The original ELIZA source code is rediscovered
In 2021 Weizenbaum's original 1960s ELIZA source code was rediscovered in the MIT archives, letting the first famous chatbot be restored.
The Pile: An Openly Documented LLM Training Dataset
EleutherAI released The Pile, an 825 GiB English text dataset built from 22 documented sources, making an LLM training corpus's contents openly inspectable.
OpenAI introduces DALL-E
OpenAI unveils the original DALL-E, a GPT-3-scale model that generates images from natural-language text descriptions - the start of the generative image era.
OpenAI releases CLIP
CLIP learns visual concepts directly from natural-language supervision, becoming the bridge that made modern multimodal AI possible.
Pinecone launches the first managed vector database
On January 27, 2021, Pinecone launched a fully managed vector database, turning approximate nearest-neighbor search over embeddings into a simple cloud API.
C2PA founded to standardize content provenance
Adobe, Microsoft, the BBC and others formed C2PA to build an open standard for tracing where digital media came from.
On the Dangers of Stochastic Parrots
A 2021 FAccT paper by Bender, Gebru, McMillan-Major, and Mitchell that became the most-cited critique of the large-language-model paradigm.
The ICRC calls for new rules on autonomous weapons
In May 2021 the Red Cross urged states to adopt binding rules on autonomous weapons, including a ban on systems that target people.
Anthropic founded as an AI safety company
Anthropic is founded by siblings Dario and Daniela Amodei and colleagues as an AI safety and research company, surfacing with a 124 million dollar Series A.
Voiceitt launches speech recognition for non-standard speech
Voiceitt's 2021 app uses AI trained per user to recognize atypical speech and translate it into clear text or synthesized speech.
Holly Herndon launches Holly+, a shareable 'digital twin' of her voice
Musician Holly Herndon released a tool letting anyone sing in her AI voice, governed by a DAO that splits royalties, a consent-first model.
DeepMind and EMBL-EBI give away the AlphaFold protein database
DeepMind and EMBL-EBI launched a free, open database of AlphaFold protein structures in 2021, later expanding to over 200 million entries.
HyperCLOVA: Naver's Korean GPT-3
Naver unveiled HyperCLOVA, an 82-billion-parameter Korean language model, one of the first GPT-3-scale non-English models.
FDA authorizes Paige Prostate, the first AI for cancer in pathology
In 2021 the FDA authorized Paige Prostate, the first AI software to help detect prostate cancer on digitized biopsy slides, via the De Novo pathway.
UK National AI Strategy
The UK published its National AI Strategy in 2021, a ten-year plan built on three pillars of ecosystem, adoption, and governance.
DARPA Subterranean Challenge final event
DARPA's SubT Challenge tested autonomous robots mapping tunnels, caves, and urban underground; Team CERBERUS won the $2M systems competition.
Meta releases its internal Instagram teen well-being research
After the Facebook Files, Meta published the annotated internal slide decks on Instagram and teen well-being that the WSJ had cited.
DeepMind nowcasts rain with a deep generative model of radar
DeepMind's 2021 DGMR model nowcast rain for the next 90 minutes and was preferred by Met Office experts in 88 percent of cases.
The Facebook Files and the Frances Haugen disclosures
In 2021 whistleblower Frances Haugen released internal Meta research and told the US Senate the company put profit over safety.
DeepMind acquires and open-sources the MuJoCo physics engine
DeepMind bought the MuJoCo physics simulator in 2021, made it free, and open-sourced it in 2022, removing a key barrier for RL research.
Google's Project Relate builds personal speech models for atypical speech
Project Relate, launched by Google in 2021, trains a personalized model to understand a person's atypical speech and transcribe or restate it.
The UNESCO Recommendation on the Ethics of AI
In November 2021 all 193 UNESCO member states adopted the first global standard on AI ethics, built around human rights and human oversight.
AI helps mathematicians discover new results in knot theory
In 2021 DeepMind and mathematicians used machine learning to surface patterns that led to new theorems in knot theory and representation theory.
SenseTime lists in Hong Kong, the first major AI software IPO in Asia
On December 30, 2021, China's SenseTime listed on the Hong Kong exchange days after US sanctions forced it to exclude American investors.
John Deere reveals a fully autonomous tractor at CES 2022
John Deere unveiled a production-ready autonomous tractor using six stereo camera pairs and a deep neural network for 360-degree obstacle detection.
