I stopped using Microsoft Office bundle: - slides --> LaTeX Beamer - docs --> Markdown - excel --> csv/json + python (matplotlib) The three replacements are text only, so I can use LLM Agents to work on them Claude Code + Cursor is my new UI for everything

No!!! EC!! U deserve better!! Revive the glory of EC!!!

Kaiming once used a similar analogy: one trust turbojets not because we figured out aerodynamics or solved Navier-Stokes equations, but we tested the turbojets tens of millions of times

The “so done” and “so back” oscillations

Is the difference led by the implicit bias induced by length normalization?

That means 5-TPG would win 42 of 44 trades Guys, market might have been solved

Does MuP just work for MoE? I suspect the top-k operation increases variance, so maybe extreme value theorem shall be considered Maybe shift init further by 1/log, 1/loglog, or pi^2/6

Why Claude for Excel, When you can Claude Code + matplotlib + .csv?

I won't be surprised if Gemini is actually expert choice

TL;DR: Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3 Qwen3… (*38)

A reminder that 11 months have passed, and we still have no open-source implementation of o1-pro A multi-agent long reasoning framework that can use 100x test time compute to produce well-thought results

Have you noticed that, chatbots never say "sorry"?

It’s a good mental model to think model training cost is essentially 0 It gets amortized by the increasing demand One should only care about inference cost in the long run

I remember seeing the demo during NeurIPS 2023 Still don’t see the point

❌ Conference shitty peer review ✅ Decentralized public voting system, (e.g., huggingface daily🤗) ✅✅✅ Recommendation system for papers based on engagements I thought we solved decentralized review a while ago in social media

writing the intro of an ML paper be like: