laurenswhipple / ml / agi-to-asi

From AGI to ASI — Walkthrough

A guided path through the DeepMind report (Genewein et al., June 2026). Read in order; each note links forward and back.

Paper: paper.pdf · arxiv 2606.12683 · 57 pages · 14 authors incl. Hutter, Legg, Orseau, Leibo, Dafoe

The reading path

  1. 01 - The Big Picture — what the paper is actually claiming, and why this one matters
  2. 02 - The Intelligence Continuum — AGI / ASI / Universal AI, and why they refuse sharp definitions
  3. 03 - Effective Compute — 10x per year — the single most load-bearing number in the paper
  4. 04 - Digital Intelligence Advantages — why "human-level" is already a trap
  5. 05 - Universal AI (AIXI), informally — the theoretical ceiling, and what it does NOT tell us

The four pathways

  1. 06 - Pathway 1 — Scaling
  2. 07 - Pathway 2 — Paradigm Shifts
  3. 08 - Pathway 3 — Recursive Self-Improvement
  4. 09 - Pathway 4 — Multi-Agent Collectives ← the most interesting one

The frictions

  1. 10 - Bottleneck — Data Wall
  2. 11 - Bottleneck — Economics & Resources
  3. 12 - Bottleneck — Research Gets Harder
  4. 13 - Bottleneck — Abstraction Barrier ← the most philosophically loaded
  5. 14 - Bottleneck — Deliberate Slowdown

Strange questions the report raises

  1. 15 - Is Superintelligence Super-Creative
  2. 16 - Goals, Agency, Alignment
  3. 17 - What ASI Cannot Do

Closing

  1. 18 - Open Research Questions
  2. 19 - Insights and Takeaways ← the part most worth your time on a re-read
  3. Burnt Strawman ← polemical counter-reading; the parts that don't hold up
  4. _Glossary

One-line summary

The "single transformative step at AGI" mental model is probably wrong. AGI→ASI is a continuum traveled along four overlapping pathways, throttled by frictions whose magnitudes nobody can yet measure.

See 19 - Insights and Takeaways for the parts that actually surprised me.

↑ index The Big Picture →