High
DeepMind RT-1: the first Transformer trained on real robotics data
DeepMind releases RT-1, a robotics transformer trained on 130,000 real episodes with 13 robots, generalizing to never-seen tasks.
Robotics DeepMindRT-1Robotics Transformer
2022
52 entries
Medium
Constitutional AI: the model self-corrects without humans in the loop
Anthropic publishes Constitutional AI: instead of pure RLHF, the model critiques and revises its own responses following a written 'constitution'. Less human labeling, more transparency.
AI Security AnthropicConstitutional AIRLAIF
Medium
Boston Dynamics adds visual AI to Spot: map-free autonomy
Spot gains advanced autonomous navigation and industrial anomaly detection via visual AI, operating without pre-loaded maps.
Robotics Boston DynamicsSpotAutonomous Navigation
Landmark
★ On my workflow
ChatGPT: AI lands in everyone's browser
OpenAI launches ChatGPT, a free conversational interface on GPT-3.5 aligned via RLHF. It crosses one million users in five days.
Foundation Models OpenAIChatGPTGPT-3.5
Medium
Stable Diffusion 2.0: new architecture and OpenCLIP encoder
Stability AI releases SD 2.0 with OpenCLIP replacing CLIP, native 768x768 resolution, a new depth2img model, and improved inpainting. A controversial release due to breaking compatibility with existing LoRAs and prompts.
Image & Video Gen Stable Diffusion 2.0Stability AIOpenCLIP
Medium
Notion AI alpha: AI inside the tool you already work in
Notion launches Notion AI in private alpha, GPT integrated inside pages: summarize, rewrite, translate, brainstorm without leaving the document.
Enterprise AI NotionNotion AIProductivity
Medium
Galactica: Meta launches (and pulls in three days) a science LLM
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.
Foundation Models MetaGalacticaScience LLM
High
NVIDIA Triton Inference Server 2.x: the de facto standard for production inference
NVIDIA consolidates Triton as the open-source platform for serving PyTorch, TensorFlow, and ONNX models in production, with dynamic batching, multi-GPU support, and gRPC/HTTP APIs.
AI Infrastructure NVIDIATritonInference Server
Medium
HuggingFace Accelerate: One Python Script for CPU, GPU, TPU, and Mixed Precision
HuggingFace Accelerate provides a unified API that runs the same training code on any hardware without changes, becoming the backbone of most open LLM training pipelines.
AI Infrastructure AccelerateHuggingFacemulti-GPU
Landmark
LangChain: the framework for LLM applications is born
Harrison Chase releases LangChain, an open-source Python library to chain LLMs with prompt templates, memory, tools and external data sources. It will become the default stack of the first LLM apps.
Agents LangChainFrameworkLLM Apps
Medium
Textual Inversion: inject a custom concept into diffusion models
Weizmann Institute publishes Textual Inversion: learning a new text token representing a custom concept from 3-5 images, without modifying model weights.
Image & Video Gen Textual Inversionpersonalizationembedding
High
EnCodec: Meta AI compresses audio with neural networks and beats Opus
EnCodec compresses 24kHz stereo audio to just 1.5–12 kbps at quality surpassing Opus, becoming the standard vocoder for modern neural TTS.
Voice & Audio EnCodecNeural CodecAudio Compression
High
MT-OPT: Google trains a single robot policy on 800+ tasks and 57,000 hours of real data
Google pre-trains a single policy on over 800 real robot tasks and 57,000 hours of real-world data, demonstrating for the first time zero-shot transfer to new tasks through large-scale multi-task offline learning.
Robotics MT-OPTmulti-task robot learningoffline RL
High
GPTQ: 4-bit post-training quantization making GPT-scale inference practical
Frantar et al. (ETH Zurich) publish GPTQ: accurate 4-bit quantization without significant fine-tuning, the first technique to make inference of 175B-parameter models practical on consumer hardware.
AI Infrastructure GPTQQuantizzazione4-bit
Landmark
ReAct: the framework that unites reasoning and acting in LLMs
Yao et al. introduce ReAct, a schema alternating explicit thoughts (Thought) and concrete actions (Act) in LLMs, the theoretical foundation of all modern agents.
Agents ReActReasoningTool Use
Medium
Imagen Video and Phenaki: Google answers on text-to-video
A week after Make-A-Video, Google Research unveils Imagen Video and, around the same time, Phenaki: two different approaches to text-to-video, with longer, more coherent clips.
Image & Video Gen GoogleImagen VideoPhenaki
Medium
Make-A-Video: Meta unveils the first credible text-to-video
Meta AI shows Make-A-Video, a system that generates short animated clips from a text description by reusing a pre-existing text-to-image model.
