AIster · The model wars · Launches

OpenAI previews Sora as Google unveils Gemini 1.5

Minute-long AI video and a million-token context window, announced on the same day

Key facts

Date

15 Feb 2024

Era

The model wars

Interest

Launches

Threads

Model wars

Who

OpenAI (Sora); Google (Gemini 1.5)

Sora

Text-to-video, clips up to a minute; public access came in Dec 2024

Gemini 1.5 Pro

128,000 tokens as standard; up to 1 million in private preview

On 15 February 2024 OpenAI previewed Sora, a model that generates video from a written prompt. OpenAI said it could produce clips as long as one minute, and it published a set of samples, among them a street scene in snowy Tokyo. OpenAI described Sora as a diffusion model that uses a transformer architecture: it starts from video that looks like static noise and removes the noise over many steps.

Limited access Sora was not opened to the public. OpenAI gave access to red teamers, domain experts in areas such as misinformation, hateful content and bias, who were asked to probe it for harms, and to a number of visual artists, designers and filmmakers. OpenAI also listed weaknesses: the model could struggle to simulate the physics of a complex scene accurately, and could confuse spatial details of a prompt, such as left and right.

The previews also reached the film industry. On 22 February the filmmaker and studio owner Tyler Perry told The Hollywood Reporter that he had put an $800 million expansion of his Atlanta studio, which would have added 12 soundstages, on hold indefinitely after seeing what Sora could do.

Public access arrived on 9 December 2024, when OpenAI made Sora available to ChatGPT Plus and Pro subscribers, though not in the EU at launch.

A million tokens The same day, Google announced Gemini 1.5. Its first model, 1.5 Pro, came with a standard context window of 128,000 tokens (tokens are the small chunks of text a model reads), and an experimental version handled up to one million tokens for a limited group of developers and enterprise customers. Google said that was enough to take in 1 hour of video, 11 hours of audio, a codebase of more than 30,000 lines or more than 700,000 words in one go. It also used a Mixture-of-Experts design, in which only the parts of the network most relevant to an input are activated.

Why it mattered Neither was widely available that day, and the two announcements pointed in different directions. Sora showed that the methods behind image generators could be stretched to longer, more coherent video. Gemini 1.5 showed that a model could take in a whole book, film or codebase in a single prompt.

Questions people ask

When was Sora announced?

OpenAI previewed Sora on 15 February 2024. At first only red teamers and a number of visual artists, designers and filmmakers could use it; OpenAI opened it to ChatGPT Plus and Pro subscribers on 9 December 2024.

What does a one-million-token context window mean?

It is how much a model can read in a single prompt. Google said Gemini 1.5 Pro's experimental one-million-token window could take in 1 hour of video, 11 hours of audio, over 30,000 lines of code or over 700,000 words at once.

Sources

1. OpenAI, Sora: Creating video from text, 2024-02-15
2. Google, Our next-generation model: Gemini 1.5, 2024-02-15
3. TechCrunch, OpenAI's Sora video generator is launching for ChatGPT Pro and Plus subscribers — but not in the EU, 2024-12-09
4. The Hollywood Reporter, Tyler Perry Puts $800M Studio Expansion on Hold After Seeing OpenAI’s Sora: “Jobs Are Going to Be Lost”, 2024-02-22