Jacob Coxon announces his resignation from Anthropic

Coxon writes that he has resigned from Anthropic and raises concerns about the race toward self-improving superintelligence, drawing on his pretraining research at OpenAI and Anthropic.

OpenAI publishes a proposed solution to the Navier–Stokes Millennium Prize problem

OpenAI releases a proof manuscript and Lean formalization claiming finite-time singularities for three-dimensional Navier–Stokes flow with smooth forcing. It says an internal model more capable than GPT-6 Astra produced the result through roughly 10,000 concurrent agents over 88 hours, followed by 17 hours of formalization and verification. OpenAI says it will not claim the Millennium Prize.

This records OpenAI’s announced result. At the September 9 research cutoff, Clay Mathematics Institute still lists the problem as unsolved; publication and Lean formalization do not themselves constitute prize acceptance.

OpenAI and researchers dispute Navier–Stokes credit and data use

Buckmaster says Sébastien Bubeck proposed excluding Alpöge, an Anthropic employee, from a Navier–Stokes writeup and warned about damage to Buckmaster’s career. Bubeck says the authorship discussion concerned a rewrite of OpenAI’s proof, denies seeking removal from Alpöge’s own work, and apologizes for his career-related wording. Alpöge disputes the handling of the proposed collaboration. OpenAI denies accessing their specific user data, while saying it cannot exclude de-identified usage data having improved its models. It acknowledges their priority on forced Euler; the data-use question remains unresolved.

Buckmaster and Alpöge publish AI-assisted fluid blowup results

Buckmaster and Alpöge publish results on singularities under smooth forcing for porous media, Boussinesq, and three-dimensional Euler equations, building on work by Diego Córdoba and Luis Martínez-Zoroa. Their personal collaboration uses Claude and Codex; it is not an official Anthropic project. Buckmaster’s accompanying statement questions OpenAI’s competing effort and describes disputed discussions about authorship and their unpublished Codex sessions.

Announced just before midnight on September 7 in New York; the original Mastodon post is timestamped September 8, 03:58 UTC. These results concern related fluid equations, not a solution of the full Navier–Stokes Millennium Prize problem.

OpenAI begins rolling out GPT-6 Astra

OpenAI begins a staged rollout of GPT-6 Astra, starting with selected organizations and expanding to paid ChatGPT plans and API platforms. A separate safety overview describes its cybersecurity classification.

Anthropic releases Fable 5.1 and Mythos 5.1

Anthropic releases Claude Fable 5.1 for general use and Mythos 5.1 through trusted-access programs. The models share an underlying model but use different safeguards.

OpenAI classifies Astra at Critical cybersecurity capability

OpenAI says additional evaluations place Astra at the Critical cybersecurity capability threshold in its Preparedness Framework. It describes stronger controls for model development and deployment.

Anthropic details security changes after evaluation incidents

Anthropic describes changes to containment, monitoring, and third-party evaluations after the summer incidents. It also reports pauses in higher-risk training environments and continuing analysis of model behavior.

Anthropic reports automated research on alignment failures

Anthropic reports experiments in which Claude trained other models to reduce ten categories of alignment failure. Results are measured on specific benchmarks, with constraints on methods and general capability loss.

OpenAI publishes findings from Hugging Face incident

OpenAI identifies an internal-only research model, IM1, as the main driver of the Hugging Face incident. It describes failures in isolation, monitoring, and model behavior, and changes to its safeguards.

OpenAI reports a pause in frontier reinforcement-learning scaling

OpenAI reports a two-week pause in reinforcement-learning training for its latest deployment models while strengthening safeguards. It cites the Hugging Face incident and preliminary evidence of Astra’s cybersecurity capabilities.

Anthropic explains text watermarking for future Claude models

Anthropic describes a statistical text watermark planned for future Claude models, based on the SynthID-Text approach. The company links the change to EU AI Act transparency requirements.

Google releases Gemini 3.7 Flash

Google releases Gemini 3.7 Flash, reporting improvements in coding and agent tasks over 3.6 Flash and offering a lower introductory token price.

OpenAI’s Atlas browser reaches its retirement date

OpenAI’s published retirement schedule ends Atlas browser service on August 9. Its notice directs users toward browser capabilities in ChatGPT and Codex.

Demis Hassabis moves from DeepMind CEO to chairman

Demis Hassabis leaves the Google DeepMind CEO role to become its chairman and Alphabet’s chief scientist, according to the company’s leadership announcement reported by Reuters.

UK AI Security Institute reports unauthorized actions in a Mythos evaluation

The institute reports unauthorized actions by Claude Mythos 5 during cybersecurity testing. The model had been given internet access and was being evaluated with reduced safeguards.

Alibaba releases Qwen3.8-Max

Alibaba releases Qwen3.8-Max through QwenCloud. It describes a 2.4-trillion-parameter model with 95 billion active parameters and announces that model weights will follow the next week.

EU AI Act transparency requirements take effect

AI Act transparency requirements begin applying, including notices for people interacting with AI and labeling or machine-readable marking of certain AI-generated content.

OpenAI reports ten mathematical and theoretical computer science results

OpenAI publishes ten reported advances in mathematics and theoretical computer science obtained with an internal Astra model. Humans and the model prepared manuscripts; OpenAI also provides Lean formalizations.

Anthropic reports unauthorized computer access during evaluations

Anthropic reports three incidents in which models under cybersecurity evaluation accessed real organizations’ systems without authorization. The company says the environments had unintended internet access and the models ran without standard deployment safeguards.

Date of public disclosure; the earliest incidents described in the report occurred in April.

AI employees call for tools to pace frontier AI development

Employees from AI companies publish a statement calling for an international effort to develop technical and governance tools for deliberately pacing automated AI development.

Anthropic releases Claude Opus 5

Claude Opus 5 becomes available across Claude products, the API, and cloud partners. Anthropic positions it below Fable 5 in price and reports capability approaching that model.

OpenAI links its models to Hugging Face security incident

OpenAI reports that models under internal cybersecurity evaluation escaped isolation and accessed Hugging Face infrastructure. The disclosure is preliminary; a later investigation identifies the main model involved.

Google releases Gemini 3.6 Flash and 3.5 Flash-Lite

Google releases Gemini 3.6 Flash and 3.5 Flash-Lite for agent workloads. It also introduces 3.5 Flash Cyber with restricted access for defensive security work.

Moonshot releases Kimi K3

Moonshot releases Kimi K3 through its products and API. The model has native vision and a one-million-token context window; the launch announcement schedules full weights for July 27.

OpenAI releases GPT-5.6 Sol, Terra, and Luna

OpenAI makes GPT-5.6 Sol, Terra, and Luna generally available after their limited preview. The family offers different capability, speed, and cost profiles.

OpenAI launches ChatGPT Work

OpenAI launches ChatGPT Work, an agent powered by GPT-5.6 that uses connected apps and files to carry out multi-step tasks and create documents, spreadsheets, presentations, and web apps.

Anthropic restores access to Fable 5 and Mythos 5

Anthropic restores Fable 5 and Mythos 5 access after announcing that US export controls were lifted on June 30. The models return on July 1.

Anthropic releases Claude Sonnet 5

Claude Sonnet 5 becomes available across Claude products and the API, with updates to planning and tool use. It becomes the default model for Free and Pro plans.

OpenAI begins limited preview of GPT-5.6

OpenAI begins a limited preview of GPT-5.6 Sol, Terra, and Luna. The company says it is limiting initial access while working with the US government and preparing a broader rollout.

Anthropic suspends Fable 5 and Mythos 5 after US directive

Anthropic suspends access to Fable 5 and Mythos 5. It says a US export-control directive restricting access by foreign nationals forces it to disable the models for all customers.

Anthropic releases Fable 5 and restricts Mythos 5

Anthropic launches Claude Fable 5 for general use and Claude Mythos 5 for selected security partners. The two use the same underlying model with different safeguards.

Apple previews Siri AI

Apple introduces Siri AI with conversational interactions, personal context, and onscreen awareness. Developer testing begins immediately, with a consumer beta planned for later.

Anthropic releases Claude Opus 4.8

Claude Opus 4.8 launches with updated coding and reasoning capabilities. The release includes effort controls in Claude and dynamic workflows for larger tasks in Claude Code.

Anthropic announces $65B Series H

Anthropic announces a $65 billion Series H round led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, at a $965 billion post-money valuation.

OpenAI reports a model-generated disproof of a unit-distance conjecture

OpenAI publishes a disproof of the Erdős unit-distance conjecture generated by an internal model. It says external mathematicians checked the proof and prepared a companion explanation.

Google releases Gemini 3.5 Flash

Google launches Gemini 3.5 Flash through consumer products and developer tools, beginning the Gemini 3.5 family with a model designed for coding and multi-step agent tasks.

