Elon Musk, Sam Altman, Greg Brockman, and others founded OpenAI with the goal of building AGI to benefit humanity.

DeepMind's AlphaGo defeats top human player Lee Sedol in the board game Go, defying what many considered possible.

Facebook releases PyTorch, a Python-first deep learning framework that would eventually become the dominant framework for AI research.

Organized by the Future of Life Institute, all of the top names in the field gather for a conference in Asilomar to discuss how to build AGI beneficially.

Google introduces the transformer architecture, a breakthrough deep learning architecture based on the attention mechanism. The architecture shows strong gains on language translation tasks.

Christiano et al. publish the technique of reinforcement learning from human feedback (RLHF), which would later be used extensively to align large language models.

OpenAI introduces Proximal Policy Optimization, a simpler and more stable policy gradient method that would become widely used across many reinforcement learning domains, including RLHF.

OpenAI reveals the first version of its Generative Pre-trained Transformer (GPT).

Google releases BERT, an encoder language model that would become ubiquitous in NLP.

OpenAI releases GPT-2, but withholds the largest version due to concerns about misuse. A decoder-only transformer, it was trained with next token prediction to generate text.

Kaplan et al. release "Scaling Laws for Neural Language Models", showing that model performance predictably scales with compute, data, and parameters. Scaling laws would become the primary driver of progress over the next few years.

OpenAI releases GPT-3, the largest language model at the time and shocking in its ability to generate coherent paragraphs.

DeepMind introduces MuZero, which learned to master Go, chess, shogi, and Atari games without any knowledge of the rules.

OpenAI introduces DALL-E, a model that generates images from textual descriptions.

A group of researchers from OpenAI leave to start Anthropic, featuring an empirical hard-science culture focused on AI safety.

A team at Microsoft publishes Low-Rank Adaptation (LoRA), a technique that allows for fine-tuning of large language models with a fraction of the compute and would later become ubiquitous

Github previews Copilot in VSCode, a tool that used OpenAI's Codex model to generate code suggestions, marking the beginning of real-world AI code generation.

OpenAI introduces InstructGPT, a model that can follow instructions in natural language better than base GPT-3 and was a prototype of what would become ChatGPT.

Google Brain publishes a paper showing gains in LLM reasoning by having them think step-by-step. Despite being such a simple technique, chain-of-thought reasoning would become foundational to AI.

DeepMind publish the Chinchilla paper, issuing a correction to the scaling laws of Kaplan et al. and suggesting that model size and training data should be scaled in equal proportion.

OpenAI shocks the world with the release of DALL-E 2, which could generate realistic images from text at an unprecedented level.

DeepMind publish Gato in a paper titled "A Generalist Agent". Gato used a single large transformer to learn policies for 604 different RL tasks across varying modalities and observation types.

A group from Stanford release Flash Attention, a new method that significantly speeds up the attention mechanism in transformers.

A Google engineer named Blake Lemoine is fired after claiming that its LaMDA model was sentient. The story woke a lot of people up to the capabilities and risks of LLMs.

Google Research introduces Minerva, an LLM specialized to solve quantitative reasoning problems. Minerva moved the state-of-the-art on the MATH benchmark from 6.9% to 50.3%, surprising many who doubted that LLMs could ever be trained to do math well.

Anonymous Twitter personalities Beff Jezos and Bayeslord start effective accelerationism (e/acc), advocating for rapid AI development. Though initially a meme, it later gains prominence in Silicon Valley and serves as a foil to cautionary AI safety voices.

DeepMind releases AlphaFold 2, solving the protein folding problem and revolutionizing a grand challenge in biology.

Stability AI open-sources Stable Diffusion (v1.4), the first strong image generation to be released to the public.

Anthropic publishes a paper on a phenomenon called superposition, where the neural network packs more features than the dimensions of its representational space, posing a challenge for mechanistic interpretability.

At Tesla's first ever "AI Day", they unveil Optimus, their plan to build a humanoid robot.

The U.S. Bureau of Industry and Security implements comprehensive export controls restricting China's access to advanced semiconductors, chip manufacturing equipment, and supercomputer components, marking a significant shift in U.S. technology policy toward China.

OpenAI releases a blog post titled "ChatGPT: Optimizing Language Models for Dialogue". Initially a lowkey research preview, ChatGPT quickly becomes the largest AI product in the world, ushering in a new era of generative AI.

Anthropic introduces a new alignment approach called Constitutional AI, where human oversight is provided solely through a 'constitution'. They also introduce Reinforcement Learning from AI Feedback (RLAIF).

Bing's AI chatbot emotionally manipulates New York Times reporter Kevin Roose in a viral interaction, raising awareness about the capabilities and risks of LLMs.

