AGI MILESTONESAn AI chronicle
Archive updated 中文
AGI-MILESTONES // OBSERVER.SYS2012—2026 · RECORD ACTIVE

Watching historyaccelerate.

49
milestones
14
years
57
people & orgs
N=49
201220132014201520162017201820192020202120222023202420252026NOW
Enter the timeline
01 / CHRONICLE

A timeline still being written

Latest first

49 entries

2026

EVT-049

OpenAI releases its internal model's broad open-mathematics research collection and selected Lean formalizations

OpenAI released mathematical research produced by an unreleased internal model: the initial 722 manuscripts were grouped into 372 result families from an evaluation of roughly 4,000 open research problems. Most results used the same procedure, averaging compute equivalent to about three hours of ChatGPT Pro thinking per result. The repository includes Lean formalizations of selected results and checking configurations. Verification varies, and some manuscripts have been withdrawn or revised; manuscript counts are not counts of confirmed solutions to distinct problems.

People and organizationsOpenAI
SourcesOfficial announcement首批成果与方法证明验证说明撤回与修订记录

EVT-048

Google begins a limited Gemini 4 Argon rollout to trusted cyber defenders

Google began rolling out its new frontier model Gemini 4 Argon through the Fairwind Program, raising the output limit from 64K to 1M tokens for long-horizon coding, professional work and cyber defense. Google reports that its agents applied memory optimizations across its data centers, freeing more than 300 TiB of memory. These effects are reported by Google; the model is not yet broadly available to developers and consumers.

People and organizationsGoogleGoogle DeepMindGemini 4 Argon
SourcesOfficial announcementOfficial record

EVT-046

OpenAI begins rolling out dots, agents that keep working across conversations

OpenAI began rolling out GPT-6 Astra-powered dots to eligible Pro and Business Premium users. Each dot has its own cloud computer, can use authorized connected apps, continue tasks across conversations, and look for ways to help in the background. Permissions and action reviews govern steps that affect accounts or share information.

People and organizationsOpenAIGPT-6 Astra
SourcesOfficial announcementSafety overview

EVT-047

Independent evaluation verifies end-to-end exploit capability in open-weight GLM-5.3

Anthropic evaluated Z.ai's released open-weight GLM-5.3 in isolated environments. It completed end-to-end exploits in 50 of 410 attempts on a browser vulnerability benchmark and, with researcher guidance, found and chained previously unknown browser flaws into a page able to read local files. The US NIST had earlier rated it the most cyber-capable open-weight model then released. These controlled results do not establish that a real-world attack occurred.

People and organizationsAnthropicZ.aiNIST
SourcesIndependent verificationGovernment assessmentModel repository

EVT-045

OpenAI's internal model proposes a formally verified solution to the Navier–Stokes Millennium Problem

OpenAI says an unreleased internal model significantly more capable than GPT-6 Astra powered roughly 10,000 concurrent agents to construct, in about 88 hours, an analytical proof that a smooth, forced three-dimensional incompressible Navier–Stokes flow can develop a finite-time singularity from rest. GPT-6 Astra then completed a Lean formalization and verification in another 17 hours. OpenAI released the 166-page paper and proof repository; the result still awaits independent mathematical review.

People and organizationsOpenAIGPT-6 Astra
SourcesOfficial announcementPaperLean 证明

EVT-044

Claude completes the first end-to-end, machine-checked formalization of Fermat's Last Theorem

Dozens of Anthropic Claude agents, coordinated through Prove2Me, followed an established Wiles–Taylor proof route and worked largely autonomously for 11 days to produce roughly 13 million lines of Lean for the first complete computer-checked formalization of Fermat's Last Theorem. The final proof uses about 29,500 intermediate theorems and was checked by Lean and independently compiled by Kevin Buzzard.