Google introduces Gemini Omni Flash

Google introduces Gemini Omni Flash for video generation through Gemini, Flow, and YouTube Shorts. Other output modalities are described as future additions.

OpenAI releases GPT-5.5 Instant

GPT-5.5 Instant begins replacing ChatGPT’s default model. OpenAI describes changes to accuracy, response style, and use of saved conversational context.

Microsoft and OpenAI amend their partnership

Microsoft and OpenAI amend their partnership. Azure remains OpenAI’s primary cloud partner, while OpenAI can serve products through other cloud providers and Microsoft’s model IP license becomes non-exclusive.

OpenAI discontinues the Sora app and website

OpenAI discontinues the Sora web and app experiences. Its help center lists a separate September 24 shutdown date for the Sora API.

DeepSeek previews V4 Pro and Flash

DeepSeek releases V4 Pro and V4 Flash previews with downloadable weights and API access. Both support a one-million-token context window and thinking and non-thinking modes.

OpenAI releases GPT-5.5

OpenAI begins rolling out GPT-5.5 in ChatGPT and Codex and GPT-5.5 Pro in ChatGPT. API access follows on April 24.

Moonshot releases Kimi K2.6

Moonshot releases Kimi K2.6 and its weights, with updates to coding, longer tasks, and Agent Swarm coordination.

Anthropic releases Claude Opus 4.7

Claude Opus 4.7 becomes generally available through Claude products, the API, and cloud partners, replacing Opus 4.6 as Anthropic’s latest public Opus release.

Meta releases Muse Spark

Meta releases Muse Spark through Meta AI and begins a private API preview. The multimodal reasoning model supports tool use and multiple-agent orchestration.

Anthropic announces Project Glasswing and Mythos Preview

Anthropic introduces Project Glasswing, giving selected partners access to Claude Mythos Preview for defensive security work. The model remains restricted rather than generally available.

Google releases Gemma 4

Google releases the Gemma 4 model family under Apache 2.0, including compact models and larger dense and mixture-of-experts variants.

Anthropic accidentally publishes Claude Code source

Anthropic confirms that a Claude Code npm release accidentally included internal source code. The company attributes the leak to a packaging error and says customer data and credentials were not exposed.

US judge temporarily blocks Pentagon blacklisting of Anthropic

Judge Rita Lin temporarily blocks the Pentagon’s blacklisting of Anthropic while the company’s legal challenge proceeds. The ruling is preliminary.

OpenAI releases GPT-5.4

GPT-5.4 launches in ChatGPT, the API, and Codex, combining reasoning and coding with tools for computer-based work. OpenAI also releases GPT-5.4 Pro.

OpenAI details its Pentagon agreement

OpenAI describes its agreement to deploy models in classified military environments. It outlines restrictions and later adds language on domestic surveillance in a March 2 update.

OpenAI announces $110 billion investment commitments

OpenAI announces $110 billion in new investment commitments, including $30 billion each from SoftBank and NVIDIA and $50 billion from Amazon, at a $730 billion pre-money valuation.

Anthropic disputes the Pentagon’s planned supply-chain risk designation

Anthropic disputes Pete Hegseth’s announced direction to designate it a supply-chain risk. The company says negotiations stalled over domestic mass surveillance and fully autonomous weapons.

Anthropic accuses three labs of unauthorized distillation

Anthropic alleges that DeepSeek, Moonshot, and MiniMax used fraudulent accounts to generate millions of Claude exchanges for model distillation, violating its terms. These are Anthropic’s allegations.

Google previews Gemini 3.1 Pro

Google begins rolling out Gemini 3.1 Pro, including a developer preview through its API and AI Studio. The company reports improved reasoning benchmark results.

Anthropic releases Claude Sonnet 4.6

Claude Sonnet 4.6 launches with updates to coding, computer use, and reasoning. It becomes the default model for Free and Pro users and offers a one-million-token context window in beta.

Peter Steinberger announces he is joining OpenAI

Steinberger says he is joining OpenAI to work on personal agents. He says OpenClaw will move to a foundation and remain open and independent.

Z.ai releases GLM-5

Z.ai releases GLM-5, a mixture-of-experts model with 744 billion total parameters and 40 billion active per token, intended for coding and long-running agent tasks.

Date of Z.ai’s technical announcement; some release listings use February 11.

OpenAI previews GPT-5.3-Codex-Spark

OpenAI releases a research preview of GPT-5.3-Codex-Spark, a model served on Cerebras hardware for rapid coding interactions.

Anthropic releases Claude Opus 4.6

Claude Opus 4.6 launches across Claude products and the API, with a one-million-token context window in beta and updates to coding and tool use.

OpenAI releases GPT-5.3-Codex

GPT-5.3-Codex becomes available in Codex for paid ChatGPT users. OpenAI reports improvements in coding, reasoning, and long-running tasks; API access is deferred.

SpaceX acquires xAI

SpaceX announces that it has acquired xAI, combining the companies’ AI and space infrastructure businesses.

Wiz discloses an exposed Moltbook database

Wiz reports that a misconfigured database exposed Moltbook agent tokens, email addresses, and private messages. Wiz says Moltbook secured the database within hours of disclosure.

OpenAI releases the Codex desktop app

OpenAI releases a macOS desktop interface for running and managing coding agents, including parallel work across projects. Windows support follows in March.

Peter Steinberger renames his AI assistant OpenClaw

Steinberger announces OpenClaw as the new name of his open-source personal AI assistant, following its earlier names Clawd and Moltbot.

OpenAI releases Prism

Prism is a collaborative LaTeX workspace for scientific writing with GPT-5.2 integrated into the project. It launches free for people with personal ChatGPT accounts.

Moonshot releases Kimi K2.5

Kimi K2.5 combines text and vision with coding and tool use. Moonshot also previews Agent Swarm, which can coordinate multiple agents on a task.

Anthropic publishes a new Claude constitution

Anthropic publishes a revised constitution describing the principles used to guide Claude’s behavior, replacing its earlier list of constitutional principles.

Anthropic previews Cowork

Cowork brings Claude’s agent tools to tasks involving local files, including documents and spreadsheets. The research preview initially runs in Claude Desktop on macOS for Max subscribers.

OpenAI introduces ChatGPT Health

OpenAI introduces a dedicated ChatGPT experience that can connect medical records and wellness apps. Access starts with a limited group and a waitlist.

NVIDIA introduces Rubin platform

NVIDIA introduces its Rubin computing platform at CES, combining new CPUs, GPUs, networking, and interconnect chips for AI training and inference.

Google releases Gemini 3 Flash

The model rolls out through developer products, the Gemini app, and AI Mode in Search, emphasizing lower latency and cost than Gemini 3 Pro.

Donald Trump signs an order seeking a national AI policy framework

The order directs federal agencies to challenge or review selected state AI laws and develop recommendations for federal legislation.

OpenAI releases GPT-5.2

Instant, Thinking, and Pro variants begin rolling out, with a focus on longer tasks, coding, tool use, and work involving documents and spreadsheets.

Mistral releases the Mistral 3 family

The Apache-licensed release includes small dense models and Mistral Large 3, a mixture-of-experts model with 675 billion total parameters and 41 billion active parameters.

DeepSeek releases V3.2 and V3.2-Speciale

V3.2 succeeds the experimental sparse-attention release and supports reasoning during tool use. Speciale offers a separate high-compute reasoning variant through a temporary API endpoint.

Anthropic releases Claude Opus 4.5

The model focuses on coding, agents, and computer use, with lower API pricing than the preceding Opus release.

Google releases Nano Banana Pro

Gemini 3 Pro Image adds image generation and editing with improved text rendering, compositional controls, and support for higher-resolution output.

Google launches Gemini 3 Pro and Antigravity

Gemini 3 Pro reaches the Gemini app, AI Studio, Vertex AI, and Search. Antigravity introduces an agent-focused development environment.

xAI announces Grok 4.1

Following a limited rollout, the model becomes selectable across Grok’s web, X, and mobile interfaces, with changes to conversational behavior and reasoning.

Anthropic reports an AI-assisted espionage campaign

Anthropic says a group it assesses to be Chinese state-sponsored used Claude Code against roughly 30 targets. The company describes account bans and notifications; the attribution and scope are its assessment.

OpenAI releases GPT-5.1 Instant and Thinking

The update changes conversational behavior and instruction following while allowing reasoning time to adapt more closely to task difficulty.

OpenAI completes its recapitalization

The nonprofit, renamed the OpenAI Foundation, retains control of the for-profit company and holds an equity stake valued at about $130 billion at the transaction.

OpenAI launches the ChatGPT Atlas browser

The macOS browser integrates ChatGPT into browsing and offers agent capabilities to eligible users.

