As Rocks May Think

Bots can code now; commenters see utopia, doom, or a GPU brick wall

TLDR: Author says his AI assistant now codes and runs research end-to-end, hinting at a big shift in how work gets done. Comments erupt: some predict a golden age, others cite chip limits and a 2027 reality check, while many fear billionaire-driven automation will crush wages—stakes feel very real.

The author claims a chatbot sidekick can now code, run tests, and even write research reports on command—think “/experiment,” folders appear, graphs are made, conclusions drafted. He even name-drops tuning tricks like muP and d-muP. Cue the comment section meltdown. One camp cheers a coming “golden age” where machines handle the boring work. Another says slow down, Skynet—this is a hype train with a GPU (the chips that run AI) shortage for brakes. The top skeptic blasts the “automated & luxurious communist utopia” vibe with hard numbers: short chip lifespans, factory limits, and production walls by 2027–2028. Timeline watchers mock the vibe shift too—“AI 2027” is suddenly canon? Bold. Meanwhile, worker-anxiety hits red: a top-rated line warns of a “dark age” where billionaires skip paying human wages entirely. Oof. There’s meta-drama, too: “Where’s the thesis?” asks one, while another dunks, “doesn’t understand how LLMs (chatbots) work.” The vibe: wild optimism vs hardware reality, with labor panic under it all. The post says: bots can think and code now. The comments say: prove it, price it, and please don’t delete our paychecks.

Key Points

  • The article asserts rapid AI-driven changes since 2022, including ChatGPT’s capabilities, AI-enabled cyberattacks, emergence of home humanoids, and widespread humanoid projects.
  • It emphasizes that machines can now code and think well, with the author using Claude Code to automate coding and research tasks.
  • An “/experiment” command standardizes research workflows: creating folders, generating single-file Python scripts, saving data/figures in CSV, and producing report.md summaries.
  • A concrete experiment applies MuP and d‑muP to a GoResNet‑100M model, using Microsoft’s mup package, Ray for parallel jobs, and regular metric evaluation while tuning key hyperparameters.
  • The workflow includes iterative, serial hyperparameter optimization with defined model sizes, FLOP budgets, checkpoints, and reflection to guide subsequent experiments.

Hottest takes

"automated & luxurious communist utopia" — measurablefunc
"we are entering a dark age where the billionaires who control most of the world's capital will no longer need to suffer the indignity of paying wages to humans" — munificent
"This person doesn't understand how LLMs work." — alsetmusic
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.