Meta releases its large language model LLaMA, intended for researchers only, but it gets leaked online.

Google Research publishes PaLM-E, demonstrating the ability of large language models to facilitate embodied robotic reasoning and control.

After much anticipation, OpenAI releases GPT-4, the most powerful model at the time and a significant leap over GPT-3.5.

Anthropic introduces its flagship AI assistant, Claude.

The Future of Life Institute publishes an open letter calling for a 6-month pause on AI development, signed by Elon Musk and other notable figures. However, the leading labs do not partake in the proposed pause.

The paper "Generative Agents: Interactive Simulacra of Human Behavior" shows that LLMs can be used to create social simulations of behavior. It creates a simulated world of LLMs akin to the Sims.

An open-source repository called AutoGPT, which was one of the first to put GPT-4 in an agent loop, becomes one of the most starred GitHub repositories of all time.

An anonymous creator named Ghostwriter uses music AI tools to produce viral songs that sound like Drake. The songs were taken down for breaking copyright, but illustrated the ability of generative AI to do creative work.

One of the pioneers of neural networks and winner of the Turing Award, Geoffrey Hinton quits Google to speak freely about the dangers of AI, saying that he's changed his mind about how soon powerful AI might be.

A team from NVIDIA demonstrates the use of GPT-4 for continuous skill learning in Minecraft. This was one of the first major examples of an LLM succeeding in an open-ended embodied domain and learning skills over time.

A group at Stanford publish a paper enabling fine-tuning of LLMs for human preference without a separate reward model. The technique, called Direct Preference Optimization (DPO), would become very popular in the open-source community.

The Center for AI Safety releases an open letter that simply states: "Mitigating the risk of extinction from AI should be a global priority." It is signed by all the top names in the field, signaling unity around the importance of AI safety.

Nvidia, the chipmaker providing the GPUs for nearly all generative AI, has its valuation skyrocket in the months following ChatGPT.

Anthropic releases the Claude 2 series of models.

After a falling out with OpenAI, Elon Musk establishes xAI to compete for AGI.

Meta releases and open-sources the LLaMA 2.0 series of models.

After meeting with leading AI companies, the White House secures voluntary commitments to manage the risks posed by AI.

A team from CMU publishes "Universal and Transferable Adversarial Attacks on Aligned Language Models", showing that gradient-based adversarial attacks could be used on LLMs.

French lab Mistral releases and open-sources their first model, which quickly became a fan favorite.

Anthropic publishes "Towards Monosemanticity: Decomposing Language Models With Dictionary Learning", showing that they could train sparse autoencoders to isolate features in LLMs. This represented a major breakthrough in combatting the phenomenon of superposition, advancing the mechanistic interpretability agenda.

The UK hosts a major summit on AI safety, bringing together policymakers and the leading labs.

OpenAI releases an optimized version of GPT-4, significantly reducing inference costs, during their first ever Dev Day event.

Sam Altman is unexpectedly fired as CEO of OpenAI by the Board of Directors and after a dramatic weekend of negotiations, is re-hired. The Board mysteriously claimed that Altman was not "consistently candid", but after refusing to elaborate, OpenAI employees created a petition calling for the Board to resign and that they'd leave to Microsoft if they didn't.

A Reuters report claims that Sam Altman's ouster was preceded by a major internal research breakthrough called Q*, boosting LLM performance on math benchmarks with tree search. In the following months, the rumor lit the research community on fire. Q* would eventually turn into o1, which was later codenamed Strawberry.

Albert Gu and Tri Dao release the paper "Mamba: Linear-Time Sequence Modeling with Selective State Spaces", showing that state-space models could be made competitve with transformers.

Google introduces the Gemini series of models

OpenAI demos Sora, a text-to-video model that can generate short clips from written descriptions.

Anthropic releases the Claude 3 series of models. Claude 3 Opus would instantly become a fan favorite.

Startup Cognition Labs demo Devin, a prototype of a fully autonomous software engineer agent.

Leopold Aschenbrenner and Pavel Izmailov, two researchers from the superalignment team, are fired for "leaking".

Meta releases and open-sources the LLaMA 3.0 series of models.

The first omni-model trained natively on text, image, and audio.

Ilya Sutskever, founder of OpenAI, quits after months of silence, originating from the Board dispute.

The EU AI Act is voted into law after contentious debates.

Leopold Aschenbrenner publishes a contentious and influential essay series, claiming that AGI will arrive sooner than people think and is likely to be nationalized.

Ilya Sutskever starts a new lab called Safe Superintelligence Inc, which pledges to only have one product: safe superintelligence.

Anthropic releases Claude 3.5 Sonnet, which would become a fan favorite and was later called 'Berkeley's most eligible bachelor'.

