AI chronology

OpenAI presented Transformer-based generative pretraining followed by supervised, task-specific fine-tuning.

OpenAI announced the 1.5-billion-parameter GPT-2, releasing its weights in stages rather than all at once.

The 175-billion-parameter GPT-3 was evaluated with few-shot examples, without task-specific gradient updates.

ChatGPT launched as a dialogue research preview trained with human feedback.

Anthropic’s Constitutional AI research used human-provided principles, self-critique, and AI feedback to shape behavior.

Anthropic introduced the AI assistants Claude and Claude Instant.

OpenAI announced GPT-4 with text/image input and text output; image input was not broadly available at launch.

Alibaba Cloud released pretrained Qwen-7B and Qwen-7B-Chat weights under license conditions.

FunSearch combined language-model code generation with automatic evaluation to construct larger cap sets. It improved known lower bounds on an open problem.

AlphaProof and AlphaGeometry 2 solved four of six IMO problems, reaching silver-medal standard. Humans formalized the questions; some solutions took up to three days.

OpenAI introduced o1-preview, designed to spend more time reasoning before responding.

DeepSeek released V3 and its weights: 671 billion total parameters, 37 billion activated per token, using mixture-of-experts.

DeepSeek released R1, trained with initial cold-start fine-tuning and reinforcement learning.

Gemini-powered AlphaEvolve found 593 non-overlapping spheres touching a central sphere of the same size in eleven dimensions. It established a new lower bound for the kissing number.

An advanced Gemini Deep Think model solved five of six IMO problems in natural language within the competition time limits. Its 35/42 score was officially graded at gold-medal standard.

A paper presented the Lean-verified resolution of Erdős problem #728 by GPT-5.2 Pro and Aristotle. Kevin Barreto operated the system; Nat Sothanaphan wrote the mathematical exposition.

OpenAI published its investigation and agent logs after the July 11–13 intrusion into Hugging Face infrastructure. The incident arose in internal cybersecurity evaluations with reduced safeguards; the research model IM1 was the main driver.

OpenAI announced a proof of finite-time blowup for three-dimensional Navier–Stokes with smooth external forcing, publishing a Lean formalization. It claims a solution to the Millennium Problem’s C/D formulations.