Image & Video Gen MetaMake-A-VideoText-to-Video
Medium
Hugging Face Inference Endpoints: deploy LLMs in two clicks
Hugging Face launches Inference Endpoints, a managed service to deploy Hub models on AWS, Azure or GCP with autoscaling, on-demand GPUs and private endpoints.
AI Infrastructure Hugging FaceInference EndpointsDeployment
High
Flan-T5 and Flan-PaLM: instruction tuning scales to 1,800 tasks
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.
Foundation Models Flan-T5instruction tuningchain-of-thought
High
Whisper open source: audio transcription becomes a commodity
OpenAI releases Whisper under MIT license: a speech-to-text model trained on 680,000 hours of multilingual audio, near commercial-grade quality, runs locally.
Voice & Audio OpenAIWhisperASR
Medium
Character.AI: persona chatbots from ex-Google founders
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.
Foundation Models Character.AIChatbotPersona
High
Prompt Injection: when user input hijacks system instructions
Riley Goodside and Perez et al. formalize Prompt Injection: an attack where malicious user input overwrites an LLM's system instructions, bypassing policies and guardrails.
AI Security Prompt InjectionLLM SecurityAdversarial Attacks
High
AudioLM: Google teaches a language model to listen and continue audio
AudioLM generates long-range coherent audio using two tiers of tokens — semantic and acoustic — with no text or score conditioning.
Voice & Audio AudioLMLanguage ModelAudio Generation
High
DreamBooth: generate your subject in any style with 3-5 photos
Google Research publishes DreamBooth: fine-tune a diffusion model on 3-5 images of a specific subject to reproduce it in any context or style. Foundation of all personalized AI image generation.
Image & Video Gen DreamBoothpersonalizationfine-tuning
Landmark
Stable Diffusion: image generation goes open
Stability AI publicly releases weights and code of a text-to-image latent diffusion model that runs on a consumer GPU. AI image generation leaves the cloud.
Image & Video Gen Stable DiffusionStability AIDiffusion Models
Medium
GitHub Copilot: 40% of code in active files written by AI
GitHub publishes first real-world data: 40% of code in files with Copilot active is AI-generated. First quantitative benchmark on AI tools' actual impact on developer output.
AI Coding GitHub CopilotDeveloper ProductivityResearch
High
SayCan: grounding LLMs in robot affordances
Google Robotics shows how to combine an LLM for high-level planning with robot value functions that filter only physically executable actions.
Robotics GoogleSayCanEmbodied AI
High
diffusers v0.1: the standard library for diffusion models
Hugging Face releases diffusers, a modular Python library for diffusion models — text-to-image, audio and beyond. It quickly becomes the de facto standard.
Open Source Models Hugging FaceDiffusersLibrary
Medium
DALL-E 2 enters beta: generative image AI for the public
OpenAI opens DALL-E 2 in beta to over one million waitlist users, with a pay-per-image credit system. First large-scale consumer product for image generation.
Image & Video Gen OpenAIDALL-E 2Beta
High
BLOOM 176B: the first truly open large multilingual LLM
The BigScience collective releases BLOOM, a 176-billion-parameter model trained on 46 human languages and 13 programming languages, under an open RAIL license.
Open Source Models BigScienceBLOOMHugging Face
High
Midjourney opens public beta on Discord
Midjourney opens its public beta with a text-to-image model accessible via a Discord bot. Its strong aesthetic default and community turn image generation into a mass phenomenon.
Image & Video Gen MidjourneyDiscordText-to-Image
High
Red Teaming LLMs with LLMs: the DeepMind paper that changed safety testing
Perez et al. (DeepMind) show that an LLM can be used as an automatic attacker against another LLM, discovering undesired behaviors at a scale impossible for human teams.
AI Security Red TeamingDeepMindLLM Safety
Medium
UL2: Google unifies pretraining paradigms with Mixture-of-Denoisers
Google Research combines three major pretraining objectives into a single 20B model, outperforming GPT-3 on many benchmarks at one-eighth the parameters.
Foundation Models UL2mixture of denoiserspretraining
Medium
Tabnine 3.0: AI code completion with privacy-first and local models
Tabnine releases version 3.0 with local or cloud model support, becoming the first mature AI code completion product on the market before Copilot's rise.
AI Coding TabnineCode CompletionLocal AI
Landmark
FlashAttention: IO-aware attention that revolutionizes transformer training
Tri Dao (Stanford) publishes FlashAttention: an IO-aware implementation that avoids materializing the attention matrix in HBM, achieving 2-4x speedup and 10x less GPU memory.
