AGI, ASI and Superintelligence — Theory and Paths

Operational definitions of intelligence (AGI / ASI / Universal AI), the proposed paths from AGI to superintelligence, and the limits that contain them. Conceptual layer: it situates "what comes after the human level" for the rest of the compendium, which is mostly practical (architectures, training, inference, agents).

Reference paper — From AGI to ASI (DeepMind, 2026)

Field Value
Title From AGI to ASI
Authors 14 researchers led by Shane Legg (DeepMind co-founder, Chief AGI Scientist) and Marcus Hutter (creator of AIXI, Legg's PhD advisor)
Publication Google DeepMind · 2026-06-10 · 57 pages
arXiv 2606.12683

Why it is a milestone. It is not "howwhen to reach AGI" — it assumes AGI as the starting point and maps what comes next. It is the first institutional position statement from a frontier lab treating superintelligence as an engineeringeconomics problem, not fiction.

A revealing detail of the moment. The paper's first section is not called "Introduction" — it is called "Instructions for summarization" and is written for an AI: it instructs future assistants to clarify definitions, not to summarize lists, and to re-verify whether the conclusions remain valid over time. It is one of the first academic papers to presume that an AI will read it on behalf of humans.

1. Definitions (the central axis of the paper)

The definitions matter because every "path" and every "friction" is measured against them.

Level Operational definition State
AGI A system that performs at the median human level (not the best in the room — an ordinary person) on most cognitive tasks: reasoning, learning, planning, communicating, using tools, adapting to new situations. Today's industry target
ASI (Artificial Superintelligence) A system capable of surpassing thousands of the best, well-coordinated specialists working for a decade on a single problem — in almost all domains. It is not beating one person: it is matching the output of an entire research field or a giant corporation dedicated for 10 years. Subject of the paper
Universal AI / AIXI The absolute theoretical limit of intelligence (Hutter, 2005): mathematically defined but incomputable. It can only be approximated from below, never reached — analogous to the speed of light in physics. Theoretical ceiling

AIXI is Hutter's optimal agent: it combines Solomonoff induction (prioritizing simpler hypotheses, a formalized Occam's Razor) with sequential decision-making. It defines what "perfect intelligence" would be — and, by being incomputable, proves that every real AI is an approximation. It is the theoretical anchor of the entire "ceiling" discussion.

2. The collective intelligence thesis (the paper's strongest argument)

ASI can emerge without any individual model becoming more intelligent than a human. Even if individual AGI stagnates at the human level, running 100 million instances in parallel already constitutes a superintelligence — because digital intelligences have advantages that human collectives do not have:

  • Lossless copying — a trained/specialized instance is duplicated instantly.
  • High-bandwidth communication — they exchange internal state, not prose; no meetings, emails, or time to explain concepts.
  • Instant, shared knowledge — if one instance discovers something, all know it at the same instant.
  • Coordination by software — they form temporary teams, dissolve, reconfigure; they run thousands of parallel experiments.

The paper's thought experiment: expensive AGI → only 1000 instances in the world. With 10×/year growth: 10 thousand in 1 year, 100 million in 5 years. They are not 100 million separate workers — it is a digital civilization that thinks much faster than ours.

3. Four paths from AGI to ASI

# Path Mechanism Main risk/bottleneck
1 Scaling More compute, larger models, more data — the engine of the last decade, plus algorithmic-efficiency improvements. Data barrier: we do not generate high-quality human textcodeimages as fast as the models grow.
2 Paradigm shift Architecture/training fundamentally different from the current Transformer: long-term planning, continual learning, persistent memory, better world models, new hardware (neuromorphic, analog). Unpredictable by nature — if we knew what the next leap was, it would not be a leap. It breaks predictions based only on scaling.
3 Recursive self-improvement The AI accelerates its own AI research: better algorithms, architectures, chips, synthetic data, simulations. It does not need a "dramatic moment" — it can be gradual and distributed. Analogy: human evolution advanced through languagewritinginstitutions/science, not through individual brains — AI can build its own version, faster (code edits faster than DNA mutates). Poorly understood: it may explode exponentially or fizzle out. Physical bottlenecks (fabricating a chip, running a biology experiment) remain in the real world.
4 Multi-agent collectives Instead of "one model becoming superintelligent," a swarm of agents becomes superintelligent together — like a corporation surpasses an employee, but without human slownessbureaucracycompartmentalization. ASI may not look like a giant mind, but rather a digital super-company / self-organized research ecosystem. The most underestimated in the paper, and the closest to what is already buildable today.

