Timeline
Major AI events from the present back to the field's origins. Each event links to sources — many link to GitHub issues for active debate.
📂 Data lives in timeline.json — missing an event? Open an issue or submit a PR.
2026
Vibe Coding Becomes Leisure Activity
IndustryA cultural shift: people start replacing video games and bar time with 'vibe coding' — building things with AI for fun. Programming as entertainment, not just work. The barrier to creation approaches zero.
Energy Emerges as the Real Bottleneck
InfrastructurePer Elon Musk on Collision/Dwarkesh: energy is the binding constraint for AI scaling in the next year, with chips becoming the bottleneck in 3-4 years. Hyperscalers continue massive capex into infrastructure.
Coding Agents Write Most New Code
IndustryAI coding agents are now writing the majority of new production code. The question shifts from 'will AI replace programmers?' to 'how fast?' Context windows at 200K-1M standard.
Claude 4 / Opus 4 Released
Model ReleaseAnthropic releases Claude 4 with Opus 4 as the new frontier reasoning model. Extended thinking, deeper code understanding, and improved reliability. Continues to dominate coding workflows.
GPT-5 Series Launches
Model ReleaseOpenAI releases the GPT-5 family with significant improvements in reasoning, instruction following, and multimodal capabilities. The race at the frontier remains tight.
Grok 4 — The RLVR Milestone
Model ReleasexAI releases Grok 4. First major model where more than half of training cost was NOT pre-training. Reinforcement learning from verifiable rewards (RLVR) and synthetic data cross a critical threshold.
2025
Frontier Model Costs Drop 10x from 2024
InfrastructureCost per token for frontier-quality inference drops roughly 10x year-over-year. What cost $100 in early 2024 costs $10 by end of 2025. The deflationary spiral continues.
RAG Begins Looking Obsolete
InfrastructureWith 1M+ context windows and agents that can manage their own filesystem and memory, traditional RAG architectures feel increasingly unnecessary. Agents decide how to access data themselves.
Agents Go from Minutes to Hours
InfrastructureAI agents start running autonomously for hours, not just minutes. Still unreliable compared to humans, but 1000x cheaper. The math changes the game — reliability matters less when cost approaches zero.
Grok 3 Released
Model ReleasexAI releases Grok 3 with massive compute scaling. Trained on xAI's Memphis supercluster with 100K H100s. Competitive with frontier models across benchmarks.
Gemini 2.5 with Native Tool Use
Model ReleaseGoogle releases Gemini 2.5 Pro with native tool use and deep research capabilities. Agents can browse, code, and reason in extended chains. Google's infrastructure advantage starts showing.
Gemini 3 Pro Released
Model ReleaseGoogle launches Gemini 3 Pro with expanded context, improved reasoning, and tighter integration with Google's ecosystem. Native image generation and tool use.
OpenAI Raises at $300B Valuation
FundingOpenAI closes funding at a $300B valuation, the highest for a private company in history. But Nvidia pulling back from a rumored $100B investment suggests cracks.
Anthropic Raises $3.5B at $60B Valuation
FundingAnthropic raises $3.5B at a $60B valuation, up from $18B a year earlier. Amazon and Google compete to be primary cloud partner. Safety-first approach is commercially viable.
Claude 3.5 Sonnet Becomes Coding Standard
IndustryClaude 3.5 Sonnet becomes the de facto model for AI-assisted coding. Developers prefer it over GPT-4 for code tasks. Anthropic's quiet dominance in the developer market.
DeepSeek R1 Shocks the World
Open SourceChinese lab DeepSeek releases R1, an open-source reasoning model that rivals frontier labs at a fraction of the cost. Proves reasoning can be trained cheaply. Stock markets react.
💬 Debate this event (#16)2024
Context Windows: 200K+ Becomes Commodity, 1M+ at Frontier
InfrastructureBy end of 2024, 200K+ token context windows are commodity across frontier models, with Gemini offering 1M-2M. GPT-4 Turbo hit 128K in Nov 2023, Claude 3 launched at 200K in March 2024. The frontier pushes toward practical 1M+ usage.