AI Infrastructure FlashAttentionAttentionTransformer
Landmark
GitHub Copilot: AI for code becomes a product for everyone
GitHub announces general availability of Copilot for all developers at $10/month. It's the first mass-market AI tool living inside the daily code editor.
AI Coding GitHubCopilotOpenAI
High
SoundStream: Google's first real-time neural audio codec
SoundStream introduces Residual Vector Quantization to compress audio at 3kbps with quality surpassing Opus at 12kbps, founding the architecture of all modern neural codecs used in audio LLMs.
Voice & Audio SoundStreamneural codecRVQ
Medium
Tortoise TTS: convincing voice cloning from 3 seconds of audio
James Betker releases Tortoise TTS, an open source model with few-second voice cloning and human-like vocal quality — the first real breakthrough in accessible TTS.
Voice & Audio TTSVoice CloningOpen Source
High
Imagen: Google enters text-to-image generation
Google Research unveils Imagen, a text-to-image diffusion model that uses a frozen T5 text encoder and beats DALL-E 2 on benchmarks for photorealistic fidelity.
Image & Video Gen GoogleImagenText-to-Image
High
Gato: DeepMind tries a single agent for 600+ tasks
DeepMind unveils Gato, a 1.2-billion-parameter Transformer that with the same weights plays Atari games, controls a robot arm, captions images and chats.
Multimodal AI DeepMindGatoGeneralist Agent
High
Meta OPT-175B: the first 175-billion LLM opened to researchers
Meta AI releases OPT-175B, a language model comparable in size to GPT-3, with weights available to researchers and a public training logbook.
Open Source Models MetaOPTOpen Source
High
DeepMind Flamingo: the first few-shot visual language model
Flamingo brings few-shot learning to vision: SOTA on VQA and captioning with no task-specific fine-tuning.
Multimodal AI Visual Language ModelFew-Shot LearningVQA
High
NaturalSpeech: Microsoft achieves human parity on LJSpeech benchmark
NaturalSpeech is the first TTS system to achieve a MOS statistically indistinguishable from recorded human speech on the LJSpeech benchmark, marking a historic milestone for speech synthesis.
Voice & Audio NaturalSpeechMicrosofthuman parity
High
DALL·E 2: the quality leap in image generation
OpenAI announces DALL·E 2, a diffusion-based text-to-image model producing photorealistic 1024×1024 images. Initially waitlist-only, public access in July.
Image & Video Gen OpenAIDALL-E 2Diffusion
Medium
PaLM 540B: Google's GPT-3 answer brings chain-of-thought
Google publishes PaLM, a 540B-parameter model trained on the new Pathways system. Demonstrates emergent reasoning capabilities when guided with chain-of-thought.
Foundation Models GooglePaLMPathways
Landmark
Chinchilla: the big models were undertrained
DeepMind publishes the Chinchilla paper and shows that, given equal compute, smaller models trained on far more tokens beat oversized undertrained ones.
Foundation Models DeepMindChinchillaScaling Laws
Landmark
NVIDIA H100 and Hopper architecture: the foundation-model GPU
At GTC 2022 NVIDIA unveils the Hopper architecture and the H100 GPU, with FP8 Transformer Engine and NVLink 4. It will become the hardware substrate for nearly every large LLM of the following years.
AI Infrastructure NVIDIAH100Hopper
High
Self-Consistency: sample multiple reasoning paths for better answers
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.
Foundation Models Chain of ThoughtSelf-ConsistencyReasoning
High
AlphaCode: DeepMind takes on competitive programmers
DeepMind unveils AlphaCode, a system that generates code for competitive programming problems and ranks in the top half of human participants on Codeforces.
AI Coding DeepMindAlphaCodeCompetitive Programming
Medium
Coqui TTS: open source speech synthesis for everyone
Coqui TTS is an open source Python library for quality text-to-speech, forked from Mozilla TTS, supporting over 1100 languages and adopted by the HuggingFace community.
Voice & Audio CoquiTTSOpen Source
High
InstructGPT: the fine-tuning that teaches GPT to obey
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).
Foundation Models OpenAIInstructGPTRLHF
Medium
UnifiedIO (AI2): first unified sequence-to-sequence model for text, images, audio, and video
AI2 and University of Washington present UnifiedIO: the first sequence-to-sequence model capable of handling text, images, audio, video, and structured data as both inputs and outputs through a single architecture, trained on 80+ tasks simultaneously.
Multimodal AI UnifiedIOmultimodalunified model