Chronicle

Requirements for stand-alone high-risk systems in employment, biometrics, justice and border control apply no later than this date.

The limited grace period for machine marking and detection of synthetic content ends for systems released before 2 August 2026.

The next group of rules applies from 2 August 2026, while specific high-risk-system duties follow their own transition schedule.

Moonshot AI published Kimi K3 weights with launch instructions for Transformers, vLLM and SGLang. This is an infrastructure-scale release, not a model for an ordinary home PC.

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.

ChatGPT Work carries out longer tasks across files, apps and finished deliverables.

Anthropic updated its mainstream model for coding, agent workflows and everyday knowledge work.

Anthropic released a model for long-running coding and knowledge-work tasks designed around hours or days of agentic work rather than a single answer.

The government established a voluntary secure early-access framework and classified evaluation of advanced models' cyber capabilities.

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

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

An open physical-AI omnimodel combined vision, world simulation and action generation in one architecture.

NVIDIA combined physical reasoning, world generation and action trajectories in one open system.

METR measures a rising horizon for software tasks frontier agents complete at a given reliability; estimates beyond 16 hours remain uncertain.

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.

Figure reported producing more than 350 third-generation robots and increasing throughput to one per hour.

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.

Meta said its new accelerator is already in production and is the first of four in-house silicon generations being deployed over two years.

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

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

One neural system controls a humanoid's full body from pixels, including walking, balance and manipulation.

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

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

In the AI Futures sample, estimates updated in early 2026 shifted earlier, but this is not a scientific consensus.

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.

Google, Amazon and Microsoft brought their own accelerators into cloud and production systems, widening the choice beyond NVIDIA GPUs.

Anthropic donated the tool-connection protocol to the Agentic AI Foundation; its SDKs had passed 97M monthly downloads.

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

AgentKit turned agent applications into an engineering layer with versions, evaluations and traces.

Real-time interactive worlds with scene consistency lasting several minutes.

Google opened a protocol for agents from different vendors to exchange tasks and results.

Google brought reasoning into the main Gemini family and combined it with long context and multimodal input.

The Responses API and Agents SDK combined reasoning, tools, orchestration and observability.

METR proposed measuring agents by the human task duration they complete at a given reliability.

Coding systems read the whole project.

Strong reasoning at a radically lower training cost.

A dense open family for every hardware budget.

Anthropic opened a standard for connecting models to data and work tools.

Microsoft contracted 835 MW for 20 years; Constellation plans to return the unit to service in 2028.

Extra inference-time compute became a distinct way to improve difficult-task performance.

One model began working across text, images and audio in real time.

Anthropic's family established sustained frontier competition across reasoning, coding and image analysis.

o1 showed that extra inference-time compute can materially improve hard problem solving.

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.

The model began predicting joint structures of proteins, DNA, RNA, small molecules and ions — a step from protein shape toward interaction mechanisms.

Data centres used about 1.5% of global electricity; the IEA expects demand to more than double by 2030.

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

An open-weight sparse mixture of experts in which only part of the parameters runs for each token.

Google introduced a model family designed from the start to work across text, images, audio and video.

Vision as a standard part of a frontier model.

One model learned to segment arbitrary objects from a point, box or prompt without training for each class.

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

GNoME proposed 2.2 million new crystal structures, including 380,000 potentially stable candidates. These are computational candidates, not finished materials.

The robotic A-Lab planned syntheses, ran experiments and revised recipes, producing 36 of 57 target compounds.

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

A dialogue interface made instruction following a mainstream way to use a language model.

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

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

Data beats size — every training budget gets rebuilt.

Instruction tuning — the model starts following orders.

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

AlphaFold DB published predictions for nearly every catalogued protein known to science. It accelerates research but does not replace experimental validation.

Predicting a protein's 3D structure from its amino-acid sequence approached experimental accuracy for a broad class of problems.

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

Performance was linked by power laws to model size, data and compute.

Translation, classification and question answering were reduced to one format: text in, text out.

Coherent text generation stops being a toy.

Language pre-training becomes the industry default.

Pretraining a transformer on a large corpus and then adapting it to a task became a repeatable transfer-learning recipe.

The Transformer replaced recurrence with attention and scaled better on parallel compute.

Beats Lee Sedol: search plus reinforcement learning.

Residual connections make hundred-layer networks trainable.

One network encoded a sequence and another generated a new one, creating a general template for translation and generation.

GANs: machines start convincingly generating, not just recognising.

Word meaning became geometry: similar words occupied nearby regions of a learned vector space.

A sharp ImageNet gain demonstrated the power of deep networks trained on GPUs.

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

A convolutional network learned to recognise handwritten digits in a real banking system.

Multilayer networks gained a practical way to distribute error through layers and learn internal representations.

Rosenblatt demonstrated a system that adjusted its weights from examples and learned simple classifications.

McCulloch and Pitts described a simple mathematical neuron and showed how networks of them could implement logical operations.