A few words about AI

before we start with algorithms

ADS2 · 5. 10. 2026 · Martin Koutecký

AI is here

The three AI pills
AI-pilledknow what AI is and can do today, and act like it
AGI-pilledexpect AGI to arrive, and act like it
ASI-pilledexpect ASI to arrive, and act like it

Get at least AI-pilled!

Zvi Mowshowitz, The Three AI Pills (Aug 2026)

Don't listen to people with bad track records

  • Example: Ed Zitron, well-known AI skeptic
  • Dan Luu went through his dated predictions:
    27 predictions, 26 wrong (1 "possibly right")
  • "generative AI is a dead-end technology that has peaked" — Aug 2024
  • Gemini at 500M users is "so unrealistic" — Feb 2025 (it reached 750M)
  • People who don't want to take the AGI/ASI pill say it doesn't exist.
    That doesn't make it not exist.

Dan Luu, Ed Zitron's AI prediction track record

What I learned building Tine

  • I built Tine, a note-taking app (Logseq-compatible), with AI
  • 450+ GitHub stars
  • 4.5k commits, ~500 branches, ~460k lines of code

… and I did not see a single line of the code.

The Bitter Lesson

Music video by Ethan Mollick (made with Claude + Suno), from The Dot and the Swarm · Rich Sutton, The Bitter Lesson (2019): "general methods that leverage computation are ultimately the most effective, and by a large margin."

The Overhang

Models can do much more than almost anyone is doing with them.
What you bring:

Deep knowledgeexpertise in a field: you can judge quickly and correctly
Wide knowledgeknowing which ideas from elsewhere apply (AI knows them, but won't bring them up unless asked)
Tastewhich outputs to keep, which to throw away
Agencytrying things yourself instead of waiting to be told

Ethan Mollick, The Overhang (Sep 2026)

AI: good or bad for learning?

  • AI is very sensitive to how well you communicate — and that takes deep + wide knowledge
    • learn to communicate well → work with other students
    • use the TAs! They are great and care about you a great deal
  • AI can easily make you feel like you understand without any understanding
    • AI will always be happy to oblige — that doesn't mean it does the job
    • apply agency: decide if / when / how to use AI for learning
    • test: can you explain it / solve the next one without AI?

AI may kill us

  • AI 2027 — for real. Again: listen to people with a track record
    • Kokotajlo's 2021 forecast predicted chain-of-thought, inference scaling, chip export controls, $100M training runs
  • We could do some research on AI generalization — talk to me if you are interested
  • Call your representative if you are American, and pray if you are not

Takeaways

  • Get AI-pilled: find out what it can do today
  • Learning this deeply pays off whether AI stalls, keeps improving, or goes all the way — in the last case, because someone has to understand what we built
  • Use AI deliberately; don't let it fake your understanding

Now: algorithms.