OpenAI holds a model back and turns for-profit
On 14 February, OpenAI announced GPT-2 and said it would not release the largest model, citing the risk of misuse, publishing smaller versions first instead. It challenged the convention of publishing everything, and opinion divided. The full 1.5-billion-parameter model followed on 5 November, with OpenAI saying the misuse it had worried about had not materialised.
In March, OpenAI created OpenAI LP, a for-profit subsidiary structured to cap investor returns at a fixed multiple, with Sam Altman as CEO. On 22 July, Microsoft invested $1 billion and became OpenAI's exclusive cloud provider. The judgment that a non-profit could not fund compute at that scale had taken structural form.
In April, OpenAI Five beat OG, the Dota 2 world champions. In October, a robot hand solved a Rubik's Cube one-handed.
Google put BERT into Search in October, calling it the biggest change since 2015 — the previous year's paper reaching the product. An AI ethics council it had launched in April was disbanded after ten days amid criticism of its membership. In December, Sundar Pichai took on the role of Alphabet CEO as well.
This block is written by the editors. It is kept separate from the sourced record below.
OpenAI announced GPT-2, a 1.5-billion-parameter language model trained solely to predict the next word on 40GB of text from 8 million web pages. Without task-specific training, it showed broad abilities in text generation, question answering, and summarization. Citing misuse concerns, OpenAI withheld the full model and released only a smaller version as an experiment in staged release, sparking industry debate over publication norms.
OpenAI announced OpenAI LP, a new "capped-profit" company. Returns for first-round investors are capped at 100x their investment, with any excess owned by the nonprofit OpenAI entity, enabling the fundraising needed for large-scale compute and talent. The LP is governed by the nonprofit's board of directors, with Sam Altman as CEO, Greg Brockman as chairman and CTO, and Ilya Sutskever as chief scientist.
Google dissolved the Advanced Technology External Advisory Council (ATEAC), its external AI ethics advisory body, roughly ten days after forming it. The trigger was a petition signed by some 2,400 employees demanding the removal of Kay Coles James, president of the conservative Heritage Foundation, from the council; the inclusion of a drone company executive had also revived concerns about military uses of AI, and one member had already resigned. Google said it had "become clear that in the current environment, ATEAC can't function as we wanted. So we're ending the council and going back to the drawing board." The episode exposed how hard it is for an AI company to design external oversight.
OpenAI Five, OpenAI's reinforcement learning system, won two back-to-back games against OG, the reigning Dota 2 world champion team, at the OpenAI Five Finals event. It was the first time an AI beat the world champions in an esports game. The result demonstrated that scaled-up self-play reinforcement learning could master a complex team-based game, avenging OpenAI Five's losses at The International the previous year.
Microsoft announced a $1 billion investment in OpenAI and a partnership aimed at building AGI (artificial general intelligence). The companies agreed to jointly develop AI supercomputing technologies on Azure, with Microsoft becoming OpenAI's exclusive cloud provider and preferred partner for commercializing its pre-AGI technologies. A single $1 billion commitment from a company to a research organization pursuing AGI was an unusual sum at the time.
OpenAI presented research in which a human-like robot hand solved a Rubik's Cube. The neural networks were trained entirely in simulation, using the same reinforcement learning code as OpenAI Five together with a new technique called Automatic Domain Randomization (ADR), and coped with perturbations never seen in training. The work showed that reinforcement learning can tackle physical-world tasks requiring high dexterity.
Pandu Nayak, Google's VP of Search, announced that BERT had been applied to Search, improving its handling of longer, conversational queries where small words such as prepositions carry the meaning. Google said it would affect one in ten queries and called it the biggest change to search since RankBrain in 2015. The post used the query "2019 brazil traveler to usa need a visa" as an example, explaining that previous algorithms had missed that it concerned a Brazilian traveling to the US. Research the company had published a year earlier was now running inside one of the world's largest services.
As the final stage of its staged release, OpenAI published the model weights and code for the largest version of GPT-2, with 1.5 billion parameters. Since the February announcement, OpenAI had gradually released larger versions while studying misuse risks, and it also released a model to help detect GPT-2-generated text, framing the process as a test case for responsible publication of powerful models.
Google co-founder Larry Page stepped down as CEO of Alphabet, with Google CEO Sundar Pichai taking on the parent company's top job as well. Co-founder Sergey Brin also left his role as Alphabet president, which was eliminated. Both kept their board seats but stepped away from day-to-day operations. Page and Brin wrote that it was "the natural time to simplify our management structure." With both founders stepping away from daily operations, decision-making for Google and Alphabet was consolidated under Pichai.