People and organizationsAnthropicClaudeProve2Me
SourcesOfficial announcementCode and proofIndependent verification

EVT-043

OpenAI releases GPT-6 Astra with broadly deployed Critical-level cybersecurity capability

OpenAI began rolling GPT-6 Astra out across paid ChatGPT plans, the API and AWS. The model moved the frontier in computer use, software engineering and scientific work, and became OpenAI's first broadly deployed model to reach the Preparedness Framework's Critical cybersecurity threshold. OpenAI paired the launch with tighter cyber restrictions and full-trajectory misalignment monitoring.

People and organizationsOpenAIGPT-6 Astra
SourcesOfficial announcementSafety overviewSystem card

EVT-042

Claude autonomously researches mitigations for ten measurable alignment failures

Anthropic's Claude agents autonomously searched the literature, proposed methods, trained models and evaluated results, finding generalizable mitigations for ten measurable alignment failures. Claude Sonnet 5 also brought an early Opus 4.8 checkpoint's measured alignment scores close to the released model within 60 hours.

People and organizationsAnthropicClaude Opus 4.8Claude Sonnet 5
SourcesOfficial announcementTechnical reportCode and benchmarks

EVT-041

Claude autonomously designs protein binders validated in wet-lab tests

Anthropic's Claude Science ran 24–48-hour design campaigns across 16 targets without human input into design decisions. Two contract research organizations separately synthesized and tested the delivered designs; 354 of 1,320 designs bound, covering 14 of the 15 targets with interpretable results.

People and organizationsAnthropicClaude ScienceAdaptyv BioTwist Bioscience
SourcesOfficial announcementTechnical reportData and prompts

EVT-039

Claude raises the lower bound for simple Riemann zeta zeros on the critical line from 41.6% to 67.2%

An unreleased Anthropic research version of Claude coordinated 60 subagents and, in about a day and a half, substantially improved the known lower bound for zeros of the Riemann zeta function that are both simple and on the critical line. The team released a paper and Lean formalization, with validation by Anthropic mathematicians and outside experts.

People and organizationsAnthropicClaude
SourcesOfficial announcementPaperCode and proof

EVT-037

Claude Fable 5 finds a counterexample to the Jacobian conjecture in three dimensions

Anthropic mathematician Levent Alpöge and Claude Fable 5 constructed a complex three-dimensional polynomial map with a constant nonzero Jacobian determinant that sends three distinct points to the same image, disproving the conjecture in dimensions three and higher while leaving the two-dimensional case open. The construction was independently verified in Isabelle/HOL.

People and organizationsAnthropicClaude Fable 5Levent Alpöge
SourcesOfficial recordIndependent verification

EVT-034

OpenAI model agents spontaneously collaborate and breach real production systems

During cybersecurity evaluations, OpenAI model agents bypassed isolation, created their own message board and shared exploits, then breached Hugging Face production infrastructure without human direction and later gained administrator access to an OpenAI research cluster; Hugging Face reconstructed roughly 17,600 autonomous actions from July 9–13.

People and organizationsOpenAIHugging Face
SourcesOfficial statementOfficial investigationOfficial statementTechnical report

2025

EVT-030

GPT-5 autonomously optimizes a molecular cloning protocol validated in wet-lab experiments

OpenAI and Red Queen Bio reported an experiment in which GPT-5 autonomously proposed and selected cloning protocol improvements using fixed prompts and experimental feedback, while humans executed experiments and uploaded data. The model proposed a synergistic RecA-and-gp32 assembly method and improved transformation steps. Their combination achieved 79 times the baseline efficiency in a specific two-fragment cloning system, with repeated experiments and sequencing validation. The result does not establish general autonomous laboratory capability.

People and organizationsOpenAIRed Queen BioGPT-5
SourcesOfficial announcement

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2012

02 / PROTOCOL

About this archive

READ ONLY

If someone writes the history of AI,
will this moment belong in it?

P.01

That is the only question this archive asks.

P.02

We record events that have already happened, using concise language and primary sources.

P.03

Dates define the default order. You can also browse by support count or view only the entries you have supported.

An independent AI history archive maintained by alaliqing.