InstructGPT brings RLHF to GPT-3
OpenAI's InstructGPT used reinforcement learning from human feedback to make GPT-3 follow instructions, the direct precursor of ChatGPT.
EleutherAI releases GPT-NeoX-20B, the largest open-weight model of its day
EleutherAI released GPT-NeoX-20B in 2022, then the largest dense autoregressive language model with publicly available weights.
DeepMind controls fusion plasma with reinforcement learning
DeepMind and EPFL trained a neural network to control the magnetic coils of a real tokamak and shape fusion plasma.
FourCastNet brings Fourier neural operators to weather forecasting
NVIDIA's 2022 FourCastNet used adaptive Fourier neural operators to forecast global weather at 0.25 degrees, orders of magnitude faster than physics models.
China's Algorithm Recommendation Provisions Take Effect (2022)
China's rules on recommendation algorithms grant users opt-out rights and require providers to file their algorithms with regulators.
John Deere launches See and Spray Ultimate, a vision system that targets weeds
John Deere's See and Spray Ultimate uses computer vision to spray herbicide only on weeds, cutting non-residual herbicide use sharply.
NVIDIA Hopper and the H100: The Transformer Engine
NVIDIA's Hopper H100 added an FP8 Transformer Engine and HBM3, becoming the chip behind the generative AI boom.
DeepMind's Chinchilla scaling result
DeepMind's 2022 Chinchilla paper showed most LLMs were undertrained and that model size and training data should grow together for compute-optimal results.
PaLM, Google's 540-billion-parameter Pathways model
PaLM was a 540-billion-parameter Google model that showcased the Pathways system and showed chain-of-thought prompting unlocked multi-step reasoning at scale.
OpenAI announces DALL-E 2
OpenAI unveils DALL-E 2, a system that generates and edits high-resolution images from natural-language text prompts.
Meta releases OPT-175B with a public training logbook
Meta released OPT-175B, an open GPT-3-scale model, with its weights and a candid logbook of the infrastructure problems hit during training.
Clearview AI settles with the ACLU under Illinois biometric law
In May 2022 Clearview AI agreed to stop selling its faceprint database to most private companies nationwide to settle an ACLU BIPA suit.
Cassie sets a Guinness record for a bipedal robot 100-meter dash
Oregon State's Cassie, a two-legged robot trained with machine learning, ran 100 meters in 24.73 seconds for a Guinness World Record.
Wysa wins FDA Breakthrough Device status for an AI mental-health agent
In May 2022 the FDA gave Wysa a Breakthrough Device Designation for an AI chatbot treating chronic pain with depression and anxiety.
Cosmopolitan runs a magazine cover generated by DALL-E 2
Cosmopolitan published a cover image made with OpenAI's DALL-E 2, billing it as the first AI-generated magazine cover.
Microsoft retires emotion recognition and restricts its Face API
In June 2022 Microsoft retired AI that infers emotion, gender, and age from faces and locked down access to its Azure Face API.
BirdNET identifies birds by sound for millions of citizen scientists
Cornell's BirdNET app uses a neural network to identify over 3,000 bird species by sound, drawing millions of public submissions.
FIFA uses semi-automated offside at the 2022 World Cup
FIFA's 2022 World Cup used limb-tracking AI and a sensor ball to flag offside, the first men's World Cup to do so.
No Language Left Behind: one model translating 200 languages
Meta's NLLB-200, open-sourced in 2022 and published in Nature in 2024, translates directly among 200 languages, many low-resource.
BigScience releases BLOOM, a 176B open multilingual model
BLOOM, a 176-billion-parameter open multilingual model from the BigScience collaboration of hundreds of researchers, was released in July 2022.
Midjourney opens public beta
Midjourney, an independent lab text-to-image generator run through Discord, opened to the public in July 2022 and became one of the dominant image AI tools.
Thaler v. Vidal: an AI cannot be a patent inventor
The Federal Circuit ruled that only a human can be named an inventor on a US patent, rejecting Stephen Thaler's AI system DABUS.
The CHIPS and Science Act funds US chip manufacturing
Signed in August 2022, the CHIPS and Science Act appropriated roughly $50 billion to bring advanced semiconductor manufacturing back to the United States.