Mistral Raises at $6B Valuation
FundingMistral AI raises $640M at a $6B valuation, barely a year after founding. Europe's AI champion. Open-weight models (Mistral 7B, Mixtral) prove you don't need Bay Area scale.
AI Coding Tools Explode
IndustryCursor, GitHub Copilot, and AI coding assistants cross critical adoption thresholds. Developers start reporting 2-5x productivity gains. Coding becomes AI's killer app.
GPT-4o — Multimodal Native
Model ReleaseOpenAI releases GPT-4o with native voice, vision, and text in one model. Real-time voice conversation feels like science fiction. Context windows hit 128K standard.
Llama 3 Released
Open SourceMeta releases Llama 3 (8B and 70B), setting new open-source benchmarks. The 70B model rivals GPT-3.5. Open source is no longer a generation behind.
EU AI Act Passed — World's First Comprehensive AI Regulation
PolicyThe European Parliament passes the EU AI Act, the most comprehensive AI regulation in the world. Classifies AI systems by risk level, bans certain uses (social scoring, real-time biometric surveillance), and requires transparency for foundation models. Sets the global regulatory template.
💬 Debate this event (#42)Claude 3 Family Released
Model ReleaseAnthropic launches Claude 3 Opus, Sonnet, and Haiku. Opus matches or beats GPT-4 on most benchmarks. Anthropic's breakout moment — proves the frontier isn't OpenAI-only.
OpenAI Announces Sora
Model ReleaseOpenAI reveals Sora, generating photorealistic video from text prompts. Hollywood takes notice. The gap between 'AI can't do X' and 'AI does X well' keeps shrinking.
Gemini 1.5 Pro — 1M Context Window
Model ReleaseGoogle announces Gemini 1.5 Pro with a 1 million token context window. Shatters previous limits by 10x. At Google I/O in May, expanded to 2M tokens. Raises questions about whether RAG is even necessary anymore.
2023
OpenAI Board Crisis
IndustryOpenAI's board fires Sam Altman, triggering a 5-day corporate drama that ends with his reinstatement. Exposes tensions between safety and commercialization that define the industry.
Elon Musk Founds xAI, Launches Grok
Lab FormationMusk's xAI launches Grok, trained partly on X/Twitter data with a 'rebellious' personality. Adds a fourth major player to the frontier lab race.
UK AI Safety Summit at Bletchley Park
Policy28 countries sign the Bletchley Declaration on AI safety. First major international AI governance gathering. Hosted at the historic codebreaking site. AI safety moves from academic concern to geopolitical agenda.
Biden AI Executive Order
PolicyPresident Biden signs sweeping executive order on AI safety, requiring disclosure of training details for powerful models. First major US government AI regulation attempt.
Llama 2 Released Open Source
Open SourceMeta releases Llama 2 with a commercial license, legitimizing open-source as a viable path for frontier-adjacent models. Fine-tuning community explodes.
Inflection AI Raises $1.3B
FundingInflection AI (founded by Mustafa Suleyman) raises $1.3B from Microsoft, Nvidia, and others. Peak AI funding mania — the company would later be effectively absorbed by Microsoft.
Mistral AI Founded
Lab FormationArthur Mensch, Timothée Lacroix, and Guillaume Lample (ex-Meta/DeepMind) found Mistral AI in Paris. Europe's answer to the American AI lab duopoly. Raises €105M seed — a European record.
'Pause Giant AI Experiments' Open Letter
PolicyOver 1,000 signatories including Elon Musk and Steve Wozniak sign an open letter calling for a 6-month pause on training AI systems more powerful than GPT-4. Sparks fierce debate about AI safety vs. progress.
GPT-4 Released
Model ReleaseOpenAI launches GPT-4, a massive leap in reasoning and multimodal capability. Passes the bar exam, scores 90th percentile on SATs. Sets a new frontier that competitors race to match.
