How AI Learned To Think

From Farzad.

Master Plan: What Elon Musk is Actually Building book: https://a.co/d/04C1qu8Z
Abundance or Collapse book: https://a.co/d/0cQgFdGH

Reach out to Farzad to implement AI in your small business: https://farzad.fm/consulting
Farzad Mesbahi (@farzad-fm) — humanity’s analyst on disruptive innovations and their 1–20 year implications. https://farzad.fm
I unpack why physical AI and humanoid robots like Tesla Optimus matter more than doomsday AI headlines — from car-factory supply chains that already prove the hard parts, to what happens when robots take over the work that built the industrial world, and why almost nobody is pricing how close this actually is.

Want to sponsor my channel? Email: sponsors@dcsocialmedia.co.uk

A computer can check an AI’s math proof line by line, but nobody can check a wedding toast. That one difference explains where AI is superhuman, where it is clumsy, and which work and which companies feel it first.

In this video: how language models learn by imitation, why imitation hits a ceiling, how practice against a machine-checkable answer key (reinforcement learning with verifiable rewards) broke it, and what that means for jobs and for investors.

CHAPTERS
0:00 The flat toast
1:00 How chatbots learn: imitation
2:53 Learning from a score: RLVR and the answer key
6:09 The answer to your guess
7:58 The proof you can check
8:31 Why AI is jagged by design
11:15 Back to the toast

HOW THIS VIDEO WAS MADE
The research, the script, the voice (a clone of mine), the animation and the edit were made by AI. The idea and the analogy are mine.

VIDEO CREDITS
Stock footage via Pexels (Pexels License): "Students Taking Examination in Classroom" by Andy Barbour, plus clips of a smartphone, homework, crumpled paper, folders on shelves, a server and a family dinner toast.
Photo: "Mining operation on Dominion Creek, Yukon Territory" by Eugene Hegg, c. 1898, via Wikimedia Commons (public domain).
All other visuals are original animations made for this video.

SOURCES
– Allen Institute for AI, Tulu 3 (introduces RLVR), Nov 22, 2024: https://arxiv.org/abs/2411.15124
– DeepSeek-R1, Nature, Sept 17, 2025 (R1-Zero, AIME 2024 pass@1 15.6% to 77.9%, GRPO): https://www.nature.com/articles/s41586-025-09422-z
– OpenAI, Learning to reason with LLMs (o1-preview), Sept 12, 2024: https://openai.com/index/learning-to-reason-with-llms/
– OpenAI, o1 system card, Dec 5, 2024
– Google DeepMind, AI solves IMO problems at silver-medal level (AlphaProof, IMO 2024): https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/
– OpenAI, fluid-equations result, Sept 8, 2026; Clay Mathematics Institute statement, Sept 11, 2026: https://www.claymath.org/news/
– OpenAI, InstructGPT (training from human feedback), Mar 4, 2022: https://arxiv.org/abs/2203.02155
– OpenAI, GDPval: https://openai.com/index/gdpval/

Analysis, not financial advice.

MUSIC
"LoFi Sweet song." by u_sr0ywmbcog, via Pixabay (Pixabay Content License).