2018
Pre-training becomes the standard, and employees object to military work
In June, OpenAI announced the first GPT: train on a large body of text first, then adapt to each task. In October, Google published BERT, reading context in both directions at once and rewriting the leaderboards across the main benchmarks. Reusing a pre-trained model became the standard way to work on language.
In April, Google employees sent a letter demanding the company withdraw from the Pentagon's Project Maven. Thousands signed, and the contract was not renewed. In June, Google published its AI Principles, stating it would not build weapons. Engineers had spoken up about what their work was used for, and the company answered in writing.
That same June, Microsoft acquired GitHub for $7.5 billion, taking hold of where developers gather.
OpenAI published its Charter in April, committing to stop competing and start assisting if another organisation got close to safe AGI first. Elon Musk had left its board in February. OpenAI Five, announced in June, lost to professionals at The International in August. July brought the Dactyl robot hand, and May an analysis tracking the growth in compute.
In December, DeepMind's AlphaFold took first place at CASP13 in protein structure prediction.
This block is written by the editors. It is kept separate from the sourced record below.
- February1Musk steps off the OpenAI board
- April2Employees oppose a military contract, OpenAI puts its promises in writing
- May2An AI places a phone call, and compute doubles every 3.4 months
- June4Microsoft buys GitHub, Google sets out AI principles
- July1A hand trained only in simulation turns a block in the real world
- August1OpenAI Five falls short against professionals on the big stage
- October1BERT rewrites the leaderboards across the board
- December1AlphaFold takes first place in protein prediction