Category
88 entries
OpenAI releases GPT-5.5, GPT-5.5 Thinking, and GPT-5.5 Pro: designed as an "agent runtime" for persistent multi-step workflows. 23% more factually correct vs GPT-5.4. File Library, side-by-side shopping, improved image gen.
Google releases Gemini 3.1 Pro: 77.1% on ARC-AGI-2 (more than double Gemini 3 Pro), 80.6% SWE-Bench Verified, 94.3% GPQA Diamond. Same price as 3 Pro: $2/M input.
Anthropic releases Opus 4.6: first Opus with 1M-token context in beta, agent teams in Claude Code, leadership on Terminal-Bench 2.0 and Humanity's Last Exam. Pricing unchanged at $5/$25.
Google DeepMind announces Gemini 3 with Pro and Flash variants: improved reasoning, native long context, deeper integration into Workspace and Android.
Anthropic releases Claude Haiku 4.5: performance equal to Claude Sonnet 4 (May 2025) at a third of the price and double the speed. Changes the cost/quality ratio for high-volume agentic tasks.
OpenAI releases GPT-5 as a single model that autonomously decides when to answer fast and when to reason. Family: GPT-5, mini, nano, Pro. Default in ChatGPT, including free tier.
xAI launches Grok 4 and Grok 4 Heavy (variant running multiple parallel instances, like o1-pro). SuperGrok Heavy tier at $300/month. High but contested benchmark numbers.
Anthropic launches Claude Opus 4 and Sonnet 4. Opus 4 reaches 72.5% on SWE-bench Verified (vs 49% for Sonnet 3.7), can work autonomously on coding tasks for hours. 'Extended thinking' built in.
Mistral launches Medium 3, claimed 8× cheaper than Claude Sonnet at similar performance and deployable self-hosted on 4 GPUs. Positioned on the European 'sovereign enterprise' niche.
OpenAI ships o3 (full) and o4-mini as reasoning models with native access to all ChatGPT tools: web search, Python, image gen, vision. First real 'agentic reasoning'.
Google DeepMind ships Gemini 2.5 Pro, first model in the 2.5 family with built-in 'thinking'. 1M context window, reasoning capabilities competitive with o1/o3.
OpenAI releases GPT-4.5 (codename Orion) as a 'research preview'. The largest model the company ever trained with traditional scaling, but expensive — marking the end of the pure pre-training era.
xAI launches Grok 3, trained on the Colossus 200K H100 cluster in Memphis. Includes a 'Think' reasoning mode and 'DeepSearch' agentic web research. Available to X Premium subscribers.
Google makes Gemini 2.0 Flash generally available, introduces cheaper Flash-Lite, and previews Gemini 2.0 Pro Experimental with a 2M-token context window.
Stanford/UW paper: with 1000 curated examples and a technique called 'budget forcing' they fine-tune Qwen2.5-32B to compete with o1-preview on math. Training cost: <$50.
Microsoft Research publishes MatterGen in Nature: a diffusion model generating stable crystal structures conditioned on target properties (magnetism, conductivity). Experimental synthesis of a new material confirmed.
Moonshot AI releases Kimi k1.5, a reasoning model with 128k context and RL-trained long chain-of-thought that matches OpenAI o1 on AIME and MATH-500, with a user-controllable 'long-thinking' mode.
OpenAI announces o3 and o3-mini: SWE-bench 71.7%, FrontierMath 25.2%, ARC-AGI 87.5% (with high compute budget). Huge jump on hard reasoning. GA expected in 2025.
Google DeepMind deploys Gemini Nano (1.8B and 3.25B) on Pixel 8 Pro and Galaxy S25, offline execution on NPU via Android AICore API. First time a frontier lab puts an LLM directly on the device.
AllenAI releases OLMo 2 at 7B and 13B with staged mid-training and specialized data mixing, outperforming Llama 3.1 and Qwen 2.5 on instruction following while preserving full transparency on data, code, and checkpoints.