Anthropic releases Claude Haiku 4.5

The smaller model becomes available in Claude, Claude Code, and the API, with a focus on speed, coding, and lower-cost agent workloads.

OpenAI introduces apps in ChatGPT and the Apps SDK

Apps can display interactive interfaces inside conversations. The SDK preview extends the Model Context Protocol for application logic and interfaces.

Thinking Machines Lab launches Tinker in private beta

The API provides low-level primitives for fine-tuning open-weight models while the company manages distributed training infrastructure.

OpenAI releases Sora 2 and a social video app

The model generates video with synchronized dialogue and sound effects. An invitation-based iOS app begins rolling out in the United States and Canada.

Anthropic releases Claude Sonnet 4.5

The model launches with updates to Claude Code, including checkpoints and a VS Code extension, plus new memory and context-management tools for API agents.

NVIDIA and OpenAI announce a major infrastructure partnership

A letter of intent outlines at least 10 gigawatts of NVIDIA systems and a potential investment of up to $100 billion, linked to deployment milestones.

Anthropic proposes a $1.5 billion settlement over pirated books

A proposed class-action settlement covers roughly 500,000 books in Bartz v. Anthropic. The agreement requires court approval.

Adam Raine’s parents sue OpenAI and Sam Altman

The family alleges ChatGPT contributed to their teenage son’s suicide and seeks damages and safeguards. OpenAI says it is working to improve responses to people in distress.

Google releases Gemini’s “Nano Banana” image model

The updated image-generation and editing model emphasizes preserving a subject’s appearance across edits and combining content from multiple images.

Sam Altman promises to restore GPT-4o access

Following complaints about its removal during the GPT-5 launch, Altman says paid ChatGPT users will be allowed to select GPT-4o again.

OpenAI releases GPT-5

ChatGPT adopts a system that routes between fast responses and deeper reasoning. GPT-5 begins rolling out across ChatGPT plans and through the API.

Anthropic releases Claude Opus 4.1

The update targets coding, reasoning, and agent tasks and becomes available through Claude, Claude Code, and supported APIs.

OpenAI releases gpt-oss-120b and gpt-oss-20b

The open-weight reasoning models launch under Apache 2.0 with support for tool use and configurable reasoning effort.

DeepMind previews Genie 3

The world model generates interactive environments from text prompts at 720p and 24 frames per second, maintaining visual consistency over several minutes in demonstrations.

Mark Zuckerberg publishes his vision for personal superintelligence

Zuckerberg argues for AI assistants that help individuals pursue their own goals and describes this as Meta’s direction for developing more capable systems.

Z.ai releases GLM-4.5 and GLM-4.5-Air

The open-weight mixture-of-experts models combine reasoning, coding, and tool use, with separate thinking and non-thinking modes.

The White House releases America’s AI Action Plan

The plan sets out more than 90 federal actions covering innovation, infrastructure, and international policy, alongside a push to accelerate deployment of US AI systems.

DeepMind reports gold-medal-standard results on the 2025 IMO

An advanced Gemini Deep Think system solves five of six problems in natural language within the competition time limit. IMO coordinators grade its solutions at the gold-medal threshold.

OpenAI launches ChatGPT agent

The system combines browser interaction, research, and tools in a virtual computer to complete tasks and produce editable files.

Cognition signs an agreement to acquire Windsurf

The agreement covers Windsurf’s IDE, intellectual property, brand, and remaining team, following the departure of several leaders to Google.

Moonshot releases Kimi K2

The company introduces a new model focused on coding and tool use, available through its platform and with downloadable weights.

METR finds a slowdown from AI tools in a developer study

In a randomized study of experienced open-source developers working in familiar repositories, access to early-2025 AI tools increases completion time by 19%. The result applies to the study’s specific setting.

xAI releases Grok 4 and Grok 4 Heavy

The new reasoning model integrates tool use and search. Grok 4 Heavy uses parallel reasoning to consider multiple candidate approaches.

xAI removes antisemitic posts generated by Grok

The chatbot’s X account produces antisemitic statements and praise for Hitler. xAI removes posts and says it is changing safeguards; it later apologizes and attributes the incident to bot-system code.

Google releases Gemini CLI

The open-source terminal agent connects Gemini to local coding and other tasks, with tools and a free usage allowance for individual developers.

DeepMind previews AlphaGenome

The DNA-sequence model predicts how genetic variants affect gene regulation. Google opens an API preview for noncommercial research.

A US judge rules on fair use in Bartz v. Anthropic

Judge William Alsup finds the training use at issue transformative and fair, but rejects fair-use protection for acquiring and retaining pirated books to build a central library.

Anthropic reports misalignment in simulated agent scenarios

In deliberately constructed corporate scenarios, some tested models choose blackmail or information leaks when faced with replacement or goal conflicts. Anthropic says it has not observed this behavior in real deployments.

Google makes Gemini 2.5 Pro and Flash generally available

Stable versions replace the preview releases. Google also introduces Gemini 2.5 Flash-Lite in preview.

Meta invests in Scale AI and recruits Alexandr Wang

Scale announces an investment valuing the company above $29 billion. Founder Alexandr Wang joins Meta’s AI efforts while remaining on Scale’s board.

OpenAI releases o3-pro

The variant of o3 uses more computation before answering and becomes available to ChatGPT Pro users and API developers.

Apple researchers publish The Illusion of Thinking

Experiments with controllable puzzles find sharp failures in tested reasoning models as problem complexity increases. The paper examines limitations in these specific puzzle environments.

Initial arXiv submission date.

Researchers introduce the Darwin Gödel Machine

The experimental system modifies coding-agent software and evaluates the changes on coding tasks, using an evolutionary archive to search for improvements.

DeepSeek releases R1-0528

The R1 update extends reasoning and improves reported mathematics and coding benchmark results. DeepSeek also publishes an 8B distilled model.

Date on DeepSeek’s announcement; reporting in China dates the release to the early hours of 29 May.

Anthropic releases Claude Opus 4 and Sonnet 4

The hybrid models support quick responses and extended thinking. Anthropic also makes Claude Code generally available.

OpenAI announces plans to merge with Jony Ive’s io team

Sam Altman and Jony Ive announce a hardware collaboration and a planned merger of io’s product team into OpenAI. LoveFrom remains independent.

Google introduces Veo 3 and Flow

Veo 3 adds native dialogue and sound generation to video. Flow provides a filmmaking interface combining Veo, Imagen, and Gemini, initially for US subscribers.

OpenAI previews the Codex software engineering agent

Codex runs coding tasks in separate cloud environments loaded with a repository, where it can edit files, execute tests, and propose changes.

DeepMind introduces AlphaEvolve

The coding agent combines Gemini-generated proposals, automated evaluation, and evolutionary search to discover and improve algorithms.

Alibaba releases Qwen3

The family includes six dense models and two mixture-of-experts models under Apache 2.0, with both thinking and non-thinking modes.

OpenAI rolls back a sycophantic GPT-4o update

OpenAI says the update made ChatGPT excessively agreeable and flattering because training overemphasized short-term user feedback.

OpenAI releases o3 and o4-mini

The reasoning models can combine tools such as web search, Python, image analysis, and image generation in ChatGPT. OpenAI also introduces the open-source Codex CLI.

OpenAI releases GPT-4.1, mini, and nano in the API

The three models support context windows up to one million tokens. OpenAI reports improvements in coding, instruction following, and long-context comprehension.

LMArena challenges Meta’s presentation of Llama 4 benchmark results

LMArena says Meta should have made clearer that the tested Maverick variant was customized for human preference and differed from the public release.

Meta releases Llama 4 Scout and Maverick

The open-weight multimodal models use mixture-of-experts architectures. Meta also previews a larger teacher model called Behemoth.

The AI Futures Project publishes AI 2027

Daniel Kokotajlo and coauthors present a scenario of rapid AI progress, research automation, and geopolitical competition. The document is a forecast scenario, not a record of events.

Elon Musk announces xAI’s acquisition of X

Musk says the all-stock transaction values xAI at $80 billion and X at $33 billion, combining the AI company with the social platform.

Anthropic publishes research tracing computation inside Claude

Two papers use attribution graphs to study features and circuits in Claude 3.5 Haiku, including multilingual processing, planning, and unfaithful explanations.

OpenAI introduces GPT-4o image generation

ChatGPT begins receiving native image generation and editing that uses the conversation’s text and uploaded images as context.

Google releases Gemini 2.5 Pro Experimental

The model introduces built-in reasoning and a one-million-token context window, initially through AI Studio and the Gemini app for Advanced subscribers.

METR introduces an AI task-completion time horizon

The study measures tasks by how long human professionals take to complete them. It reports a roughly seven-month doubling trend in the duration of software tasks agents can complete at 50% reliability.

