The year generative models reached the public

Three releases put generative models in ordinary hands inside eight months. OpenAI announced DALL-E 2 in April, and Stability AI published Stable Diffusion in August under a license that allowed commercial use, with the weights available for anyone to download. On 30 November, OpenAI opened ChatGPT as a free research preview reached through a web page rather than an API key. Criticism arrived with the tools. Artists objected to training on their work, teachers to homework written by machine, and researchers to fluent answers that were wrong.

18 entries recorded

November

  1. OpenAI releases ChatGPT as a free research preview

    OpenAI published ChatGPT on 30 November 2022: a conversational interface to a model trained with human feedback and described as a sibling of InstructGPT. It was free during the research preview and reached through a web page, with no API key and no waitlist.

  2. Meta’s Cicero reaches human-level play in Diplomacy

    Meta AI announced Cicero on 22 November 2022, an agent that paired a language model with strategic planning and reinforcement learning to play the negotiation game Diplomacy. Across 40 anonymous games on webDiplomacy.net it more than doubled the average human score.

  3. Meta withdraws its Galactica science model after three days

    Meta AI released Galactica on 15 November 2022, a language model trained on scientific text and intended to help write papers and summaries, with a public demo. Users showed it producing fluent but false claims and invented citations, and Meta took the demo down on 17 November.

October

  1. The US restricts advanced chip exports to China

    The US Bureau of Industry and Security issued an interim final rule on 7 October 2022 restricting exports to China of advanced computing chips, chip-making equipment, and related technology. Its provisions took effect in stages on 7, 12, and 21 October.

  2. White House publishes a Blueprint for an AI Bill of Rights

    The White House Office of Science and Technology Policy published the Blueprint for an AI Bill of Rights on 4 October 2022. The non-binding framework set out five principles covering safety, algorithmic discrimination, data privacy, notice and explanation, and human fallback.

September

  1. Tesla shows its first Optimus humanoid robot prototype

    At Tesla’s AI Day in Palo Alto on 30 September 2022, Elon Musk and the engineering team showed a working prototype of the Optimus humanoid robot, nicknamed “Bumble C”, which walked onto the stage and waved. The 2021 event had shown only a dancer in a costume.

  2. Getty Images bans AI-generated content over copyright

    Getty Images said on 21 September 2022 that it would prohibit the upload and sale of AI-generated images. Chief executive Craig Peters said the policy was meant to protect customers from legal and reputational risk while the copyright questions stayed unresolved.

  3. OpenAI publishes Whisper under an MIT license

    OpenAI released Whisper on 21 September 2022, a speech-recognition system trained on 680,000 hours of multilingual audio collected from the web. Both the code and the pretrained weights were published under the permissive MIT license.

August

  1. An AI-generated image wins a Colorado State Fair prize

    Jason Allen took first place in the emerging-artist digital-arts category of the Colorado State Fair’s fine art competition with “Théâtre D’opéra Spatial”, an image he produced from hundreds of Midjourney prompts and then edited and upscaled.

  2. Stability AI publishes the Stable Diffusion weights

    Stability AI, working with LMU Munich researchers, RunwayML, EleutherAI and LAION, published the full weights of Stable Diffusion on Hugging Face on 22 August 2022 under the CreativeML OpenRAIL-M license. The model ran on a single consumer graphics card.

  3. Biden signs the CHIPS and Science Act into law

    President Joe Biden signed the CHIPS and Science Act on 9 August 2022. It authorized roughly 280 billion dollars in spending and appropriated 52.7 billion dollars in direct funding for domestic semiconductor manufacturing, research, and workforce development.

July

  1. AlphaFold database expands to over 200 million structures

    DeepMind and EMBL’s European Bioinformatics Institute expanded the AlphaFold Protein Structure Database on 28 July 2022, from roughly 1 million predicted protein structures to more than 200 million, covering nearly every sequence catalogued in UniProt.

  2. Google fires the engineer who called LaMDA sentient

    Google fired the engineer Blake Lemoine on 22 July 2022. He had told executives that the company’s LaMDA conversational model had become sentient after extended chat sessions, and had been placed on paid administrative leave on 11 June for sharing confidential material.

  3. BigScience releases BLOOM with 176 billion parameters

    The BigScience collective, hundreds of volunteer researchers coordinated around Hugging Face and using France’s Jean Zay supercomputer, released BLOOM on 12 July 2022. The 176-billion-parameter model wrote 46 natural languages and 13 programming languages.

  4. Midjourney opens its image generator to public beta

    Midjourney, a small independent lab led by David Holz, opened its text-to-image tool to public beta in July 2022. Anyone could join its Discord server and generate images from written prompts with a limited number of free credits. There was no standalone web app.

April

  1. OpenAI announces DALL-E 2, its second text-to-image model

    OpenAI announced DALL-E 2 on 6 April 2022. It produced higher-resolution and more coherent images from written descriptions than the 2021 DALL-E, and could edit existing pictures or generate variations of them. Access ran through a waitlist.

  2. Google introduces PaLM, a 540-billion-parameter model

    Google Research introduced the Pathways Language Model on 4 April 2022, a 540-billion-parameter dense transformer trained across 6,144 TPU v4 chips using Google’s Pathways system. Google reported the best few-shot scores it had published on most of the benchmarks tested.

March

  1. Chinchilla finds most large language models are undertrained

    DeepMind posted “Training Compute-Optimal Large Language Models” on 29 March 2022. Its 70-billion-parameter Chinchilla, trained on four times the data used for the 280-billion-parameter Gopher at equal compute cost, outperformed Gopher and GPT-3 across a broad set of benchmarks.