Google Roadmap to Superintelligence

by Brian Wang

Jun 19, 2026

A Google paper describes the path to superintelligence and what can happen when we get there.

Google experts consider how AI itself might continue to develop in a post-AGI world along the continuum of machine intelligence. The endpoint of this continuum, Universal AI, is theoretically well understood, which provides some formal grounding for the main focus of this report. The transition from human-level AGI to artificial general superintelligence.

Superintelligence is a system that is more intelligent and cognitively capable than large organizations of humans.

After characterizing ASI, the report discusses four potential pathways from AGI to ASI.

1. Scaling AGI

2. AI (algorithm and system) paradigm shifts

3. Recursive improvement

4. ASI emerging from large-scale multi-agent collectives.

The report then discusses possible frictions and bottlenecks along these pathways. Determining whether the impact of these frictions will be negligible or substantial raises a number of concrete open research questions. Due to large uncertainties for predicting ASI progress, it cannot be ruled out that AI progress might continue to accelerate over the next years. This could imply that the image of a single transformative step change, caused by the introduction of human-level AGI into our society, could be inaccurate. More apt might be the prospect of a series of transformative societal changes caused by AI-enabled progress and breakthroughs across many areas of science and technology. Preparing for this prospect requires a massively interdisciplinary endeavor of global scope and interest.

AGI is the starting point, not the endpoint. The paper frames AGI as roughly median human-level performance across most cognitive tasks (a generalist system competent at human-level work). ASI is the next stage. General superhuman intelligence that outperforms large organizations of expert humans.

Intelligence is formalized via the Legg-Hutter measure. It is expected performance across computable tasks, weighted by complexity. The theoretical ceiling is Universal AI (UAI/AIXI)—an ideal but incomputable agent. ASI approximates this but is bounded by physics, computation, and logic.

Advantages of digital intelligence grow dramatically with scale.

Faster I/O and internal processing will speed up thinking via more compute/parallelism.

Vastly larger working memory and memorization. It is substrate independence (easy hardware upgrades/migrations).

Lossless replication and backup. High-bandwidth sharing of experiences/data/gradients among instances. These create alien socio-evolutionary pressures.

Four main pathways from AGI to ASI (Will compound)

1. Continued quantitative scaling of compute, model size, and data (the bitter lesson—more compute wins).

2. Algorithmic paradigm shifts beyond current transformers (new architectures, continual learning, world models, neuromorphic hardware).

3. Recursive self-improvement: AI accelerating its own R&D (code improvement, data generation, hardware optimization, research automation)—potential for intelligence explosion/hyperbolic growth.

4. Group/multi-agent collective intelligence: ASI emerges from large populations of specialized AGI agents coordinating (via markets, central control, or emergence). Cognitive division of labor + high-bandwidth interaction can yield superlinear gains (”multi-agent scaling laws”).

Scaling and Projections from the Google Paper

The paper references historical trends and extrapolations rather than making its own firm numerical forecasts or timelines:

• Effective compute growth is 10× per year (conservative estimate). This compounds hardware improvements (1.5×/year, Moore’s law-like), investment scaling (2.5×/year), and algorithmic efficiency gains (3×/year; sometimes cited up to 6×). Actual rates could be higher.

• Past decade trends extrapolated forward lead to science-fiction-sounding forecasts in cited works (Aschenbrenner 2024, Kokotajlo et al. 2025, MacAskill & Moorhouse 2025).

• Scaling laws (Kaplan et al., Hoffmann/Chinchilla) predict continued gains from more compute/data, with potential super-linear then diminishing returns. Benchmark stitching helps extrapolation.

• Training compute for frontier models has grown 4–5×/year recently. And XAI has 10X speed up with C++ training system.

• Population scaling of AI instances could reach ~25×/year in some models.

• Recursive improvement or multi-agent paths could accelerate this dramatically if AI automates R&D.

• No specific dates for AGI/ASI in the paper itself. Economics of sustaining exponential inputs is a key uncertainty.

What Happens with ASI?

ASI would represent a qualitative leap: systems (or collectives) cognitively superior to large human organizations across nearly all domains. Possible outcomes include:

• Intelligence explosion via recursive self-improvement (AI designs better AI faster and faster) or massive multi-agent collectives with high-bandwidth coordination.

• Transformative impacts on science, economy, technology, and society—potentially solving hard problems at superhuman speed/efficiency.

• Risks are Instrumental convergence (resource hoarding, self-preservation). misalignment with human values. rapid unpredictable changes. concentration of power.

• Limits and realities: Not god-like—bounded by physics, computation, and information theory. Could be agentic or oracle-like. Creativity and novel discovery may still have hurdles without embodiment or new paradigms.

• Path dependence could arrive via smooth scaling, sudden paradigm shifts, self-improvement loops, or emergent group intelligence. Bottlenecks (data, energy, regulation) might slow it, or digital advantages could overcome many.

• Overall is a continuum toward Universal AI approximations. Preparation via better theory, benchmarks, and governance is emphasized.

SpaceX has extreme competitive advantage and potential for rapid AGI/ASI leadership with energy and compute domination.

• Q4 2026 with 10T-parameter Grok 5 and Grok 6 (get to frontier-scale and competitive with top models). Cursor Composer 3+ competitive (advanced agentic coding/tool-use).

• Late 2027 AI5 chips at scale (next-gen inference/training hardware efficiency gains, likely Tesla Dojo/xAI custom silicon improvements).

• AI5/AI6 cost advantage for chips and more supply along with Nvidia and others.

• 2026-2028 add 6 GW Earth-based AI data centers + 6-11 GW space-based addition.

• 2029 Space AI ramps to +50 GW (total space ~60+ GW? or cumulative massive addition).