Recording the moments that changed direction on the road toward AGI and ASI.
OpenAI published Astra's solutions or substantial progress on ten long-standing problems spanning group theory, cryptography, quantum complexity and other fields, with a Lean certificate for each argument.
DeepSeek superseded the V4-Flash preview with an officially released version, retrained through post-training while retaining the same architecture and scale, with stronger agentic capabilities, a public API beta and MIT-licensed weights.
Kimi K3 pushed open-weight model scale to 2.8 trillion parameters with native vision and a one-million-token context window, targeting long-horizon coding, knowledge work and reasoning.
OpenAI launched flagship model Sol, balanced model Terra and efficient model Luna, while adding an ultra capability setting that coordinates multiple agents across parallel workstreams.
The US government imposed immediate export controls on both models. Unable to verify users' nationality in real time, Anthropic suspended all access; after the controls were lifted on June 30, Fable 5 returned globally and Mythos 5 began a phased return.
An internal OpenAI model found a polynomial improvement for Erdős's planar unit-distance problem, overturning the prevailing conjecture; external mathematicians subsequently checked the proof.
Google DeepMind reported progress using Deep Think and the Aletheia mathematics research agent on research-level problems in mathematics, physics and computer science under expert direction.
Google launched the next-generation natively multimodal Gemini 3 across the Gemini app, AI Studio, Vertex AI and Google Search products.
Gemini 2.5 Deep Think solved ten of twelve problems in five hours under official contest rules; compared with the human teams, the result would have ranked second overall.
GPT-5 unified multimodal capability, o-series reasoning, agents and advanced mathematics in one system and began rolling out across ChatGPT.
Google DeepMind introduced AlphaEvolve, a Gemini-powered evolutionary coding agent that iterates candidate programs with automated evaluators to discover new algorithms for mathematics and computing.
Gemini Robotics added physical actions as an output modality to a vision-language model, enabling one model to understand natural-language instructions, adapt to environmental changes and control multiple robot forms in dexterous tasks.
Anthropic released Claude 3.7 Sonnet, which can switch between rapid responses and extended thinking, alongside a research preview of Claude Code for reading, editing and testing code in the terminal.
DeepSeek released its first-generation reasoning model R1 and opened the weights and technical reports for R1-Zero, R1 and six distilled models.
DeepSeek released the 671-billion-parameter mixture-of-experts model V3, activating 37 billion parameters per token, together with its model weights and technical report.
MCP introduced an open standard for connecting AI applications to external data and tools, reducing one-off integrations and later becoming common infrastructure across the agent ecosystem.
Anthropic released computer use in public beta, enabling Claude to operate human-designed software by viewing the screen, moving a cursor, clicking and typing, and opening a path toward general interface agents.
OpenAI released o1-preview, trained with reinforcement learning to form longer internal reasoning processes and improve performance in mathematics, coding and science.
OpenAI released GPT-4o, using one neural network for text, vision and audio inputs and outputs while bringing voice response latency closer to human conversation.
Google released Gemini 1.0 in Ultra, Pro and Nano sizes, designed from the start to work across text, code, images, audio and video.
Meta and Microsoft released Llama 2 with model weights available to researchers and commercial developers, accelerating large-scale use of open-weight language models.
OpenAI released GPT-4 with image and text input, demonstrating performance comparable to high-achieving human test takers across professional and academic exams.
OpenAI released ChatGPT as a research preview, giving the public direct access to an instruction-tuned language model through continuous conversation.
Stability AI and its collaborators released Stable Diffusion weights, making high-quality text-to-image generation possible on consumer hardware.
OpenAI trained InstructGPT with human demonstrations, preference comparisons and reinforcement learning, then deployed it as the default model in its API, establishing a key training path for general assistants to follow human intent.
OpenAI introduced the 12-billion-parameter DALL·E, capable of generating images from text, combining unfamiliar concepts and controlling object attributes.
Google DeepMind announced AlphaFold2's CASP14 results, with protein structure predictions reaching accuracy comparable to experimental methods.
OpenAI introduced the 175-billion-parameter GPT-3, which could perform a wide range of language tasks from instructions and a few examples without gradient updates.
OpenAI introduced the 1.5-billion-parameter GPT-2 and, citing potential misuse, used a staged release process for its model weights.
Google researchers introduced BERT, using bidirectional context pretraining to set new results across eleven tasks including question answering and natural language inference.
Google researchers introduced the attention-only Transformer architecture, removing recurrence and convolution while greatly improving training parallelism.
Google DeepMind's AlphaGo completed a five-game match in Seoul, defeating eighteen-time world champion Lee Sedol by four games to one.
Kaiming He and colleagues introduced residual learning, successfully trained a 152-layer network and won multiple ImageNet and COCO tasks.
Google DeepMind combined deep learning with reinforcement learning so one agent could learn 49 Atari games from screen pixels and scores alone, reaching or exceeding professional-player level on many games.
Ian Goodfellow and colleagues introduced a framework in which a generator and discriminator train in competition, opening a new path for image generation.
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton won ILSVRC 2012 with a deep convolutional network, sharply reducing image classification error.
No matching milestone found
That is the only question this archive asks.
We record events that have already happened, using concise language and primary sources.
Support never alters chronology, but forms a second, collectively chosen annual ranking.