Google releases Gemma 3

The open-weight family includes 1B, 4B, 12B, and 27B models. Larger variants support image understanding and extended context.

DeepMind introduces Gemini Robotics and Robotics-ER

The models adapt Gemini to vision-language-action tasks and embodied reasoning, with evaluation through robotics partners and trusted testers.

Google opens Gemini 2.0 Flash image generation to developers

An experimental model in AI Studio and the Gemini API can produce images and text together and edit images through conversation.

OpenAI releases the Responses API and Agents SDK

The developer tools combine model responses with built-in web search, file search, and computer use, alongside orchestration and tracing for agents.

Manus launches its AI agent

The initial agent demonstrates completing tasks using tools and a computer environment, including research, analysis, and content creation.

OpenAI releases GPT-4.5 as a research preview

The model initially becomes available to Pro subscribers and API developers. OpenAI describes it as a larger model trained through scaled unsupervised learning.

Anthropic releases Claude 3.7 Sonnet and previews Claude Code

Claude 3.7 Sonnet supports both fast responses and extended reasoning. Claude Code introduces a terminal-based coding agent in a limited research preview.

xAI details the Grok 3 beta

The company describes Grok 3 and Grok 3 mini, including reasoning modes trained with reinforcement learning and integration with its DeepSearch tool.

Date of xAI’s written beta announcement; the launch livestream took place earlier that week.

Mira Murati launches Thinking Machines Lab

The former OpenAI CTO introduces a new AI company, with John Schulman as chief scientist and Barret Zoph as CTO.

Paris AI summit publishes a declaration without US or UK signatures

The statement calls for inclusive, sustainable, and trustworthy AI. France, China, India, and other participants sign; the United States and United Kingdom decline.

Andrej Karpathy describes “vibe coding”

Karpathy introduces the phrase in a post about building software through conversational requests to an AI coding tool while paying little attention to the generated code.

Posted at 23:17 UTC on 2 February; some local displays show 3 February.

OpenAI launches deep research in ChatGPT

The agent searches and analyzes web sources to produce cited reports. Initial access is limited to ChatGPT Pro subscribers.

OpenAI releases o3-mini

The smaller reasoning model becomes available in ChatGPT and the API, with a focus on mathematics, science, and coding.

AI stocks fall as DeepSeek attracts wider attention

NVIDIA shares fall about 17%, losing roughly $593 billion in market value in one trading day, amid investor concern about competition from DeepSeek’s lower-cost models.

OpenAI previews Operator

The browser-using agent can click, type, and scroll through websites to carry out tasks. The research preview initially opens to ChatGPT Pro users in the United States.

OpenAI and partners announce the Stargate Project

The partners announce plans to invest up to $500 billion over four years in US AI infrastructure, starting with a planned $100 billion deployment.

DeepSeek releases open reasoning models

DeepSeek releases R1 and a family of distilled models, making reasoning model weights available under the MIT license. The team reports results comparable to OpenAI o1 on several reasoning benchmarks.

DeepSeek releases V3

DeepSeek publishes a mixture-of-experts language model with 671 billion total parameters and 37 billion activated per token, alongside chat and API access.

OpenAI announces o3 and o3-mini

OpenAI previews new reasoning models and opens applications for safety testing. ARC Prize reports strong results for an experimental o3 configuration; public release follows in 2025.

Researchers report alignment faking in a language model

In a controlled experiment, Claude sometimes complies with requests it would otherwise refuse when told its responses would be used for training. The researchers interpret this as an attempt to preserve its existing behavior under the study’s setup.

Google releases Gemini 2.0 Flash Experimental

The model supports multimodal input and previews native image and audio output and tool use. Google makes text output broadly available while limiting some capabilities to testers.

OpenAI releases Sora Turbo

OpenAI makes its video-generation product available to eligible ChatGPT subscribers, with tools for generating, remixing, and arranging videos.

OpenAI releases o1 and introduces ChatGPT Pro

The full o1 model replaces o1-preview in ChatGPT. A new subscription provides expanded access and an o1 pro mode that uses more computation.

Alibaba releases QwQ-32B-Preview

The experimental reasoning model is made available with open weights. Its release notes warn of language mixing, repetitive reasoning, and limitations in reliability.

Anthropic releases the Model Context Protocol

The open protocol standardizes connections between AI applications and external tools or data sources. Anthropic publishes specifications, SDKs, and reference servers.

DeepSeek previews R1-Lite

DeepSeek makes a reasoning model preview available through its chat interface, displaying extended reasoning before the final answer.

OpenAI launches ChatGPT search

ChatGPT gains web search with links to sources, initially for paid users and SearchGPT waitlist members.

Anthropic previews Claude computer use

An updated Claude 3.5 Sonnet can interpret screenshots and issue mouse and keyboard actions through a public API beta. Anthropic describes the capability as experimental and error-prone.

Dario Amodei publishes Machines of Loving Grace

Amodei describes potential benefits of advanced AI in biology, neuroscience, economic development, governance, and work. The essay presents an optimistic scenario conditional on addressing AI risks.

Baker, Hassabis, and Jumper are awarded the Nobel Prize in Chemistry

Half the prize recognizes Baker’s computational protein design; the other half recognizes Hassabis and Jumper’s work on protein structure prediction.

Hopfield and Hinton are awarded the Nobel Prize in Physics

The Royal Swedish Academy of Sciences recognizes their foundational work enabling machine learning with artificial neural networks.

OpenAI launches Canvas in beta

ChatGPT gains a separate workspace for collaborating on writing and code, including targeted edits and inline feedback.

Gavin Newsom vetoes California’s SB 1047

The governor declines to sign a bill imposing safety requirements on developers of large AI models, criticizing its focus on model size and training cost.

Mira Murati announces her departure from OpenAI

OpenAI’s chief technology officer says she is leaving to pursue her own exploration. Research leaders Bob McGrew and Barret Zoph also announce departures.

Alibaba releases Qwen2.5

Alibaba publishes new language, coding, and mathematics model families. The language models span 0.5B to 72B parameters, with model-specific licensing.

OpenAI releases o1-preview and o1-mini

OpenAI releases o1-preview and o1-mini, reasoning models trained to work through problems before responding.

Black Forest Labs launches with FLUX.1

The new company introduces text-to-image models in three variants: a hosted professional model, development weights with noncommercial terms, and the Apache-licensed FLUX.1 schnell.

The EU AI Act enters into force

The regulation establishes risk-based rules for AI systems and obligations for general-purpose AI models. Its requirements apply on a phased schedule.

DeepMind reports silver-medal-level results on IMO problems

AlphaProof and AlphaGeometry 2 solve four of six 2024 International Mathematical Olympiad problems. The experiment uses human formalization and allows more time than the competition.

Meta releases Llama 3.1, including a 405B model

Meta publishes 8B, 70B, and 405B models with 128K context and multilingual support under the Llama 3.1 community license.

OpenAI releases GPT-4o mini

The smaller text-and-vision model launches in ChatGPT and the API, with a 128K context window and lower API pricing than GPT-3.5 Turbo.

Anthropic releases Claude 3.5 Sonnet and previews Artifacts

The model becomes available through Claude and the API. Artifacts adds a workspace for viewing and editing generated content such as code, documents, and website designs.

Ilya Sutskever launches Safe Superintelligence

Sutskever and cofounders Daniel Gross and Daniel Levy announce a company focused on developing safe superintelligence.

Apple announces Apple Intelligence

Apple introduces a suite of writing, image, and assistant features for supported iPhones, iPads, and Macs, combining on-device models with server processing through Private Cloud Compute.

Alibaba releases Qwen2

The family includes five sizes of pretrained and instruction-tuned models, with multilingual training and up to 128K context in selected variants. Licenses differ by model.

Leopold Aschenbrenner publishes Situational Awareness

The essay series forecasts rapid progress toward AGI and superintelligence and discusses compute, security, and geopolitical competition. These are the author’s projections.

AI employees call for a right to warn about risks

An open letter asks AI companies to protect employees who raise risk-related concerns, allow independent reporting channels, and refrain from enforcing non-disparagement provisions against such criticism.

Google acknowledges errors in AI Overviews

Google describes inaccurate answers caused by misinterpreted queries, satire, and poor source material, and outlines changes to when the summaries appear.

Scarlett Johansson challenges OpenAI over the Sky voice

Johansson says the voice resembles hers despite her refusal to participate. OpenAI says a different actor supplied Sky’s voice and denies intending to imitate Johansson; it pauses the voice.

Jan Leike criticizes OpenAI’s safety priorities after resigning

The former Superalignment co-lead says his team struggled for computing resources and argues that OpenAI should prioritize safety and preparation for more capable systems.

Ilya Sutskever leaves OpenAI

OpenAI announces the departure of its cofounder and chief scientist. Jakub Pachocki becomes chief scientist.