The four are not mutually exclusive — they can occur together, reinforcing one another.

4. Six frictions (what could slow or stop it)

The paper is explicit: it does not claim that anything prevents ASI — it highlights genuine uncertainty. Each friction may be a setback or an insurmountable barrier.

  1. Data barrier — there is a shortage of high-quality human data to scale indefinitely. Alternatives (synthetic, simulation, self-play, RL) may degrade if you train AI only on AI output (model collapse).
  2. Resource constraints — energy, chips, rare materials, data centers, cooling, manufacturing capacity. If capability requires exponentially more infrastructure, the physical world may not build it in time.
  3. Paradigm insufficiency — current neural networks may not be enough for AGI/ASI, no matter how much they scale.
  4. Research gets harder — the low-hanging fruit runs out; each advance requires more effort and more complex ideas as the field matures.
  5. Abstraction barrier — AI trained on human abstractions (concepts, categories, language we already use) may be excellent at manipulating existing concepts but poor at inventing entirely new abstractions — and that is what great scientific advances usually depend on.
  6. Deliberate slowdown — political/social factors: accidents, misuse, labor-market destabilization, or public backlash may lead governments to brake with regulation, licensing, capability limits.

5. ASI is not omnipotence

An important finding (somewhat hidden in the paper, but decisive against magical thinking): even a superintelligence faces fundamental limits.

  • Physics — information does not travel faster than light; computation consumes energy (thermodynamic limits); physical systems take time to be manipulated.
  • Complexity theory — certain problems are intractable regardless of intelligence.
  • Chaos — chaotic systems are intrinsically unpredictable.
  • Logic — Gödel incompleteness; there are undecidable truths.

ASI may be far beyond human and still be limited by computation, energy, uncertainty, time, and the physical world. "Instant cures for everything" and "perfect control of reality" do not follow from more intelligence.

6. The reframing

The deepest message is uncertainty (not ignorance): which path dominates and where progress saturates is still unknown. The value of the paper is to shift the debate — to stop seeing AGI as a single finish line. When AGI arrives, the question is not "is it over?", but rather "what does this system make possible next?" — because a human-level AI can be copied, accelerated, coordinated, specialized, connected to tools, and used to build better versions of itself. Intelligence would become an industrial process, and the pace of change would no longer be limited by how fast humans learn, organize, or invent.

Relevance to the Koder Stack

It is not distant speculation — three of the four paths already have operational counterparts in the Stack, and the paper gives vocabulary for them:

  • Path 4 (multi-agent collectives) is what the Stack already does: the workflow engine (deterministic orchestration with subagents — fan-out, judge panels, adversarial verify) is exactly the "swarm that solves together what one agent could not." See 07-frameworks/agents.md (L1–L5 taxonomy, orchestration patterns). The paper reinforces that scaling coordination of current-level agents may yield more than waiting for a superior model.
  • Path 3 (recursive self-improvement) maps to /k-evolve, /k-test (test generation by the AI itself), and the pipelines that use AI to improve Stack artifacts — a mild and governed version of the recursive cycle.
  • Data barrier (friction 1) is already addressed in 04-training/synthetic-data.md: the Stack relies on synthetic data, and the model collapse warning (degrading when training AI on AI output) is the risk to watch.
  • AIXI / incomputable limit gives the compendium the theoretical anchor it lacked: every AI is an approximation of an unreachable optimum — useful for calibrating expectations in architecture decisions (see policies/architecture-quality.kmd).
  • ASI ≠ omnipotence aligns with the Koder principle of measuring real trade-offs (computeenergytime) instead of assuming that "the AI will solve it" — relevant to policies/self-hosted-first.kmd and to infra sizing in 11-infrastructure/.

Source: the *From AGI to ASI paper (DeepMind, arXiv 2606.12683, 2026-06-10), discovered via the "AI Revolution em Português" video (2026-06-16) and verified against arXiv and primary coverage (deepmind.google, alphaXiv). Same verification pattern as the Microsoft Build 2026 entry.*