The Swedish Academy awards the 2024 Chemistry Nobel to David Baker (protein design) and to Demis Hassabis and John Jumper at DeepMind for AlphaFold — the first time an industry AI system co-stars in a scientific Nobel.
The Royal Swedish Academy awards the 2024 Physics Nobel to John Hopfield and Geoffrey Hinton for their foundational work on artificial neural networks, formally recognizing AI as a discipline.
NVIDIA releases Nemotron-4 340B optimized for high-quality synthetic data generation, enabling enterprises to train smaller domain-specific models without collecting real data.
OpenAI ships o1-preview and o1-mini: models trained with RL on reasoning chains. On math, physics, competitive coding they beat GPT-4o by a huge margin. Paradigm shift.
DeepMind announces that AlphaProof (on Lean) and AlphaGeometry 2 solved 4 of 6 problems at the 2024 International Mathematical Olympiad, reaching silver-medal threshold.
OpenAI ships GPT-4o mini at $0.15/$0.60 per 1M tokens, 60% cheaper than GPT-3.5 Turbo, MMLU 82%. Moves the 'baseline model' bar for most use cases.
Mistral AI and NVIDIA release Mistral Nemo 12B: 128k context window, trained with NeMo toolkit, designed as a direct replacement for Mistral 7B in production.
Anthropic releases Claude 3.5 Sonnet: outperforms Claude 3 Opus (the previous flagship) at Sonnet pricing ($3/$15). Introduces 'Artifacts': side-panel output for code, documents, charts.
DeepMind and Isomorphic Labs publish AlphaFold 3 in Nature: it predicts the structure and interactions of proteins, DNA, RNA, ligands, and ions — vastly extending the domain beyond AlphaFold 2.
Snowflake releases Arctic, a MoE with 480B total and 17B active parameters per token, SOTA on enterprise SQL and coding, Apache 2.0, trained with 3.5M GPU-hours on H100.
Anthropic ships the Claude 3 family in three sizes. Opus, the flagship, beats GPT-4 on MMLU, HumanEval, MATH. Native multimodal vision. For the first time GPT-4 is no longer the outright leader.
Mistral AI announces Mistral Large, a closed flagship model with near-GPT-4 performance, and Le Chat (consumer interface). In parallel it signs a strategic Microsoft partnership for Azure distribution.
Google announces Gemini 1.5 Pro: Mixture of Experts architecture, 128K standard context, 1M in preview. New benchmark: near-perfect 'needle in a haystack' retrieval over long inputs.
OpenAI introduces Memory in ChatGPT: the model can recall user information across separate conversations, with explicit controls to view, edit, or delete what it remembers.
Alibaba Cloud releases Qwen-1.5, a 0.5B-to-110B parameter family with native 32k context support, GQA, bilingual EN/ZH, instructions in 30+ languages, and RLHF chat.
AllenAI releases OLMo with weights, the full Dolma dataset (3T tokens), training code, and all intermediate checkpoints, making the entire LLM training process scientifically reproducible for the first time.
DeepMind publishes AlphaGeometry in Nature, a neuro-symbolic system that solves International Mathematical Olympiad geometry problems at medal level, without human-annotated training data.
Google announces Gemini Ultra/Pro/Nano, the first family of natively multimodal models (text, images, audio, video). Ultra beats GPT-4 on MMLU 90.0% vs 86.4%. Controversial demo video.
01.ai by Kai-Fu Lee releases Yi-34B: 34B parameters trained on 3.1T tokens, modified Llama-2 architecture, bilingual EN/ZH, top-3 open weight in November 2023.
Anthropic ships Claude 2.1: 200K-token context window (~500 pages), 2× reduction in false statements on borderline questions, tool use in beta. Reply to GPT-4 Turbo 128K.
Upstage presents Solar 10.7B, created by merging intermediate layers of two fine-tuned LLaMA-2 models (depth upscaling), winning the MBTI-OpenLLM leaderboard in November 2023.
