https://laurenswhipple.com/ml/agi-to-asi/ A guided walkthrough of the DeepMind report "From AGI to ASI" (Genewein et al., June 2026). en-us Sat, 13 Jun 2026 06:02:35 +0000 https://laurenswhipple.com/ml/agi-to-asi/ https://laurenswhipple.com/ml/agi-to-asi/ Sat, 13 Jun 2026 06:02:35 +0000 A guided path through the DeepMind report (Genewein et al., June 2026). Read in order; each note links forward and back. https://laurenswhipple.com/ml/agi-to-asi/01-the-big-picture.html https://laurenswhipple.com/ml/agi-to-asi/01-the-big-picture.html Sat, 13 Jun 2026 06:02:35 +0000 A position report from inside DeepMind, with Hutter (AIXI), Legg (Legg-Hutter score), Orseau, Leibo (multi-agent), Dafoe (governance), Graepel and others. It is **not** a technical paper proposing a method. It is a **landscape map** of what could come after human-level AGI, written by people who built much of the conceptual vocabulary the field uses. https://laurenswhipple.com/ml/agi-to-asi/02-intelligence-continuum.html https://laurenswhipple.com/ml/agi-to-asi/02-intelligence-continuum.html Sat, 13 Jun 2026 06:02:35 +0000 The paper deliberately refuses sharp thresholds. It uses the **Legg-Hutter score** (average performance across all computable tasks, weighted by inverse Kolmogorov complexity — see 05 - Universal AI (AIXI), informally) as a *continuum*, then picks coarse labels along it. https://laurenswhipple.com/ml/agi-to-asi/03-effective-compute.html https://laurenswhipple.com/ml/agi-to-asi/03-effective-compute.html Sat, 13 Jun 2026 06:02:35 +0000 The report's quantitative forecasting hinges on one estimate from Epoch AI: **effective compute grows by ~10× per year.** https://laurenswhipple.com/ml/agi-to-asi/04-digital-advantages.html https://laurenswhipple.com/ml/agi-to-asi/04-digital-advantages.html Sat, 13 Jun 2026 06:02:35 +0000 This sounds banal. It is not. Every advantage that follows is a direct consequence of being *code*, and every one of these advantages **intensifies with more compute**. Biological humans cannot scale into any of them. https://laurenswhipple.com/ml/agi-to-asi/05-universal-ai.html https://laurenswhipple.com/ml/agi-to-asi/05-universal-ai.html Sat, 13 Jun 2026 06:02:35 +0000 To set a **theoretical ceiling** on the AGI→ASI discussion. You can either project from below (extrapolate trends) or bound from above (formal limits). AIXI is the upper bound; ASI lives somewhere between today and AIXI on the continuum (02 - The Intelligence Continuum). https://laurenswhipple.com/ml/agi-to-asi/06-pathway-scaling.html https://laurenswhipple.com/ml/agi-to-asi/06-pathway-scaling.html Sat, 13 Jun 2026 06:02:35 +0000 Just keep doing what works. Bigger transformers, more data, more compute, more test-time thinking. This is the only pathway with **historical extrapolation data** (03 - Effective Compute — 10x per year). https://laurenswhipple.com/ml/agi-to-asi/07-pathway-paradigm-shifts.html https://laurenswhipple.com/ml/agi-to-asi/07-pathway-paradigm-shifts.html Sat, 13 Jun 2026 06:02:35 +0000 A "true" paradigm shift = a dramatic architectural or training change that **isn't just a smooth evolution** of pretraining + fine-tuning + test-time scaling. https://laurenswhipple.com/ml/agi-to-asi/08-pathway-rsi.html https://laurenswhipple.com/ml/agi-to-asi/08-pathway-rsi.html Sat, 13 Jun 2026 06:02:35 +0000 AI accelerates AI R&D, which produces better AI, which accelerates AI R&D further. This is the classical **intelligence explosion** hypothesis (Good 1965, Solomonoff 1985, Kurzweil, Bostrom). https://laurenswhipple.com/ml/agi-to-asi/09-pathway-multiagent.html https://laurenswhipple.com/ml/agi-to-asi/09-pathway-multiagent.html Sat, 13 Jun 2026 06:02:35 +0000 This is the Leibo / Dafoe / Tomašev fingerprint. It is the pathway the report treats with the most fresh thinking, and the one where DeepMind's own