AGI-to-ASI Pathways

Sources#

• AI models have likely reached parity with superforecasters on ForecastBench
• From AGI to ASI

Summary#

The spine of the "From AGI to ASI" report (Section 5): four technological pathways by which AI might progress from human-level AGI to ASI, plus a catalogue of frictions that could slow or halt each. The pathways are not mutually exclusive and likely run in parallel, possibly compounding (not just adding). Only the first has historic data to forecast from. Crucially, the report's stance is that the significance of each friction is an open research question — it deliberately refuses to predict which dominate. This is DeepMind's theory-first counterpart to Anthropic's When AI builds itself.

The four pathways (Table 3)#

1. Scaling compute, models & data — continue the exponential that produced today's progress. Main uncertainty: how increases in scale translate into capability (spiky vs. smooth, emergent capabilities, diminishing returns). The quantitative engine and its frictions live in Effective Compute Scaling. The only pathway with historic data to fit forecasts.
2. Algorithmic paradigm shifts — sharp deviations from today's pretrain-transformer-via-log-loss paradigm (the report distinguishes evolutions — context/recurrency, continual learning, world models, linear-time architectures like Mamba/S4, tool-augmented planning — from true shifts like neuromorphic/analog/spiking or RL-pretraining). By nature unpredictable, so forecasts past a shift are "vacuous"; the antidote is advancing paradigm-agnostic theory (Universal AI (AIXI)).
3. Recursive (self-) improvement — AI accelerates AI R&D, plausibly compounding into an intelligence explosion. No historic precedent to fit. See Recursive Self-Improvement for the mechanism and Intelligence Explosion Dynamics for the growth dynamics.
4. ASI via group agent formation — superintelligence as an emergent collective property of many coordinated AGI agents (orchestrated or self-organizing markets). Emergence in complex multi-agent systems is poorly understood. See Multi-Agent Collective Intelligence.

The six frictions (Table 4)#

For each, the report names the friction and the factors that might counteract it — the balance (and how it shifts with scale) is the open question:

• Data wall — running out of high-quality pretraining data. Countered by synthetic/simulated/self-generated data and data-efficiency paradigm shifts. (Detail in Effective Compute Scaling.)
• Economic & natural-resource demand grows too fast — investments, chips, energy, datacenter siting, rare earths can't be sustained. Countered by rising AI economic returns, AI-driven efficiency gains, and infrastructure build-out.
• Neural paradigm is insufficient — AGI may be unreachable with pretrained nets + post-training + test-time scaling + SGD. Countered by continued research (even sub-AGI AI accelerates it) toward evolutions and shifts.
• Research gets harder — Bloom et al.'s "ideas are getting harder to find" (keeping Moore's law needs ~18× more researchers than the 1970s). Countered decisively by cheap artificial researchers: multiplying digital researchers 20× takes hours-to-weeks vs. years to train humans — likely only a minor friction unless progress stops before AI researchers are useful.
• Abstraction barrier — AI trained on human concepts may be unable to discover novel primitives from raw data; introduces a physical, real-time slowdown via the embodied bottleneck. Countered without being solved: even if the barrier caps any single system near human level, continued scaling and group agent formation could push collective capability past AGI — so the table's own counter routes the block into pathway 4 rather than removing it; addressing it directly is said to need a paradigm shift (interactive learning & RL). See The Abstraction Barrier.
• Deliberate slowdown / regulation / societal backlash — rogue use, accidents, military/political abuse, or backlash could cap progress. Countered (overridden) by economic/military race dynamics, regulatory arbitrage, and "military–economic adaptationism" (Dafoe) under international anarchy — making multilateral coordination "elusive, perhaps unrealistic." See Frontier Pause Verification and Responsible Scaling Policy Evaluations.

Forecasting & benchmarking-beyond-human#

Reducing uncertainty along the pathways requires two new disciplines the report repeatedly calls for: quantitative forecasting (couple effective-compute growth to capability and macroeconomics — Epoch's GATE model, Davidson et al.'s explosive-growth model — with ensembling and continual re-estimation) and benchmarking beyond AGI (current benchmarks saturate at human level: GPQA, SWE-bench, FrontierMath; needed are non-saturating, low-human-input methods — multi-agent/zero-sum competition, setter-solver auto-design, general compression benchmarks, indirect economic-productivity measures, and "multi-agent scaling laws"). Both are paired with benchmark stitching (Ho et al. 2025) for cross-model extrapolation.