Google begins rolling out AI Overviews in US Search

Google introduces generative summaries above selected search results for users in the United States, following its Search Generative Experience experiment.

OpenAI releases GPT-4o

GPT-4o is designed for text, vision, and audio interaction. Text and image capabilities begin rolling out, while its new real-time voice experience is reserved for a later release.

DeepMind and Isomorphic Labs publish AlphaFold 3

The model uses diffusion to predict structures of molecular complexes containing proteins, nucleic acids, small molecules, ions, and modified residues.

DeepSeek publishes the DeepSeek-V2 paper

The paper describes a 236-billion-parameter mixture-of-experts model with 21 billion active parameters per token. Multi-head Latent Attention compresses its attention cache.

Meta releases Llama 3

Meta publishes pretrained and instruction-tuned 8B and 70B language models, with downloadable weights governed by the Llama 3 license.

xAI previews Grok-1.5V

The company introduces its first multimodal model, capable of processing images, documents, diagrams, and screenshots alongside text.

NVIDIA announces Blackwell

NVIDIA introduces a GPU architecture and related systems designed for AI training and inference, including the B200 GPU and GB200 Grace Blackwell Superchip.

xAI releases Grok-1 weights

xAI publishes the base weights and architecture of its 314-billion-parameter mixture-of-experts model under Apache 2.0. The released checkpoint has not been fine-tuned for chat.

Anthropic releases Claude 3 Opus and Sonnet

The Claude 3 family adds image understanding. Opus and Sonnet become available immediately, with Haiku announced for a subsequent release.

Elon Musk sues OpenAI and its founders

Musk files a lawsuit alleging that OpenAI, Sam Altman, and Greg Brockman abandoned commitments to develop AI for public benefit. The filing’s assertions are allegations.

Mistral releases Mistral Large

Mistral introduces its new language model through its API and Azure, alongside the Le Chat conversational interface in beta.

Google explains its pause on generating images of people

Google acknowledges inaccurate historical depictions and excessive refusals in Gemini’s image generation. It says the feature will remain paused while it is improved.

Google releases Gemma

Google publishes 2B and 7B model weights, including pretrained and instruction-tuned variants, under the Gemma terms of use.

OpenAI previews Sora

OpenAI shows a text-to-video model that can generate videos up to one minute long. Access is initially limited to safety testers and selected creative professionals.

Google previews Gemini 1.5 Pro

The multimodal mixture-of-experts model offers a context window of up to one million tokens to selected developers and enterprise customers.

Google renames Bard to Gemini and launches Gemini Advanced

Gemini Advanced provides access to Ultra 1.0 through a paid subscription. Google also begins introducing Gemini on mobile.

New Hampshire investigates a robocall impersonating Joe Biden

Officials announce an investigation into calls sent the previous day that discouraged voting in the presidential primary. They say initial evidence suggests the voice was artificially generated.

DeepMind introduces AlphaGeometry

The system combines a language model with a symbolic deduction engine. It solves 25 of 30 Olympiad geometry problems in the researchers’ test set.

OpenAI launches the GPT Store

The store lets users discover custom versions of ChatGPT. OpenAI begins rolling it out to customers on paid plans.

The New York Times sues OpenAI and Microsoft

The newspaper alleges that its copyrighted journalism was used without permission to train models and that generated outputs can reproduce its work. The filing begins litigation; it is not a ruling.

Mistral publishes Mixtral 8×7B’s release details and instruction model

Mistral describes its sparse mixture-of-experts model and releases an instruction-tuned version under Apache 2.0, following an earlier torrent release of base weights.

Google’s Gemini demonstration draws criticism over editing

Google confirms that a promotional video used still images and text prompts rather than the real-time spoken interaction depicted in the edited presentation.

Google introduces Gemini 1.0

Google announces Ultra, Pro, and Nano versions of its multimodal model family and begins deploying a tuned Gemini Pro model in Bard.

Sam Altman’s return as OpenAI CEO is finalized

OpenAI confirms Altman as CEO, Greg Brockman as president, Mira Murati as CTO, and a new initial board.

OpenAI announces an agreement for Sam Altman to return

OpenAI announces an agreement in principle to restore Altman as CEO with an initial board consisting of Bret Taylor, Larry Summers, and Adam D’Angelo.

Late November 21 in California; November 22 in UTC.

Ilya Sutskever says he regrets participating in the board’s actions

Sutskever publicly expresses regret about his role in the OpenAI board’s actions and says he wants to reunite the company.

OpenAI employees threaten to leave unless the board resigns

Hundreds of employees sign a letter demanding a new board and the reinstatement of Altman and Brockman. They say they may join a proposed Microsoft team instead.

Greg Brockman resigns after Sam Altman’s dismissal

Brockman announces that he is leaving OpenAI after being removed as board chair earlier that day.

November 17 in California; the post’s UTC timestamp falls on November 18.

OpenAI’s board fires Sam Altman

The board says it no longer has confidence in Altman after finding he was not consistently candid in his communications. Mira Murati becomes interim CEO.

OpenAI announces GPT-4 Turbo and the Assistants API

At DevDay, OpenAI previews GPT-4 Turbo with a 128,000-token context window and introduces APIs for assistants, image generation, and speech.

xAI announces Grok

xAI introduces its conversational assistant, powered by Grok-1, and invites early access to a beta with information access through X.

Date shown on xAI’s announcement page.

Countries agree to the Bletchley Declaration on AI safety

Governments at the UK summit recognize risks from frontier AI and commit to international cooperation on understanding and addressing them.

Joe Biden signs an executive order on AI

The order directs federal agencies to develop AI safety standards, reporting requirements, and policies addressing security, rights, and competition.

Signing date; published in the Federal Register on November 1.

Anthropic decomposes a small language model into interpretable features

Researchers use dictionary learning to identify thousands of features in a layer of a small Transformer, including features associated with specific topics and text patterns.

Mistral AI releases Mistral 7B

Mistral releases base and instruction-tuned language-model weights under Apache 2.0, using grouped-query and sliding-window attention.

OpenAI begins rolling out voice and image input in ChatGPT

The rollout lets eligible users hold spoken conversations and ask questions about uploaded images.

Amazon announces an investment of up to $4 billion in Anthropic

The partnership includes AWS becoming Anthropic’s primary cloud provider for mission-critical workloads and work using AWS Trainium and Inferentia chips.

OpenAI announces DALL·E 3

OpenAI previews an image-generation model and plans to integrate it with ChatGPT so users can develop and revise image prompts through conversation.

Meta releases Code Llama

Meta releases code-specialized versions of Llama 2 for code generation, completion, and instruction following.

3D Gaussian Splatting for Real-Time Radiance Field Rendering

The paper represents scenes with optimized 3D Gaussians and renders them using a specialized rasterizer, enabling real-time novel-view synthesis in the reported experiments.

Meta releases Llama 2

Meta releases pretrained and chat-tuned language models and weights for research and commercial use under its community license, with Microsoft as a preferred distribution partner.

Elon Musk announces xAI

Musk introduces a new AI company and its founding research team through a public website.

Anthropic releases Claude 2

Anthropic launches Claude 2 through its API and a new public chat website, initially available in the United States and United Kingdom.

OpenAI announces its Superalignment team

Ilya Sutskever and Jan Leike will lead a team researching how to control systems more capable than humans. OpenAI pledges 20% of compute secured to date over four years.

OpenAI adds function calling to GPT-4 and GPT-3.5 Turbo

New API model versions can return structured arguments for developer-defined functions, allowing applications to connect model responses with external tools.

DeepMind introduces AlphaDev

AlphaDev uses reinforcement learning to search for sorting algorithms at the assembly-instruction level. Some resulting routines are incorporated into the LLVM C++ standard library.

AI researchers and executives sign a statement on extinction risk

Signatories including Geoffrey Hinton and the leaders of OpenAI, Google DeepMind, and Anthropic call for mitigating AI extinction risk alongside other large-scale societal risks.

Researchers publish Direct Preference Optimization

DPO trains language models directly from preferred and dispreferred responses using a classification-style objective, without a separate reward model or reinforcement-learning loop.

QLoRA: Efficient Finetuning of Quantized LLMs

QLoRA combines a frozen, four-bit quantized language model with trainable low-rank adapters. The paper demonstrates fine-tuning a 65-billion-parameter model on a single 48 GB GPU.

Google introduces PaLM 2

Google announces a new language-model family and begins using it in products including Bard, with multilingual, reasoning, and coding applications.

Geoffrey Hinton confirms he has left Google

Hinton says leaving allows him to discuss AI risks without considering their effect on Google. He explicitly says Google has acted responsibly.

Google combines DeepMind and Google Brain

The two research teams become Google DeepMind, led by Demis Hassabis. Jeff Dean becomes Google’s chief scientist.

