT-2 implementation entirely to the browser. Link in the comments.
Ishan Anand (@ianand) is coming to #CascadiaJS to demystify LLM internals for web developers using a real LLM implemented entirely in vanilla JavaScript that runs in the browser. No Ph.D. needed! 🤓 #JavaScript #LLMs
Chrome is like my phone. I rarely restart either and doing so to install updates is a pain.
The argument about whether Vibe coding is real coding is like debate from 20 yrs ago about whether DJs are musicians. cc @ianand
Unit tests are like a driving exam. You expect to pass or fail. Evals are like a tape measure. You want it longer than the thing you are measuring.
Engineers think this is great. Every ML, stats, data person knows this is highly suspect (and likely very bad) Don't overfit your evals! Not linking to blog b/c not about shaming. Very common mistake.