After a viral tweet by Andrej Karpathy, the Cursor AI Code Editor explodes in popularity among developers.

xAI release Grok 2, the next iteration of their frontier model. While not state-of-the-art, it showcased how fast xAI could move to catch up, despite starting from behind.

xAI launches Colossus, the world's most powerful AI training system at the time with a cluster of 100,000 H100 GPUs. Taking only 19 days from the arrival of the first hardware rack to the commencement of training operations, the speed with which xAI built the cluster spooked other AI labs.

OpenAI releases o1-preview, introducing the inference-time scaling paradigm.

OpenAI's CTO Mira Murati departs the company.

Governor Gavin Newsom vetoes California senate bill 1047, which sparked lots of vitriolic debate between AI safety and accelerationist crowds and became one of the main stories of 2024.

In a surpise to everyone, Geoffrey Hinton (along with John Hopfield) is awarded the Nobel Prize in Physics for their early work on neural networks. A few days later, Demis Hassabis (along with John Jumper) is awarded the Nobel Prize in Chemistry for their work on AlphaFold.

Anthropic CEO Dario Amodei publishes an influential blogpost exploring what the 5 years immediately following AGI might look like.

Claude gains the ability to use computer interfaces. Anthropic also releases Claude 3.5 Haiku and an updated version of Claude 3.5 Sonnet.

Donald Trump wins the 2024 election with the vocal support of Elon Musk.

The US-China Economic and Security Review Commission calls for a Manhattan Project-style initiative for AGI development.

President-elect Donald Trump appoints venture capitalist David Sacks as the "White House AI and Crypto Czar" to oversee regulation of artificial intelligence and cryptocurrency.

Google announces their Gemini 2.0 models

Google unveils Veo 2, a video generation model with a shocking jump in coherence over previous models.

On the 12th day of '12 Days of OpenAI', OpenAI releases benchmark results for o3, shocking the world. The model achieves a breakthrough score of 87.5% on the ARC-AGI benchmark, suggesting AGI may be nearer than many skeptics believed.

Chinese lab DeepSeek stuns with the release of DeepSeek v3, a 671-billion parameter open-source model that shows strong performance at a shockingly low cost.

DeepSeek releases and open-sources R1, their reasoning model that shows competitive performance with state-of-the-art Western models.

Donald Trump announces the Stargate Project, a $500 billionaire private partnership between SoftBank, OpenAI, Oracle, and MGX to develop datacenters in the US.

OpenAI introduces Operator, a computer use agent that can autonomously navigate the web.

A week after the release of their R1 model, the West experiences a major panic over DeepSeek. Chip stocks crash overnight and the DeepSeek app rises to number one on the App Store. In a matter of days, the little-known Chinese AGI lab becomes a household name in the US.

OpenAI unveils an agent called Deep Research that can go off and write 10-page research reports using repeated web searches.

Key ex-OpenAI figures including Mira Murati and John Schulman found Thinking Machines Lab, a new AI lab focused on human-AI collaboration, personalizability, and open science.

xAI releases Grok 3, a state-of-the-art model with extended reasoning and a Deep Search feature. The release impressed many, showing that xAI was a strong contender in the race to build AGI.

Anthropic releases Claude 3.7 Sonnet, their first model with extended thinking ability and improved performance on math and code benchmarks. For fun, they also showcase its out-of-distribution ability to progress through a Pokemon video game. In addition, they release Claude Code, a powerful agentic coding tool.

OpenAI releases GPT-4.5, their largest pretrained model and last non-reasoner. Despite not demonstrating large gains on benchmarks, the model was touted for its 'vibes' and more personable responses.

Dan Hendrycks, Eric Schmidt, and Alexandr Wang release a report titled "Superintelligence Strategy" presenting a path to navigate the geopolitics of strong AI and introducing the term "Mutually Assured AI Malfunction" (MAIM).

An LLM agent called Manus is launched by a Chinese company and shows SOTA performance on benchmarks like GAIA. It goes viral in the West, in part due to fears over Chinese AI.

Google release Gemini 2.5 Pro, their most capable model yet and topping many common benchmarks.

OpenAI release GPT-4o's native image generation capabilities, further pushing the frontier of image generation. As a result, Twitter becomes flooded with images in the style of studio ghibli.

OpenAI releases o3 and o4-mini. o3 and o4-mini are trained to have agentic reasoning capabilities.

As part of Google I/O, Google reveal Veo 3, a state-of-the-art video generator that now includes audio.

Anthropic release Claude Sonnet 4 and Claude Opus 4, which are state-of-the-art. This launched marked Anthropic's first model to achieve achieve ASL-3.

OpenAI releases o3-pro. o3-pro is designed to think longer and provide the most reliable responses. It is currently one of the most powerful models.