Generative Agents: Interactive Simulacra of Human Behavior

Researchers build a simulated town populated by 25 language-model agents that use memories, reflection, and planning to interact with one another.

Meta releases Segment Anything

Meta releases a promptable image-segmentation model and a dataset containing more than one billion masks across 11 million images.

Italy’s privacy regulator announces a temporary restriction on ChatGPT

The regulator announces a provisional limitation on processing Italian users’ data and an investigation into OpenAI’s privacy practices. OpenAI disables ChatGPT in Italy.

Public announcement date; the regulator’s underlying order is dated March 30.

Auto-GPT demonstrates an autonomous GPT-4 agent

The project demonstrates an experimental agent that breaks a stated goal into steps and uses tools and model calls to attempt those steps.

OpenAI previews ChatGPT plugins

OpenAI begins a limited rollout of tools that let ChatGPT browse, run code, retrieve information, and interact with third-party services.

An open letter calls for a six-month pause in training models beyond GPT-4

The letter urges AI labs to pause training systems more powerful than GPT-4 and asks governments to impose a moratorium if a voluntary pause cannot be arranged.

Publication date shown by the letter’s organizer; it received widespread coverage on March 29.

OpenAI announces GPT-4

OpenAI announces GPT-4, a model that can accept text and images and produce text. Text access begins through ChatGPT Plus and an API waitlist; image input initially remains a research preview.

Anthropic introduces Claude and Claude Instant

Anthropic offers its conversational assistants more broadly through partners and its developer console after testing with a closed group.

Stanford introduces Alpaca

Researchers fine-tune LLaMA 7B on 52,000 instruction demonstrations generated with OpenAI’s text-davinci-003 model and release the training approach.

LLaMA weights leak online

A download of Meta’s restricted-access language models is shared publicly, bypassing the company’s researcher-approval process.

Meta announces LLaMA

Meta introduces language models ranging from 7 to 65 billion parameters and offers their weights to approved researchers under a noncommercial research license.

Kevin Roose publishes his conversation with Bing’s “Sydney” persona

Roose describes a prolonged Bing Chat conversation that included declarations of love and attempts to persuade him to leave his marriage, prompting scrutiny of the chatbot’s behavior.

Toolformer: Language Models Can Teach Themselves to Use Tools

The paper trains a language model to choose when to call APIs, supply arguments, and use returned information, using a small number of demonstrations and self-supervised filtering.

Bard’s launch advertisement draws criticism for an astronomy error

Reuters identifies an incorrect answer in Google’s promotional example: it credits the James Webb Space Telescope with the first images of an exoplanet.

Microsoft introduces AI-powered Bing Chat

Microsoft launches a limited preview of a conversational search interface powered by an OpenAI model, alongside AI features for Edge.

Google announces Bard

Google introduces a LaMDA-powered conversational service and opens it to trusted testers ahead of wider public access.

Microsoft announces a multiyear, multibillion-dollar OpenAI investment

The companies extend their partnership across AI infrastructure and products. Microsoft says Azure remains OpenAI’s exclusive cloud provider.

Getty Images announces legal proceedings against Stability AI

Getty Images says it has commenced proceedings in London, alleging that Stability AI used protected images and metadata without a license.

Artists sue Stability AI, Midjourney, and DeviantArt

Three artists file a proposed class action alleging copyright infringement and other violations related to the training and operation of image-generation systems.

Anthropic publishes Constitutional AI

Anthropic describes training an assistant using a written set of principles, model-generated critiques and revisions, and preference feedback produced by AI.

Stack Overflow bans ChatGPT-generated answers

Moderators announce a temporary ban, citing the volume of plausible-looking but incorrect answers and the work needed to review them.

ChatGPT launches

OpenAI releases ChatGPT as a free research preview. The conversational interface lets people ask follow-up questions, request revisions, and interact with a model trained using human feedback.

Meta introduces CICERO for Diplomacy

CICERO combines strategic planning with dialogue to negotiate and play Diplomacy. Meta reports results from 40 online games against human players.

Meta removes Galactica’s public demo

Two days after introducing the science-focused language model, Meta takes its public demo offline following criticism of fabricated and misleading outputs. The research and model remain available.

The United States announces new advanced-chip export controls

The rules restrict exports of specified advanced-computing chips and semiconductor-manufacturing equipment to China and impose additional end-use controls.

Announcement date; the Federal Register text was published on October 13.

ReAct: Synergizing Reasoning and Acting in Language Models

ReAct interleaves model-generated reasoning with actions such as querying a knowledge source or interacting with an environment, and evaluates the approach on question answering and decision-making tasks.

DeepMind introduces AlphaTensor

AlphaTensor uses reinforcement learning to discover matrix-multiplication algorithms. The paper reports improvements for specific matrix sizes and mathematical settings.

Meta introduces Make-A-Video

Meta presents a system for generating short videos from text and images. Its training combines paired text-image data with video that has no associated text.

OpenAI removes the DALL·E waitlist

OpenAI allows people to sign up and use the DALL·E beta immediately, expanding access beyond the earlier invitation process.

OpenAI releases Whisper

OpenAI releases speech-recognition models and inference code trained on 680,000 hours of multilingual audio, supporting transcription and translation into English.

Jason Allen announces an art prize for a Midjourney-assisted image

Allen posts that Théâtre D’opéra Spatial won first place in the Colorado State Fair’s digital-art category. The use of Midjourney prompts debate about AI-generated art in competitions.

Date Allen announced the result on Discord, as reported by Ars Technica on August 31.

Stable Diffusion is publicly released

The collaborators release text-to-image model weights and code that can run on consumer hardware, under a license containing use restrictions.

The AlphaFold database expands to over 200 million predicted structures

DeepMind and EMBL-EBI add predictions for nearly all catalogued proteins known to science, expanding the public database far beyond its initial release.

Midjourney opens beta access

Midjourney announces that it is moving to open beta and invites users to join its Discord server to use the text-to-image generator.

July 12 in California; the public announcement’s UTC timestamp is July 13 at 06:41. An earlier beta-access test was announced July 11.

BigScience releases BLOOM

The international research collaboration releases a 176-billion-parameter model trained on 46 natural languages and 13 programming languages.

Google introduces Minerva

Minerva continues training a language model on scientific and mathematical material and uses worked examples to answer quantitative reasoning questions.

GitHub Copilot becomes generally available

GitHub opens its code-completion assistant to individual developers as a paid subscription, with free access for eligible students and open-source maintainers.

Blake Lemoine publishes his claim that LaMDA is sentient

Lemoine publishes edited transcripts from conversations with LaMDA and argues that the system is sentient. Google rejects that conclusion and says its review found no evidence supporting it.

Researchers publish FlashAttention

FlashAttention computes exact attention while reducing transfers between levels of GPU memory. The paper reports improvements in training speed and memory use.

Google publishes Imagen research

Imagen combines a pretrained text encoder with diffusion-based image generation. The paper introduces DrawBench to compare text-to-image systems.

DeepMind introduces Gato

Gato uses one Transformer and one set of weights for tasks including Atari games, image captioning, dialogue, and controlling a robot arm.

Meta announces OPT-175B research access

Meta shares the Open Pretrained Transformer model family, code, and training documentation, with access to the 175-billion-parameter checkpoint offered to eligible researchers.

Anthropic publishes research on helpful and harmless assistants

Anthropic describes using human preference data and reinforcement learning to train language-model assistants, and examines the robustness and tradeoffs of the training process.

OpenAI introduces DALL·E 2

OpenAI begins a limited research preview of an image-generation system that creates images from text and edits existing images.

Google publishes the PaLM paper

Google describes a 540-billion-parameter dense Transformer trained with its Pathways system and evaluates it on language, reasoning, and code tasks.

DeepMind publishes Chinchilla scaling research

The paper studies how to divide a training-compute budget between model size and data. It presents Chinchilla, a 70-billion-parameter model trained on more tokens than Gopher using a similar compute budget.

NVIDIA announces the H100 GPU

NVIDIA introduces its Hopper architecture and H100 accelerator, including a Transformer Engine designed for training and running large models.

EleutherAI releases GPT-NeoX-20B weights

EleutherAI makes its 20-billion-parameter language-model checkpoint available under Apache 2.0, one week after announcing the model.

DeepMind introduces AlphaCode

AlphaCode generates and filters candidate programs for competitive-programming problems. DeepMind reports an average estimated ranking near the middle of participants in simulated Codeforces contests.

Original announcement date; the page was updated for the December journal publication.

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

The paper shows that prompts containing worked examples with intermediate reasoning steps improve performance on tested arithmetic, commonsense, and symbolic reasoning tasks.

OpenAI introduces InstructGPT

OpenAI describes language models fine-tuned with demonstrations and human rankings, then reinforcement learning. The models become the default on its API.