Anthropic Launches Claude
Model ReleaseAnthropic releases Claude, built on Constitutional AI principles. Positions itself as the 'safety-first' alternative to GPT-4. Begins the multi-lab frontier race.
Stanford Alpaca — Fine-Tuning for $600 Ignites Open-Source Explosion
Open SourceStanford releases Alpaca, a LLaMA fine-tune created for under $600. Proves frontier-adjacent capabilities are accessible to individual researchers. Spawns Vicuna, Koala, GPT4All, and dozens of derivatives within weeks. The open-source Cambrian explosion begins.
💬 Debate this event (#44)Meta Releases LLaMA
Open SourceMeta releases LLaMA to researchers. It leaks within a week, kickstarting the open-source LLM movement and democratizing access to powerful models.
Google Launches Bard
Model ReleaseGoogle rushes out Bard as a ChatGPT competitor. A factual error in the launch demo wipes $100B off Alphabet's market cap. The AI race officially begins.
Microsoft Invests $10B in OpenAI
FundingMicrosoft commits $10 billion to OpenAI in a multi-year deal, signaling that Big Tech sees generative AI as the next platform war. Azure becomes OpenAI's exclusive cloud partner.
2022
ChatGPT Launches
Model ReleaseOpenAI releases ChatGPT, reaching 100M users in two months. The moment AI went from research curiosity to mainstream phenomenon. Everything before this was prologue.
💬 Debate this event (#13)Stable Diffusion Released — Open Source Image Generation
Open SourceStability AI open-sources Stable Diffusion, making high-quality image generation freely available. Triggers explosion of AI art tools and debate about creative AI. Open-source AI becomes a movement.
DALL-E 2 — Photorealistic AI Art
Model ReleaseOpenAI releases DALL-E 2 with dramatically improved image quality. Photorealistic generations stun the public. AI art debate explodes. Artists begin worrying about displacement.
Chinchilla Scaling Laws Reshape Model Training
ResearchDeepMind's Chinchilla paper proves GPT-3 and similar models were vastly under-trained on data relative to their size. Compute-optimal training requires far more data than assumed. Directly inspires LLaMA and the open-source model explosion.
💬 Debate this event (#34)InstructGPT Establishes RLHF as the Alignment Standard
ResearchOpenAI publishes InstructGPT, showing that reinforcement learning from human feedback (RLHF) dramatically improves model helpfulness and safety. This technique directly enables ChatGPT later that year and becomes the standard alignment approach across all frontier labs.
💬 Debate this event (#38)2021
GitHub Copilot Technical Preview
IndustryGitHub launches Copilot, powered by OpenAI Codex. AI-assisted coding enters developer workflows for the first time at scale. The seed of the AI coding revolution.
DALL-E — Text-to-Image Generation
Model ReleaseOpenAI reveals DALL-E, generating images from text descriptions. AI creativity enters mainstream discourse. The gap between 'AI can't be creative' and 'AI generates art' closes overnight.
2020
AlphaFold 2 Solves Protein Folding at CASP14
ResearchDeepMind's AlphaFold 2 effectively solves the 50-year protein folding grand challenge, achieving accuracy comparable to experimental methods. Revolutionizes structural biology. Later shares the 2024 Nobel Prize in Chemistry.
💬 Debate this event (#33)GPT-3 — The Scaling Leap
Model ReleaseOpenAI releases GPT-3 with 175B parameters — 100x GPT-2. Few-shot learning actually works. Developers build products on the API. The age of 'prompt engineering' begins. First real evidence that scale = capability.
2019
GPT-2 — 'Too Dangerous to Release'
Model ReleaseOpenAI releases GPT-2 (1.5B parameters) but initially withholds the full model, citing concerns about misuse. Generates coherent paragraphs of text. The 'too dangerous to release' framing becomes a meme — and a prescient debate.
