(Credit: Bloomberg/Contributor/Getty Images)
Google has released an update for its Gemma AI models that promises to improve their overall ability to follow instructions and provide facts, but still hasn't rolled out a fix for Gemini over six weeks after it faced controversy for failing to create historically accurate depictions of people.
The new versions of Gemma—1.1 Instruct 2B and “7B”—are "better across the board," but "the gains are larger for '7B' than 2B," says Google DeepMind researcher Robert Dadashi.
Google also fixed a bug that occurred if users tried to switch topics while conversing with Gemma and reduced occurrences of the word "Sure," which Gemma often used to start sentences.
Overuse of specific words or phrases, especially ones in response to a query, is a common issue with generative AI models. It's made it easier for journalists and researchers to detect AI use in academic papers (OpenAI's ChatGPT often uses "As of my last knowledge update," for example).
Gemma's training data didn't change much with the update, according to Dadashi, but its reinforcement learning (RL) algorithm was switched as part of the update. Amazon defines RL as the machine learning version of the human trial-and-error process, deploying a "reward-and-punishment paradigm" to train AI models.
Dadashi says that some of Gemma's tendencies can now also be tweaked with more specific prompts. For instance, asking Gemma to put responses in paragraphs can reduce the AI's tendency to respond with bullet points. Giving Gemma a word count goal can also prevent the AI from rambling on for too long.
Gemma was made with "the same research and technology" as Gemini and first launched in February for developers to add to their AI tools. Gemma is "open" but not "open-source," meaning it's available for third-party use but has more technical restrictions. Nvidia is offering Gemma as part of its Chat With RTX tool, and Opera recently announced that Gemma will be one of the AI models its web browser's developers can use to build new features.
While Gemma gets updates, Gemini continues to see restrictions on the kinds of content it's allowed to produce. Last month, Google barred Gemini from being about to answer any questions about political elections.
Gemini's ability to create images of people remains blocked as well. When its image-generating controversy began, Google SVP Prabhakar Raghavan said the AI model had "become way more cautious than we intended," resulting in historical inaccuracies. PCMag has reached out to Google for comment on when a fix for Gemini might be rolled out to the public.