At OpenAI's first developer conference: GPT-4 Turbo (128K context, lower prices), GPTs (shareable custom ChatGPTs), Assistants API (managed agents). Product + dev pivot.
Elon Musk's xAI launches Grok-1, a model integrated with X (Twitter) for real-time information, with a 314B MoE architecture released as open weights in March 2024.
HuggingFace trains Zephyr-7B with dSFT + Direct Preference Optimization on Mistral 7B base, achieving an MT-Bench score higher than Llama-2-70B-chat with 10x fewer parameters.
The Technology Innovation Institute releases Falcon-180B, the largest openly available model at 180 billion parameters trained on 3.5 trillion tokens, topping the HuggingFace Open LLM Leaderboard.
Microsoft Research shows that 1.3B parameters trained on 'textbook quality' synthetic data produce multi-step reasoning comparable to models five times larger.
Microsoft Research trains Orca 13B on step-by-step GPT-4 explanations (explanation traces), outperforming ChatGPT on BigBench and AGIEval with 13 billion parameters.
Anthropic launches Claude 2 with a 100,000-token context window (~75,000 words) and opens claude.ai to the general public (initially US and UK). Long-context enters the mainstream.
WizardLM uses Evol-Instruct — instructions automatically simplified and complicated by GPT-4 — achieving 97% of ChatGPT on WizardEval with a 70B model.
Microsoft Research releases Phi-1, 1.3B parameters trained on high-quality synthetic data ('textbooks'), outperforming models 10x larger on HumanEval.
The Technology Innovation Institute UAE releases Falcon 40B: trained on 1T tokens of RefinedWeb, it beats LLaMA 65B on benchmarks with a commercial license.
At Google I/O 2023, PaLM 2 replaces LaMDA in Bard. Four sizes (Gecko, Otter, Bison, Unicorn), strong multilingual support and improved reasoning. Spawns Med-PaLM 2 and Sec-PaLM.
MosaicML launches MPT-7B under Apache 2.0 with a 65,000-token context window via ALiBi, the first open model explicitly designed for unrestricted commercial deployment.
LMSYS fine-tunes LLaMA-13B on 70,000 ShareGPT conversations and produces an open-source chatbot that GPT-4, used as judge, rates at 90% of ChatGPT quality.
Google opens Bard public preview in US and UK, based on a lightweight LaMDA. Reception is lukewarm: slow, cautious, less useful than ChatGPT.
Anthropic launches Claude, an AI assistant trained with Constitutional AI. Same day as GPT-4. Two versions: Claude (full) and Claude Instant (faster and cheaper).
OpenAI releases GPT-4, multimodal (text + image), with reasoning, coding, and reliability clearly beyond GPT-3.5. Passes bar, medical, and coding exams.
OpenAI ships the ChatGPT API (gpt-3.5-turbo) at one tenth the price of text-davinci-003, plus Whisper API for speech-to-text. The wrapper era begins.
Microsoft integrates conversational AI into Bing (later revealed to run on pre-release GPT-4) that answers with direct citations from web pages. The Google 'code red' moment.
OpenAI launches ChatGPT, a free conversational interface on GPT-3.5 aligned via RLHF. It crosses one million users in five days.
Meta unveils Galactica, a 120B-parameter model trained on 48 million scientific papers. The public demo is pulled after three days under a wave of criticism for authoritative hallucinations.
Google scales instruction tuning to 1,800 tasks and 540B parameters, open-sources Flan-T5, and proves that chain-of-thought reasoning is teachable via fine-tuning.
Noam Shazeer and Daniel De Freitas, fathers of LaMDA, launch Character.AI: a platform letting anyone create and chat with AI characters, from Einstein to anime personas.
Google Research combines three major pretraining objectives into a single 20B model, outperforming GPT-3 on many benchmarks at one-eighth the parameters.