recent papers (Virtual Agent Economies, Distributional AGI Safety, Intelligent AI Delegation) are cited. https://laurenswhipple.com/ml/agi-to-asi/10-bottleneck-data-wall.html https://laurenswhipple.com/ml/agi-to-asi/10-bottleneck-data-wall.html Sat, 13 Jun 2026 06:02:35 +0000 Running out of high-quality text to pretrain on. Model size growth outpaces global production of novel text. Villalobos et al. (2024) estimates this happens *within the current decade*. https://laurenswhipple.com/ml/agi-to-asi/11-bottleneck-economics.html https://laurenswhipple.com/ml/agi-to-asi/11-bottleneck-economics.html Sat, 13 Jun 2026 06:02:35 +0000 Sustained 10×/yr effective compute requires sustained **2.5×/yr investment growth**, plus matching hardware production, plus matching energy supply, plus matching rare earth and water inputs. https://laurenswhipple.com/ml/agi-to-asi/12-bottleneck-research-harder.html https://laurenswhipple.com/ml/agi-to-asi/12-bottleneck-research-harder.html Sat, 13 Jun 2026 06:02:35 +0000 Bloom et al. (2020): in mature fields, productivity per researcher declines. **Keeping up Moore's law today requires ~18× more researchers than in the 1970s.** https://laurenswhipple.com/ml/agi-to-asi/13-bottleneck-abstraction-barrier.html https://laurenswhipple.com/ml/agi-to-asi/13-bottleneck-abstraction-barrier.html Sat, 13 Jun 2026 06:02:35 +0000 Computation alone (the argument goes) cannot **instantiate or discover novel conceptual primitives without an experiencing agent to map physical reality to symbols and back.** https://laurenswhipple.com/ml/agi-to-asi/14-bottleneck-slowdown.html https://laurenswhipple.com/ml/agi-to-asi/14-bottleneck-slowdown.html Sat, 13 Jun 2026 06:02:35 +0000 Capability progress might be **deliberately capped** by: - Public backlash from visible AI harms or accidents - Government regulation (EU AI Act, compute-threshold licensing, mandatory eval) - Voluntary moratoria from labs - Liability regimes shifting after high-profile failures - Geopolitical coordination (unlikely but possible) https://laurenswhipple.com/ml/agi-to-asi/15-creativity.html https://laurenswhipple.com/ml/agi-to-asi/15-creativity.html Sat, 13 Jun 2026 06:02:35 +0000 Does more intelligence → more creativity? The paper takes this question more seriously than most. https://laurenswhipple.com/ml/agi-to-asi/16-goals-agency-alignment.html https://laurenswhipple.com/ml/agi-to-asi/16-goals-agency-alignment.html Sat, 13 Jun 2026 06:02:35 +0000 That admission is itself worth flagging. The whole report is conditional on alignment being solvable. They do **not** prove this. https://laurenswhipple.com/ml/agi-to-asi/17-what-asi-cannot-do.html https://laurenswhipple.com/ml/agi-to-asi/17-what-asi-cannot-do.html Sat, 13 Jun 2026 06:02:35 +0000 The paper is unusually clear that even arbitrarily-advanced AI is bounded by physical, complexity-theoretic, and logical limits. These bounds come from 05 - Universal AI (AIXI), informally (Table 2 in the paper). https://laurenswhipple.com/ml/agi-to-asi/18-open-research-questions.html https://laurenswhipple.com/ml/agi-to-asi/18-open-research-questions.html Sat, 13 Jun 2026 06:02:35 +0000 Section 7.1 of the paper. This is the part that turns the report from a survey into a call-to-action. Worth scanning if you ever want to choose a research direction here. https://laurenswhipple.com/ml/agi-to-asi/19-insights-and-takeaways.html https://laurenswhipple.com/ml/agi-to-asi/19-insights-and-takeaways.html Sat, 13 Jun 2026 06:02:35 +0000 The parts of the report that actually surprised me on a careful read — not the bullet-point summary but the load-bearing ideas. https://laurenswhipple.com/ml/agi-to-asi/glossary.html https://laurenswhipple.com/ml/agi-to-asi/glossary.html Sat, 13 Jun 2026 06:02:35 +0000 Verbatim from Appendix B with page anchors back to the relevant notes.