2018
Google Releases BERT — Bidirectional Transformers Dominate NLP
ResearchGoogle publishes BERT, proving bidirectional pre-training crushes NLP benchmarks. Becomes the backbone of Google Search and enterprise NLP for years. While GPT bet on generation, BERT bet on understanding — and won the near-term.
💬 Debate this event (#32)GPT-1 Released — Language Modeling at Scale
Model ReleaseOpenAI releases the first GPT (Generative Pre-trained Transformer). 117M parameters. Modest by later standards, but proves that unsupervised pre-training + fine-tuning works. The GPT lineage begins.
2017
'Attention Is All You Need' — The Transformer Paper
ResearchGoogle researchers publish the Transformer architecture. Self-attention replaces recurrence. This single paper enables GPT, BERT, Gemini, Claude, and every modern LLM. Arguably the most consequential ML paper ever written.
2016
AlphaGo Defeats Lee Sedol
Model ReleaseDeepMind's AlphaGo beats world Go champion Lee Sedol 4-1. Go was considered decades away from AI mastery. Move 37 in Game 2 — a move no human would play — becomes iconic. AI capability timelines compress.
2014
Generative Adversarial Networks (GANs) Published
ResearchIan Goodfellow introduces GANs — two neural networks competing to generate increasingly realistic data. Opens the door to AI-generated images, video, and audio. The generative AI seed is planted.
2013
Word2Vec Introduces Dense Word Embeddings
ResearchTomas Mikolov et al. at Google publish Word2Vec, shifting NLP from sparse representations to dense vector embeddings. 'King - Man + Woman = Queen' captures the public imagination. The critical precursor to neural NLP and eventually Transformers.
💬 Debate this event (#34)2012
AlexNet Wins ImageNet — Deep Learning Revolution Begins
ResearchAlex Krizhevsky's deep neural network crushes the ImageNet competition by a massive margin. The deep learning era begins. GPUs become essential AI hardware. This is the real inflection point for modern AI.
2011
IBM Watson Wins Jeopardy!
Model ReleaseWatson defeats champions Ken Jennings and Brad Rutter on Jeopardy! Natural language understanding reaches mainstream awareness. IBM pivots to enterprise AI — with mixed results.
1997
Deep Blue Beats Kasparov
Model ReleaseIBM's Deep Blue defeats world chess champion Garry Kasparov. First time a machine beats a reigning world champion in a full match. Massive cultural moment — but it's brute force, not intelligence.
1987
Second AI Winter
IndustryExpert systems prove brittle and expensive to maintain. The AI market collapses. Lisp machine companies go bankrupt. Funding disappears again. AI researchers avoid using the term 'AI' for years.
1980
Expert Systems Boom
IndustryRule-based expert systems like MYCIN and XCON generate billions in commercial value. AI is hot again. Japan launches the Fifth Generation Computer Project. The second AI boom — which will also crash.
1973
First AI Winter Begins
IndustryThe Lighthill Report (UK, 1973) and subsequent DARPA funding cuts trigger the first AI winter. Earlier, the ALPAC report (1966) had killed machine translation funding specifically. Early optimism crashes against the reality that AI problems are harder than expected. Funding dries up for a decade.
1966
ELIZA — First Chatbot
ResearchJoseph Weizenbaum creates ELIZA at MIT, simulating a Rogerian psychotherapist. Users form emotional attachments to it despite its simplicity. First demonstration of the 'ELIZA effect' — humans anthropomorphizing machines.
1956
Dartmouth Workshop — 'Artificial Intelligence' Coined
ResearchJohn McCarthy, Marvin Minsky, Claude Shannon, and others convene at Dartmouth College. The term 'Artificial Intelligence' is born. The field officially begins with boundless optimism.
1950
Turing Publishes 'Computing Machinery and Intelligence'
ResearchAlan Turing asks 'Can machines think?' and proposes the Turing Test. The foundational paper that frames the entire field of AI. Every debate since is downstream of this question.
This page is open source
Missing an event? Think a date or description is wrong? Open an issue or submit a PR. All debates happen on GitHub. What gets merged becomes the record.