Chronicle

Stand-alone high-risk systems — hiring, scoring, biometrics, justice, borders — moved from August 2026 to December 2027.

Systems generating intimate imagery of an identifiable person without consent join the prohibited list.

High-risk requirements and transparency obligations take effect. Fines reach €35M or 7% of global turnover.

The new generation of fast, high-volume models.

Terra for everyday work at half the cost, Luna for speed and price.

A flagship aimed at agentic work in code, biology and cybersecurity.

A multi-agent ensemble solved all six problems of IMO 2026 with formal Lean 4 proofs.

Long-run coding and tool use at $2 and $10 per million tokens through 31 Aug 2026.

Open-weight diffusion LM: 2.42× throughput at 98.7% of baseline quality.

Error recovery, state handling and verification turn agents into a repeatable process.

The programme moved from isolated cases to continuous enrolment.

One-twentieth the compute per token, 9× prefill and 15× decode on million-token context.

A 100B-parameter model trained at $1.25 per hour.

The first fully open physical-AI omnimodel: vision, world simulation and action generation in one architecture.

Global data-centre consumption climbs from 447 TWh in 2025 to 565 TWh in 2026 — up 26%.

24/7 work on live tasks: battery module assembly, parts kitting, final-assembly assist. Tesla publishes no production figures of its own.

Cosmos 3 merges scene reasoning with action generation — a general policy instead of a script library.

System security requirements move into enforcement.

Google I/O 2026 unveiled a model that creates video from any input.

Every major hyperscaler has signed nuclear contracts: 13 projects, nearly ten gigawatts.

The standalone app closed and the API ends on 24 Sep 2026 — video generation folds into general models.

The agent-to-agent protocol reached a stable release with signed agent cards.

By GTC 2026 the platform grew to seven chips in production; partner availability lands in H2.

Quantinuum on the H-series; several vendors now report 90–100 logical qubits.

IonQ and Silicon Quantum Computing lead the operation fidelity table.

Third-generation in-house silicon enters broad data-centre deployment.

Microsoft contracted 835 MW for 20 years, with restart expected in 2027.

Playable generated worlds opened to some Google AI Ultra subscribers in the US.

PRIME study participants in the US, UK, Canada and UAE; no serious device-related adverse events reported.

Six new chips as one AI supercomputer: Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6.

Systems closed several open problems that deterred strong mathematicians — the first new verifiable result rather than a restatement.

Bio and cyber risk testing became a standard part of frontier releases rather than a goodwill gesture.

Everyone who updated between January and April 2026 moved their estimate earlier.

Researcher surveys put 50% probability near 2047; professional forecasters mass on 2027–2033.

Stanford faculty insist there is no AGI in 2026 and call for evaluation over evangelism.

Systems plan and test complete changes instead of completing lines.

Triaging failed builds, fixing tests and bumping dependencies moved into the automated loop.

Autonomous labs compress the hypothesis cycle from months to days.

Xaira, Generate Biomedicines and Isomorphic Labs moved from pipeline-filling to clinical candidates.

Speech decoding, arm control and vision hold FDA breakthrough designations.

Google, Amazon and Microsoft scale their own accelerators — NVIDIA's share under real pressure for the first time.

Figure and Agility hold the best-verified deployment records in partner warehouses.

The Willow chip showed that more physical qubits mean exponentially fewer errors overall.

Correction latency dropped below a microsecond — the precondition for real-time operation.

Architectures are diverging: there is no single path to fault tolerance any more, each platform has its own.

Anthropic donated the tool-connection protocol to the Agentic AI Foundation. Around 97M monthly SDK downloads and 9,400+ public servers — the de facto standard.

Real-time interactive worlds with scene memory.

AlphaProof took IMO 2024 silver, Aristotle reached 2025 gold. Every system solving problems formally worked through Lean.

Strong reasoning at a radically lower training cost.

A dense open family for every hardware budget.

Coding systems read the whole project.

Conversation latency drops to human level.

Chain of thought as a product, not a prompt trick.

The first open model comparable in scale to the closed frontier.

A million tokens in production — a whole book or an hour of video fits inside.

The industry measure: the share of real issues from open repositories closed without a human.

Predicting complexes rather than lone proteins: protein with DNA, ligand, ion.

Data centres used about 1.5% of global electricity, with roughly 3% projected by 2030.

Human-level results on professional exams move AI out of the toy category.

Vision as a standard part of a frontier model.

Open-weight sparse mixture of experts: dense-model quality without dense-model cost.

The workaround the industry lived on while context was short. Now a complement rather than a crutch.

Data beats size — every training budget gets rebuilt.

Some skills appear in a jump past a scale threshold rather than growing smoothly.

Instruction tuning — the model starts following orders.

The fastest technology adoption in history.

Training against a written set of principles instead of hand-labelling every answer.

Speech recognition reaches the level where audio becomes an ordinary input.

Asking the model to think step by step raised accuracy — the first hint that thinking at answer time pays.

The predicted-structure database closed a half-century folding problem and became shared infrastructure for biology.

Sparse models: more parameters without a proportional compute increase.

A shared space for text and images — the foundation all image generation grew on.

The same architecture works on images — the transformer turns out to be universal.

In-context learning: a couple of examples is enough.

The curves clusters have been planned against ever since.

Coherent text generation stops being a toy.

Language pre-training becomes the industry default.

The paper that defined the next decade.

Beats Lee Sedol: search plus reinforcement learning.

Residual connections make hundred-layer networks trainable.

GANs: machines start convincingly generating, not just recognising.

ImageNet error collapses — the GPU training era begins.

14 million labelled images — the fuel without which AlexNet could not exist.