2024
Models start to see, hear and think; people leave OpenAI
The shape of the models changed in two stages. The first half belonged to multimodality and long context: Gemini 1.5 Pro claimed a one-million-token window in February, and GPT-4o handled speech, images and text in a single model in May. Claude 3 arrived in March, Claude 3.5 Sonnet in June.
In September, OpenAI released o1-preview. Built to spend longer thinking before it answered, it was read as a way to improve models other than making them bigger — the step-by-step reasoning people had been prompting for since 2022, moved into training. o3 followed in December.
Agents filled the second half of the year. Anthropic shipped computer use in October and the Model Context Protocol in November. With Gemini 2.0 in December, Google spoke of an "agentic era."
People left OpenAI. Ilya Sutskever and Jan Leike in May, John Schulman in August, CTO Mira Murati in September; Leike and Schulman both went to Anthropic. In December the company set out a plan to reorganise as a public benefit corporation. Elon Musk had sued it back in February.
Generation reached voices and moving images. OpenAI announced Sora in February, and Luma AI opened Dream Machine with a free tier in June. In January, days before the New Hampshire primary, robocalls carrying an imitation of President Biden's voice reached voters; in February the FCC confirmed that automated calls made with an AI voice were illegal under the law as it stood.
Courts and prizes weighed in. In August a federal judge ruled Google's search business an illegal monopoly, and the EU AI Act came into force. In October the Nobel Prize in Physics went to Geoffrey Hinton and the prize in Chemistry to Demis Hassabis and John Jumper.
OpenAI opened the GPT Store in January and put out GPT-4o mini and the SearchGPT prototype in July. Microsoft brought in Mustafa Suleyman from Inflection AI and released the small Phi-3 model in April. Elsewhere Meta gave away Llama 3 and the 405-billion-parameter Llama 3.1, NVIDIA announced Blackwell, and China's DeepSeek closed the year with V3.
This block is written by the editors. It is kept separate from the sourced record below.
- January4The GPT Store opens, and 'military' leaves the usage policy
- February8Google pushes forward and pulls back at once, OpenAI moves to video
- March3Anthropic claims the top with Claude 3, Microsoft hires a team whole
- April3Microsoft puts $1.5 billion into the Gulf, Google leans on DeepMind
- May9The labs reach for voice and the screen, departures continue at OpenAI
- June4Apple brings ChatGPT in, Claude gets better at the same price
- July4OpenAI reaches for search, Microsoft steps off the board
- August3A monopoly ruling for Google Search, training data heads to court
- September2OpenAI ships a model that thinks, and loses its leadership
- October6Two Nobels for AI research, Claude starts using a computer
- November3Anthropic steps into defence and opens up how models connect
- December7OpenAI ships for twelve days straight, Google bets on agents