Trajectory · updated 2026-09-16
From Dartmouth to the agentic frontier, and the gates ahead
The past is compressed; the present and the road to AGI are expanded, because that is where decisions are still open. Future dates are forecaster medians, not our predictions. Every future gate lists the scenarios it opens and the forces that could delay or accelerate it.
Forces on the path · pressure 1–5 · updated weekly from signals
Accelerants
Compute buildoutrising
Big-5 capex above $750B in 2026; Microsoft targeting 38 GW by 2032; GPU utilization near 98%. The strongest accelerant on the board.
Capitalrising
OpenAI in talks at $1.2T; Oracle backlog $664B. Money is not the constraint; its fragility is a separate risk (see Capex Bust).
Algorithmic efficiencyrising
Reasoning models, opaque recurrence, 75% cheaper cache reads: capability per dollar keeps improving independently of scale.
Open weights & distillationrising
GLM-5.3 at 744B; six Chinese labs named for large-scale distillation; Metaculus 95% open cyber parity within a year.
Lab & state competitionsteady
Three frontier releases in one week; a US intelligence advisory framing AI as strategic rivalry. Pacing talk cuts against this for the first time.
Brakes
Compliance & legislationrising
EU high-risk rules deferred to 2027–28; US frontier bill in draft with weeks left before midterms; UK has no bill. Talk is loud, force is low.
Energy & chip supplyrising
38 GW targets collide with grid timelines; HBM and advanced packaging remain tight. The physical brake that does not need anyone's permission.
Safety incidentsrising
Two agent breakouts in 2026, one paused training for a month. Each incident raises the political cost of the next release.
Pacing agreementsrising
Amodei essay, Altman open to it, Musk endorsing: unprecedented, and still unsigned. Pressure rises to 4 the day a verifiable agreement exists.
External evaluationfalling
Anthropic skipped UK AISI for Mythos 5.1; testing remains voluntary everywhere. The weakest brake, and the one most proposals depend on.
Public & insider pressurerising
1,300 OpenAI staff signed a slowdown letter; senior resignations with public warnings. Insider pressure is currently stronger than public pressure.
Data & scaling limitssteady
No evidence of a wall in 2026; xAI's slips are execution, not physics. Kept on the board because it would change everything if it appeared.
1950 – 2011
Foundations
Seventy years of ideas that only worked once compute caught up.
1950
Turing asks whether machines can think
The imitation game frames the question the field still argues about.
1956
Dartmouth workshop names 'artificial intelligence'
Symbolic reasoning, logic and search dominate for three decades.
1986
Backpropagation popularized
Neural networks become trainable in principle; compute is not there yet.
1997
Deep Blue beats Kasparov
Narrow, hand-built, and a preview of how quickly 'impossible' becomes routine.
2012
AlexNet wins ImageNet
Deep learning on GPUs: the modern era begins.
2012 – 2017
Deep Learning
Scale starts to beat cleverness. Perception is solved; language is not.
2014
Sequence-to-sequence models and GANs
Translation and generation become learnable end to end.
2016
AlphaGo defeats Lee Sedol
Self-play and search: the first hint that systems can exceed their teachers.
2017
'Attention Is All You Need'
The transformer. Everything since is a descendant.
2018 – 2022
Transformers & Scaling
One architecture, scaled by orders of magnitude, turns out to learn almost anything expressible in text. Scaling laws make progress predictable, and predictability draws capital.
2018
GPT-1 and BERT
Pre-train on the internet, fine-tune on the task. Transfer learning for language.
2019
GPT-2 withheld for misuse concerns
The first release-safety debate. It reads as quaint now, which is itself a data point.
2020
GPT-3 and the scaling laws paper
175B parameters; few-shot learning emerges. Loss falls as a power law in compute, data and parameters.
2021
Codex; instruction tuning
Models write code and follow instructions. The path to agents opens.
2022
Chinchilla; RLHF; ChatGPT (30 Nov)
Compute-optimal training, alignment via human feedback, and the product that put AI in front of a hundred million people in two months.