Google publishes PaLM, a 540B-parameter model trained on the new Pathways system. Demonstrates emergent reasoning capabilities when guided with chain-of-thought.
DeepMind publishes the Chinchilla paper and shows that, given equal compute, smaller models trained on far more tokens beat oversized undertrained ones.
Wang et al. (Google Brain) show that sampling N diverse reasoning paths and taking the most frequent answer beats greedy decoding on all reasoning benchmarks.
OpenAI introduces InstructGPT: a GPT-3 refined with human feedback (RLHF) that follows instructions better than the 175B base model despite being much smaller (1.3B parameters).
DeepMind releases Gopher, a 280B dense model, alongside a systematic 152-task study and a companion paper on ethical considerations of foundation models.
DeepMind publishes RETRO, a 7B-parameter model that retrieves relevant passages from a 2T-token database at inference, matching the performance of models 25x larger.
Google shows that training a model on 60+ tasks framed as instructions dramatically improves zero-shot performance on unseen tasks.
Microsoft and NVIDIA announce MT-NLG, a 530B-parameter dense model trained with DeepSpeed and Megatron-LM, at the time the largest dense LM ever produced.
Stanford's Center for Research on Foundation Models publishes a 200+ page report coining the term foundation models, now standard in technical, academic and regulatory discourse.
BAAI (Beijing Academy of Artificial Intelligence) introduces Wu Dao 2.0, a 1.75 trillion-parameter multimodal Mixture of Experts model — China's response to GPT-3 and Switch Transformer.
At Google I/O, Sundar Pichai introduces LaMDA (Language Model for Dialogue Applications), a 137B-parameter model fine-tuned for dialogue, direct ancestor of Bard.
Google Brain publishes Switch Transformer, a sparse model with 1.6 trillion parameters that activates only one expert per token, proving sparse routing can scale beyond dense models.
DeepMind publishes MuZero in Nature: the RL agent learns world dynamics on its own and reaches superhuman performance on Go, chess, shogi, and 57 Atari games without being given the rules.
Google Research presents Big Bird at NeurIPS 2020, a transformer with sparse attention (local + global + random) that scales linearly, reaches SOTA on long-document QA and summarization, and proves Turing-completeness.
DeepMind announces that AlphaFold 2 has won the CASP14 competition with mean GDT >90, on par with experimental methods — widely regarded as solving the 50-year-old protein folding problem.
Google Research publishes mT5, a T5 variant pre-trained on mC4 (multilingual Common Crawl) over 101 languages, which becomes a standard baseline for many cross-lingual NLP tasks.
Allen Institute for AI releases Longformer, a transformer that combines local sliding-window attention with global attention on special tokens, scaling linearly up to 4096 tokens and beating RoBERTa on long-document tasks.
Two weeks after the paper, OpenAI opens a private beta of the first general API for its language models, available to a few hundred developers building applications directly on top of GPT-3.
OpenAI publishes 'Language Models are Few-Shot Learners' and shows that at 175B parameters a model learns new tasks from a handful of examples in the prompt.
Lewis et al. at Facebook AI publish the RAG paper, combining a dense retriever (DPR) with a seq2seq generator (BART) to answer knowledge-intensive questions without baking all facts into the weights.
Clark, Luong, Le, and Manning publish ELECTRA at ICLR 2020: instead of masked language modeling, it trains the model to detect tokens replaced by a small generator, matching BERT with a quarter of the compute.
Microsoft Research unveils Turing-NLG, the largest announced language model to date (17B), made possible by the DeepSpeed/ZeRO optimizer that drastically cuts GPU memory.
Google introduces Meena, a 2.6B-parameter conversational model trained on 341 GB of social dialogue, along with SSA, a new metric for evaluating chatbot quality.
Google Research presents Reformer, a transformer variant using LSH attention and reversible layers to go from O(n²) to O(n log n) and handle sequences up to 64k tokens.