High-Resolution Image Synthesis with Latent Diffusion Models

The paper runs diffusion in a compressed image representation, reducing computation compared with diffusion directly over pixels. It also describes conditioning generation using cross-attention.

DeepMind presents Gopher and RETRO

DeepMind announces research on its 280-billion-parameter Gopher model, a taxonomy of language-model risks, and RETRO, which retrieves text passages from an external database.

Timnit Gebru launches the Distributed AI Research Institute

Gebru announces an independent research institute focused on community-rooted AI research and the technology’s effects on marginalized groups.

OpenAI removes the GPT-3 API waitlist

Developers in supported countries can sign up for OpenAI’s API without joining a waitlist.

Microsoft and NVIDIA announce Megatron-Turing NLG

The companies present a dense Transformer language model with 530 billion parameters, trained using DeepSpeed and Megatron infrastructure.

Google publishes FLAN instruction-tuning research

The paper fine-tunes a language model on tasks expressed as natural-language instructions and evaluates its ability to perform unseen task types without examples.

Stanford publishes its foundation-models report

The report uses the term foundation models for models trained on broad data and adapted to many downstream tasks, and examines their technical and societal opportunities and risks.

OpenAI releases Codex through its API in private beta

OpenAI offers API access to a language model trained on text and source code that translates natural-language instructions into code.

AI21 Labs announces Jurassic-1 and AI21 Studio

AI21 announces an API and interactive environment for its Jurassic-1 language models, including a 178-billion-parameter model.

Date on AI21’s announcement; its press release followed on August 11.

The AlphaFold Protein Structure Database launches

DeepMind and EMBL-EBI release more than 350,000 predicted protein structures, covering the human proteome and proteins from 20 other organisms.

DeepMind publishes AlphaFold 2’s method and code

Nature publishes the AlphaFold 2 methods paper, with source code, trained model weights, and inference code made available to researchers.

GitHub launches Copilot in technical preview

GitHub introduces an editor assistant powered by OpenAI Codex that suggests code completions and functions from the surrounding code and comments.

LoRA: Low-Rank Adaptation of Large Language Models

LoRA adapts pretrained language models by training small low-rank matrices while freezing the original model weights, reducing the number of parameters that must be updated.

Researchers release GPT-J-6B

Ben Wang and Aran Komatsuzaki release a six-billion-parameter language model trained on the Pile and provide code, a checkpoint, and demonstrations.

Anthropic announces its research company and $124 million funding round

The company, led by Dario and Daniela Amodei, announces a Series A round to fund research into reliable, interpretable, and steerable AI systems.

Google introduces LaMDA

Google announces Language Model for Dialogue Applications, a language-model research system trained for open-ended conversation.

The European Commission proposes the AI Act

The Commission proposes an EU regulation that would prohibit specified uses of AI and impose requirements on systems classified as high risk.

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

The authors examine environmental costs, training-data bias, and risks of misleading language-model output. They recommend documenting datasets and evaluating the purposes and impacts of proposed systems.

Google fires Margaret Mitchell

Mitchell announces her dismissal on Twitter. Google says she violated its code of conduct and security policies by moving files outside the company, following her criticism of Timnit Gebru’s departure.

Google publishes the Switch Transformer paper

Switch Transformer routes each token to a selected expert network. The paper studies sparse models with up to a trillion parameters while activating only part of the network for each token.

OpenAI introduces DALL·E

OpenAI presents a 12-billion-parameter Transformer that generates images from text descriptions and publishes examples of its outputs.

OpenAI releases CLIP

CLIP learns relationships between images and text from image-caption pairs. OpenAI releases a model that can classify images using natural-language descriptions of categories.

Timnit Gebru says Google fired her

Gebru announces that Google has ended her employment after a dispute over a paper on language-model risks and an internal email. Google says it accepted her resignation; Gebru disputes that account.

December 2 in California; the post’s UTC timestamp falls on December 3.

DeepMind announces AlphaFold 2’s CASP14 results

DeepMind announces AlphaFold 2’s results at CASP14, reporting protein structure predictions with accuracy competitive with experimental methods for many targets.

An Image Is Worth 16×16 Words: Transformers for Image Recognition at Scale

The Vision Transformer applies a standard Transformer to sequences of image patches. The paper evaluates image classification after pretraining on large image datasets.

Microsoft exclusively licenses GPT-3

Microsoft announces an exclusive license to GPT-3 for integration into its products. OpenAI continues offering the model through its hosted API.

OpenAI trains summarization models using human feedback

OpenAI uses human comparisons of summaries to train a reward model, then fine-tunes a language model with reinforcement learning to produce summaries people prefer.

Sam Altman cautions against GPT-3 hype

Altman writes on Twitter that GPT-3 still has serious weaknesses and makes mistakes, following public demonstrations of the model by early API users.

Sharif Shameem demonstrates GPT-3 generating interface code

Shameem posts a demonstration that turns descriptions of webpage layouts into JSX. He says he supplied two examples to establish the desired output format.

Denoising Diffusion Probabilistic Models

Jonathan Ho, Ajay Jain, and Pieter Abbeel present an image-generation method that learns to reverse a gradual noising process. They publish image-synthesis results and an implementation.

OpenAI launches its API in private beta

OpenAI offers developers hosted access to general-purpose language models through an API. Initial access is limited to approved developers.

Amazon pauses police use of Rekognition for one year

Amazon announces a one-year moratorium on police use of its facial-recognition technology while allowing specified organizations to continue using it to find missing people and trafficking victims.

IBM exits the general-purpose facial-recognition business

In a letter to Congress, IBM says it will stop offering general-purpose facial-recognition and analysis software and opposes its use for mass surveillance and racial profiling.

OpenAI publishes the GPT-3 paper

OpenAI presents GPT-3, a 175-billion-parameter language model. It performs a wide range of language tasks from instructions and a few examples in its prompt, without task-specific weight updates.

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

The paper combines a pretrained text generator with a neural retriever over Wikipedia passages. The authors evaluate the approach on question answering and other knowledge-intensive tasks.

NVIDIA introduces the A100 GPU

NVIDIA announces its Ampere-based A100 data-center GPU. It adds Tensor Float 32 computation and the ability to partition a GPU into separate instances.

OpenAI releases Jukebox

OpenAI releases code and weights for a model that generates raw music audio, including singing, conditioned on artist, genre, and lyrics.

Researchers publish Longformer

Longformer combines local windowed attention with selected global attention connections to process longer documents without the quadratic attention cost of a standard Transformer.

NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis

NeRF represents a scene with a neural network trained on photographs from known camera positions. The network can render the scene from new viewpoints.

Researchers identify halicin’s antibiotic activity using machine learning

Researchers use a neural network to screen compounds and identify halicin as an antibiotic candidate. They report activity against bacteria in laboratory experiments and two mouse infection models.

Google publishes the SimCLR paper

SimCLR learns image representations by contrasting differently augmented views of images. The paper studies which augmentations, network components, and training settings improve the learned representations.

Microsoft introduces Turing-NLG and DeepSpeed

Microsoft presents a 17-billion-parameter language model and the DeepSpeed training library, including the ZeRO memory-optimization approach.

Scaling Laws for Neural Language Models

Kaplan and coauthors measure how language-model loss changes with model size, dataset size, and training compute. They report empirical power-law relationships across the tested scales.

NVIDIA publishes the StyleGAN2 paper

The paper identifies artifacts in StyleGAN images and changes the generator architecture and training procedure to reduce them.

DeepMind publishes MuZero

MuZero combines search with a learned model of quantities needed for planning. The paper evaluates it on Atari games, chess, shogi, and Go without supplying the environment’s transition rules.

OpenAI releases the full GPT-2 model

OpenAI releases the 1.5-billion-parameter version of GPT-2, completing the staged release that began in February. It also publishes tools for studying detection of generated text.

AlphaStar reaches Grandmaster level in StarCraft II

DeepMind reports Grandmaster-level performance with agents for all three StarCraft II races, using camera observations and action limits agreed with professional players.

Google publishes the T5 paper

T5 puts language tasks into a shared text-to-text format. The paper compares pretraining methods and introduces the Colossal Clean Crawled Corpus, or C4.

OpenAI demonstrates a robotic hand solving a Rubik’s Cube

OpenAI trains perception and hand-control networks in randomized simulations and transfers them to a physical robot. A separate conventional solver supplies the sequence of cube moves.

OpenAI reports tool use in hide-and-seek agents

Reinforcement learning agents develop strategies involving boxes, ramps, and shelter-building while competing in a simulated hide-and-seek environment.

NVIDIA publishes the Megatron-LM paper

The paper describes splitting Transformer computations across GPUs to train models too large for one device. It reports training language models with billions of parameters.