Connections#

• Artificial Superintelligence (ASI) — the destination these pathways lead to; supplies the AGI/ASI characterization

• Effective Compute Scaling — pathway 1's quantitative engine and its data-wall/economic frictions

• Recursive Self-Improvement — pathway 3's mechanism (Anthropic's framing); the DeepMind report is its theory-first sibling

• Intelligence Explosion Dynamics — what happens if pathway 3's loop compounds super-exponentially

• Multi-Agent Collective Intelligence — pathway 4 in full

• The Abstraction Barrier — the friction most likely to be a fundamental (not merely slowing) blocker

• Universal AI (AIXI) — the paradigm-agnostic theory that hedges against pathway 2's unpredictability

• Frontier Pause Verification — the deliberate-slowdown friction made into a governance research agenda

• Responsible Scaling Policy Evaluations — institutionalized gates (evals, licensing, incident reporting) are the slowdown friction in practice

• Task Time-Horizon Scaling — a concrete capability trendline feeding the forecasting agenda

• Fundamental Limits of ASI — the hard outer fence on every pathway; why capability forecasting must be empirical-first rather than deduced

• The Data Wall and the Validation Commons Are One Supply Constraint — the friction list revised on its first entry. The corpus's measured side (CS329A's self-training results, compiled 2026-08) does not support demoting the data wall into compute: synthetic supply is rationed by whether a verifier exists, how fast it runs, and whether the generator stays diverse — so this friction merges into the verification/oversight friction rather than dissolving, which is a tightening of the six-item list rather than an addition to it

• Measuring Beyond Accuracy Saturation — where this page's benchmarking question gets its first concrete instance: a benchmark that is non-saturating in score and saturating in comparison, because its human reference class is the thing the headline is made of

• Safety Commitments That Cannot Bind the Actor Who States Them — a seventh friction the report concedes and omits — alignment difficulty as a direct bottleneck on automated R&D — which converts into the verification friction already ranked first, and gates pathway 3 specifically

Open Questions#

• For each friction: is it a fundamental blocker (multi-year plateau) or a mere friction (slows, doesn't halt)? The report's central unresolved question. Partially answered (synthesis against Anthropic): RSI Growth Curves: Which Friction Binds First? — data-wall and research-gets-harder demote themselves into compute; economics and neural-paradigm are pathway-conditional; the abstraction barrier is the candidate fundamental (re-pacing) blocker; and deliberate slowdown is the only exogenous friction — the one Anthropic wants to install and this report doubts can be made to bind. Retagged #oq/now → #oq/source 2026-08-10: the synthesis over existing pages has been run, and what remains is a weight DeepMind itself calls "an open research question" — it needs external evidence, not another /query.
• Do the four pathways compound multiplicatively when run in parallel, and how would we detect that early?
• Can benchmarking methodology that doesn't saturate at human level be built before it's needed for ASI? Partially answered (2026-08-12) — the corpus now has an instance, and it splits the question in two. ForecastBench is non-saturating in every way the question's authors probably meant: its ground truth postdates the question so contamination is impossible, its supply auto-refreshes from live markets and time series, its Brier-type score has no ceiling short of the world's own noise, and its reference class is elite superforecasters rather than the median human. AI submissions reached that reference class in July 2026. It still stops discriminating there — the leaderboard emits significance verdicts for "superforecasters beat the model" and "model beats the public" and has no column for the reverse, and its human baseline is a 2024 elicitation extrapolated forward. So the buildable half is the task family; the unsolved half is the anchor, and the only remedy anyone has proposed is periodically re-eliciting humans, which does not survive contact with the ASI case by construction. Full treatment on Measuring Beyond Accuracy Saturation. Kept #oq/wait: whether a human-free anchor arrives before it is needed is still a claim about the future.

Sources#

• From AGI to ASI — Section 5 (pathways & bottlenecks), Tables 3 & 4, Section 7.1 (research agenda). Table-parse warning, in this page's convention: the PDF text extraction interleaves Table 4's Description and "Countered by" columns line-by-line in two rows (Data wall, Abstraction barrier) — cells run together mid-word (Synthetic data, high-fidelity simulaRunning out of sufficient…). Both rows were untangled against §5.5's prose, which restates each friction and its counter at length; nothing is dropped, only the column mapping was lost. The Abstraction-barrier counter was reconstructed on the 2026-08-19 lint pass, having been the only one of the six missing here
• AI models have likely reached parity with superforecasters on ForecastBench — Forecasting Research Institute (Substack, 2026-07-16, empirical): cited here only against the non-saturating-benchmark question. Leaderboards are PNG screenshots re-read at zoom under the image two-pass rule; the one-way significance apparatus is read from the column headers, not from the prose