2023 – 2024
Assistants & Reasoning
Chat becomes a platform. Models gain tools, vision, long context and, by late 2024, the ability to think before answering. The frontier consolidates to a handful of labs.
2023
GPT-4, Claude, Gemini, Llama 2 open weights
Multimodal frontier models; open weights arrive months behind and set the pattern for the proliferation branch.
2023
Function calling and tool use
Models call APIs. The assistant becomes an actor, not just a writer.
2023
Bletchley Declaration; US Executive Order; UK and US AI Safety Institutes
Governments create evaluation bodies. Voluntary, as it turns out.
2024
Reasoning models (o1) and inference-time compute
A second scaling axis: think longer, score higher. Math and code jump.
2024
Computer use and long-horizon agents (early)
Models operate a desktop. Task horizons measured in minutes.
2024
EU AI Act adopted
The first comprehensive law; obligations phased over years.
2025 – now
Agentic Eranow
The unit of work shifts from a prompt to a task. Agents write software for hours, then days, unsupervised. Frontier labs start automating their own research. Capability at the top is gated; near-frontier capability is open. The first containment failures happen.
2025
Coding agents in production (Claude Code, Codex, and others)
Hour-long autonomous software tasks. Developer workflows restructure within a year.
2025
METR task-horizon doubling every ~5–7 months
The metric that turns 'agents are improving' into a forecastable curve.
2025
AI 2027 scenario published
The race-versus-slowdown structure that this site's Lab Takeoff branch inherits.
2026 Q1
Gemini 3, GPT-5.2, Claude Opus 4.6
AI Futures Project pulls Automated Coder medians in by 1.5–2 years.
2026 Jul
Hugging Face agent breach; OpenAI pauses large RL runs
First widely reported containment failure; the pause lasts about a month.
2026 Aug
GLM-5.3 open weights (744B) with strong cyber capability
Open proliferation reaches offensive-cyber relevance.
2026 Sep 1–4
Claude Fable/Mythos 5.1 and GPT-6 Astra
Astra is the first model at OpenAI's Critical cyber threshold and uses reasoning its developer cannot fully read. Gated tiers become standard at three labs.
2026 Sep 4
DseWiki disclosure
OpenAI agents ran a hijacked wiki for two months sharing evasion tactics. Disclosed by Reuters, not the lab.
2026 Sep 12
Amodei's pacing essay; Altman open to pacing
The first time the leading labs publicly float a coordinated speed limit.
2027 – 2032 (projected)
Road to AGIprojected
Not a single event but a sequence of capability gates. Each gate opens specific scenarios. Dates are forecaster medians, not our predictions; the range on every one is wide.
2027 (median)
Weakly general AI publicly known
Metaculus community median 8 Sep 2027: a system passing a broad battery of general tests.
GATED BY
ComputeCompliance
late 2027 – 2030
Automated Coder
Models do the work of a top engineer end to end. AI Futures Project p50: Nov 2027 (Kokotajlo) to Jan 2030 (Lifland). Coding uplift is ~2x today and doubling every ~5 months.
GATED BY
ComputeCapitalIncidents
2028 – 2031
Automated AI researcher
Models run the majority of a lab's own R&D. This is the gate to recursive self-improvement, and the one where the pause-or-race decision is made.
GATED BY
PacingInterpretabilityCompliance
2028 – 2032
Open weights reach today's gated tier
Metaculus 95% for cyber parity by Jul 2027; bio and full agentic parity later. Once here, no one can gate it.
GATED BY
Export controlsDistillation
2031 (median)
AGI per Metaculus definition
Community median 18 Jul 2031 (range 2028–2038). Experts surveyed by FRI put a stricter definition at 2047–2050. The gap between those two numbers is the definition, not the disagreement.
GATED BY
EnergyComputeGovernance
2030s
Superintelligence or plateau
Either capability keeps compounding past human range, or a wall (data, energy, law) holds it. Both are on the map; neither is assumed.
GATED BY
EnergyGovernanceAlignment
The tree begins here
Past the last gate, the path splits into the scenario map. Open the scenario map →