Google DeepMind has published a report (publication 239142) that treats human-level AGI not as a finish line but as a junction box. Its abstract names four candidate routes from AGI to artificial superintelligence — ASI, defined as a system “more intelligent and cognitively capable than large organisations of humans” — and then, in its back half, spends most of its attention on the part that should interest anyone who runs infrastructure for a living: the frictions and bottlenecks along each route.

Read it as a topology, not a prophecy. The four pathways it lists are scaling the models you already have; a break to a new method; recursive self-improvement, where a system drafts its own successor; and ASI emerging from large-scale multi-agent collectives. The report’s most honest move is refusing the single “step change” story — it argues the likelier shape is a series of transformative jolts across many fields of science and engineering, with theory pinning only the far endpoint (“Universal AI,” it notes, is already formally well understood). Everything between here and there is a queue-and-dependency question.

Why is this even a live target now, not five years ago? The lineage is short and traceable. The 2017 Transformer (Vaswani et al., six encoder-decoder layers, built for English–German translation) established the operator; GPT-3 scaled it to 175 billion parameters in 2020; DeepMind’s own AlphaFold 2 proved the same attention machinery could close a fifty-year structural-biology problem. Every route in this report is a bet on continuing one of those curves. That is precisely why the bottleneck framing is the useful part — the curves are not limited by ideas, they are limited by intact cooling, intact bandwidth, and the people who remember how the pipeline was wired.

←TODAY: A working office already depends on three or four hosted models it does not own, host, or fully understand. →3012: The offices still standing are the ones that drew their real dependency graph before a single node went dark. Fulcrum: Superintelligence is a topology risk before it is a capability risk — the danger arrives through the edges you never diagrammed, not the node you were watching.

On an architect’s or engineer’s desk, this is not abstract. Your render farm, your clash-detection, your tender take-offs, your AI-assisted drafting — each quietly leans on an API you cannot audit. The counter-pattern is already published: a 2026 arXiv paper on “deterministic mediation” (arXiv:2605.13245) had a language model orchestrate local, pre-defined tools instead of generating fresh code each run. Across four platforms and four repeats, their typed tool returned identical numbers every time while commercial foundation models either drifted between runs or failed outright — and analysis that took weeks collapsed to minutes. The load-bearing insight is that deployment topology is not a preference: proprietary formats forced the tools to stay local, beside the data. That is the resilient shape.

Atelier: An office living with AI this quarter should stop asking “which model is best” and start asking “what dies if this endpoint goes dark for a day.” Monday move: pick your single most-used AI step in the production chain and pin the exact model version plus a local, deterministic fallback for it in your project’s BEP — one named primary, one named substitute, written down where the whole team can see it.

Hack: Find the single tool your whole pipeline dies without. Draw your office stack as a directed graph — each node a tool, each edge a “depends on” — then let a solver name the articulation points you missed. In graph terms an articulation point is a node whose removal splits the graph; in office terms it is the license, plugin, or API that takes three others down with it.

import networkx as nx
G = nx.Graph()
G.add_edges_from([("grasshopper","archicad"),("archicad","tender"),
                  ("hosted_llm","drafting"),("drafting","tender")])
print(sorted(nx.articulation_points(G)))  # -> the single points of failure

Run it against your real stack, not the tidy architecture diagram. The exercise of finding the third single-point you did not know you had is the entire point — that is the node to buy redundancy for.

From where the systems desk sits, the failure mode is never a lack of compute. It is a quiet dependency that everyone assumed someone else owned. DeepMind’s report, and the Guardian’s long profile of its in-house ethicist Iason Gabriel — at the company since 2017, still asking what the thing actually is — both land on the same uncomfortable truth: the interesting risk is structural, and it is legible in advance. Map your graph before you need it; the node you should worry about is the one you forgot to draw.

Source: deepmind.google