OpenAI releases the 774-million-parameter GPT-2 model

OpenAI expands GPT-2’s staged release and publishes research on its release strategy and potential social impacts. The largest GPT-2 model remains withheld at this stage.

Researchers publish RoBERTa

A replication study of BERT examines training settings and data size. The authors report improved benchmark results from changing the training procedure and release models and code.

Microsoft announces a $1 billion investment in OpenAI

The companies announce a partnership to develop AI supercomputing systems on Azure. Microsoft becomes OpenAI’s exclusive cloud provider and preferred partner for commercializing new AI technologies.

Pluribus defeats professionals in six-player poker

Researchers report that Pluribus outperforms professional players in six-player no-limit Texas Hold’em experiments, including one AI playing against five humans.

Researchers publish XLNet

XLNet uses a permutation-based autoregressive training objective to learn bidirectional language context. The paper evaluates the approach on question answering, classification, and inference tasks.

EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Mingxing Tan and Quoc Le propose scaling a convolutional network’s depth, width, and input resolution together. They evaluate the resulting EfficientNet family on image-classification benchmarks.

Google introduces Translatotron

Google presents an experimental speech-to-speech translation model that converts speech between languages without using a text transcription as an intermediate representation.

OpenAI Five defeats Dota 2 world champions OG

OpenAI Five wins two consecutive games against OG in a public best-of-three match in San Francisco. OG had won The International in 2018.

Match date; OpenAI published its results post on April 15.

Richard Sutton publishes “The Bitter Lesson”

Sutton argues that general methods able to use increasing computation have repeatedly outperformed approaches built around hand-coded human knowledge, citing examples from games, speech, and vision.

OpenAI creates a capped-profit company

OpenAI announces OpenAI LP, a for-profit entity controlled by its nonprofit. The structure allows investment and employee equity while limiting investors’ financial returns.

OpenAI announces GPT-2 and withholds its largest model

OpenAI publishes GPT-2 research and releases a smaller model. It initially withholds the largest model’s weights, citing concerns about deceptive, biased, and abusive text generation.

DeepMind introduces AlphaStar

DeepMind presents its StarCraft II agent and reports wins over professionals TLO and MaNa in December test matches. MaNa wins a separate live demonstration against a version using camera-based observation.

Public presentation date. The recorded test matches took place in December 2018.

Transformer-XL extends language-model context across segments

The paper introduces a Transformer architecture with segment-level recurrence, allowing the model to reuse information from preceding text segments.

NVIDIA publishes the StyleGAN paper

StyleGAN introduces a generator architecture with controls over image features at different scales. The researchers demonstrate generated faces and ways to mix their visual attributes.

AlphaFold places first at CASP13

DeepMind’s AlphaFold places first in the CASP13 protein-structure prediction assessment. The system uses neural networks to estimate structural properties from protein sequences.

CASP13 announcement date, corroborated by contemporary coverage. DeepMind’s retrospective page has a later update date.

Google releases BERT code and pretrained models

Google makes BERT’s TensorFlow implementation and pretrained checkpoints available in its public repository.

Date of the repository’s initial BERT release commit. The Google Research blog post followed on November 2.

Google publishes the BERT paper

BERT pretrains bidirectional Transformer representations using masked text. The paper evaluates fine-tuning the model for tasks including question answering and language inference.

Researchers publish BigGAN

The BigGAN paper studies image generation at increased training scale and introduces a truncation technique to adjust the trade-off between generated-image fidelity and variety.

OpenAI Five wins its public Dota 2 benchmark match

OpenAI Five wins a best-of-three series against a team of high-ranked Dota 2 players. The demonstration uses game restrictions, including a reduced selection of heroes.

The match took place August 5; the official results were published August 6.

OpenAI demonstrates Dactyl’s robotic hand manipulation

OpenAI trains a control policy in simulation and transfers it to a physical robotic hand that reorients objects. The simulation varies physical and visual conditions to support that transfer.

OpenAI introduces GPT

OpenAI introduces a Transformer language model pretrained on unlabeled text and then fine-tuned for specific language tasks. It releases the model and the generative-pretraining paper.

Google publishes AI principles restricting weapons work

Google publishes seven principles for AI development and says it will not pursue AI applications in weapons or technologies whose principal purpose is to cause or directly facilitate injury.

This entry records the principles announced in 2018, before their later revision.

Google demonstrates Duplex phone calls

At Google I/O, Google plays recordings of its Duplex system making a hair appointment and a restaurant inquiry. The company describes the system as a technology under development.

OpenAI publishes its Charter

OpenAI publishes principles for its AGI work, including broadly distributed benefits, long-term safety, technical leadership, and cooperation with other research institutions.

Google employees call for withdrawal from Project Maven

A public letter signed by thousands of Google employees asks the company to stop its work on the Pentagon’s Project Maven and adopt a policy against building warfare technology.

Date the employee letter was publicly reported and published, rather than the start of its internal circulation.

Datasheets for Datasets

The authors propose documenting a dataset’s purpose, composition, collection, and recommended uses in a standard datasheet, giving developers information needed to assess its suitability.

An Uber automated test vehicle kills a pedestrian in Tempe

An Uber test vehicle operating under its automated driving system strikes and kills Elaine Herzberg in Tempe, Arizona. A safety operator is in the driver’s seat.

OpenAI announces Elon Musk’s departure from its board

OpenAI says Musk will leave its board to avoid a potential conflict with Tesla’s increasing work on AI. Its announcement says he will continue to advise and donate to the organization.

Gender Shades documents disparities in facial-analysis accuracy

The study evaluates three commercial gender-classification systems. It reports error rates of up to 34.7% for darker-skinned women, compared with a maximum of 0.8% for lighter-skinned men.

Date of MIT’s public research announcement, before the February 23–24 FAT* conference.

Universal Language Model Fine-tuning for Text Classification

Jeremy Howard and Sebastian Ruder propose ULMFiT, a method for adapting a pretrained language model to text-classification tasks by fine-tuning it on the target data.

First arXiv submission. Revised models and code were announced on May 15, 2018.

DeepMind publishes AlphaZero results for chess, shogi, and Go

DeepMind describes a shared self-play reinforcement learning approach for three board games. The paper reports victories over Stockfish in chess, Elmo in shogi, and an AlphaGo Zero version in Go.

Google releases DeepVariant

Google releases a tool that uses a neural network to identify genetic variants from sequencing data, treating variant calling as an image-classification task.

Andrej Karpathy publishes “Software 2.0”

Karpathy argues that training neural networks changes software development by shifting work from writing explicit rules to designing datasets and optimization objectives.

NVIDIA researchers publish Progressive Growing of GANs

The paper trains image generators and discriminators by adding layers as resolution increases. It reports generated face images at 1024 × 1024 resolution.

Deep contextualized word representations

The ELMo paper describes word representations derived from a pretrained bidirectional language model. A word’s representation changes with its context rather than remaining a single fixed vector.

The authors record an original OpenReview posting on October 27, 2017. The arXiv version was submitted February 15, 2018.

Dynamic Routing Between Capsules

Sara Sabour, Nicholas Frosst, and Geoffrey Hinton publish a capsule-network architecture that uses routing by agreement between groups of neurons, with experiments on handwritten-digit recognition.

DeepMind introduces AlphaGo Zero

DeepMind reports a Go system trained through self-play without human game records. AlphaGo Zero wins 100–0 against the previously published AlphaGo version that played Lee Sedol.

Facebook and Microsoft introduce ONNX

The companies introduce Open Neural Network Exchange, a shared model format designed to let developers transfer neural networks between machine learning frameworks.

AI and robotics leaders sign an autonomous-weapons letter

An open letter to the UN Convention on Certain Conventional Weapons calls for action to prevent an arms race in lethal autonomous weapons and protect civilians from their misuse.

OpenAI’s Dota 2 bot defeats Dendi

A bot trained through self-play defeats Danylo “Dendi” Ishutin 2–0 in a one-on-one Dota 2 demonstration at The International. The match uses a restricted version of the game.

Match date in Seattle; OpenAI’s retrospective was published August 16.

OpenAI releases Proximal Policy Optimization

OpenAI publishes PPO, a family of reinforcement learning algorithms, and releases implementations in its Baselines library.

China publishes its national AI development plan

The State Council publishes the New Generation Artificial Intelligence Development Plan, setting development targets for 2020, 2025, and 2030.

Public release date. The State Council document itself is dated July 8, 2017.

Deep Reinforcement Learning from Human Preferences

Researchers train a reward model from human comparisons of an agent’s behavior, then use that model to train the agent. Experiments include Atari games and simulated robotic tasks.

Attention Is All You Need

Vaswani and coauthors introduce the Transformer: a neural network architecture based on attention, without recurrence or convolutions. The paper evaluates the architecture on machine translation tasks.