Dec 2023, model, Google DeepMind

Gemini

The first model family from the merged Google DeepMind, trained from the start on text, images, audio and video; its launch claim of beating human experts on a standard exam and an edited demo video became part of the argument over how AI capabilities are measured and shown.

What it was

On 6 December 2023 Google announced Gemini 1.0 in three sizes: Ultra, the largest; Pro, which went into the Bard chatbot that day; and Nano, for phones. It was the first model family from Google DeepMind, formed in April 2023 when Google merged DeepMind and Google Brain under Demis Hassabis. Google said Gemini was built "from the ground up to be multimodal," trained on text, code, images, audio and video together rather than joining separate models. David Silver was a co-lead of fine-tuning on the technical report.

Google's headline claim was that "With a score of 90.0%, Gemini Ultra is the first model to outperform human experts on MMLU," a 57-subject exam benchmark on which the benchmark's authors put expert performance at 89.8 percent. The technical report showed that the 90.04 percent figure used a prompting method that sampled 32 chains of reasoning; under the standard five-example method, Gemini Ultra scored 83.7 percent to GPT-4's 86.4 percent. Ultra itself was held back for what Google called "extensive trust and safety checks, including red-teaming by trusted external parties."

Read the original

What it changed

The day after launch, reporting that began with Parmy Olson at Bloomberg showed that Google's "Hands-on with Gemini" video, which appeared to show the model responding by voice in real time to drawings and objects, had been made differently. Google's explanation, quoted by TechCrunch, was that it had created the demo by "capturing footage" and then "prompted Gemini using still image frames from the footage, and prompting via text," and that the video "shows real outputs from Gemini." TechCrunch's Devin Coldewey called it "faked"; Google disputed the word.

Hassabis introduced Gemini as "the most capable and general model we've ever built." Google's developer blog published the same day a step-by-step account of the image-and-text prompts behind the demo.

The arguments it moved

Can language models reach general intelligence?

Gemini's launch showed two ways claims about capability can outrun the evidence behind them: benchmark numbers depend on how a model is prompted, and a demo depends on how it is edited. The 90.0 percent MMLU claim compared a 32-sample method for Gemini against a figure that Google's own table showed GPT-4 approached with the same method, while on the standard setting GPT-4 scored higher. The demo dispute, reported on 7 December 2023, turned on the gap between what a system produced from curated still frames and text prompts and what viewers took to be live multimodal conversation.

Compare every leader's position on this

Who should have access to powerful models?

Gemini's weights were not released, and Google DeepMind has kept its frontier Gemini models closed since. Hassabis has argued against fully open release of frontier systems, challenging advocates to say how they would handle what he calls the "bad actor problem."

Compare every leader's position on this

Positions it bears on

Demis Hassabis, Open versus closed models

Sceptical of fully open release of frontier models, challenging advocates to explain how they would stop dangerous capabilities reaching bad actors.

Gemini, built under Hassabis, was released without weights, in line with his scepticism about opening frontier models.

Critics and counterpoints

Sources

Last verified 2026-09-22.