Google is rolling out an update to the free tier of its AI-powered chatbot Gemini, focusing on improving the platform's performance and making it more broadly available, as it tries to keep up with generative AI rivals like Anthropic and OpenAI.
Starting Thursday, Gemini 1.5 Flash, the lightweight multimodal model that Google unveiled in May, will be available on web and mobile in 40 languages and about 230 countries. Google claims that Gemini 1.5 Flash offers improved quality and latency, with notable improvements especially in inference and image understanding.
The upside for Google is that it may be cheaper to run it on the backend.
In announcing Gemini 1.5 Flash, Google emphasized that the model is a “condensed” and highly efficient version of Gemini 1.5 Pro, built for generative AI workloads that the company describes as “narrow” and “high frequency.” Given the overhead of delivering a chatbot platform like Gemini (see OpenAI's ChatGPT invoice), there's no doubt that Google is eager to jump on any opportunity to reduce costs, especially if that opportunity has the fortunate side effect of improving performance in other areas.
Beyond the new base model, Google says it has expanded Gemini's context window to 32,000 tokens, which is roughly the equivalent of 24,000 words (or 48 pages of text).
Image credit: Google
Context, or context window, refers to the input data (e.g. text) that a model considers before generating an output (e.g. additional text). Some advantages of models with larger context are that they can summarize and reason about longer text snippets or files (at least in theory), and in the context of a chatbot, they are less likely to forget topics that were recently discussed.
Until now, uploading files to Gemini for analysis required Gemini Advanced, a paid version of Gemini available through Google's $20-per-month Google One AI premium plan. But Google has announced that soon all Gemini users will be able to upload files from Google Drive or local devices.
“You can do things like upload your economics study guide and ask Gemini to create practice questions for you,” Amar Subramanya, Google's vice president of engineering, wrote in a blog post shared with TechCrunch. “Soon, Gemini will also be able to analyze your data files, uncovering insights that you can visualize in charts and graphs.”
To combat hallucinations (as happens with generative AI models like Gemini 1.5 Flash), Google is previewing a feature that will display links to related web content beneath certain Gemini-generated answers. English Gemini users in certain regions will see a “tip” icon at the end of the Gemini-generated paragraph with a link to your website (or email, if you've allowed Gemini to access your Gmail inbox) and from there they can dig deeper.
The move comes after revelations that Google's generative AI models are prone to some pretty nasty hallucinations, like suggesting harmless glue for pizza recipes and concocting fake book reviews by real people. Earlier this year, Google rolled out a “Double Check” feature that highlighted whether statements made on Gemini were corroborated or contradicted by other online sources. But the related content links seem like an effort to be more transparent about which sources Gemini gets its information from.
The question this reporter has is how often and with what accuracy Gemini will surface relevant links. TBD.
But Google isn't waiting around to flood the channel.
After introducing Gemini in Messages on some devices earlier this year, Google has now rolled out the feature in the European Economic Area (EEA), the UK, and Switzerland, allowing users to chat in newly added languages such as French, Polish, and Spanish. Users can call Gemini in Messages by tapping the “Start Chat” button and selecting Gemini as the person they want to chat with.
Google is releasing the Gemini mobile app in more countries, expanding access to Gemini to teens around the world.
The company introduced a teen-focused Gemini experience in June, allowing students to sign up using their school accounts, but it wasn't available in all countries. That will change in the next week, when Gemini will become available to teens in all countries and regions where it's typically available to adults.
As part of the rollout, Google said it would introduce “additional policies and safeguards” to protect teens, without providing details. It also introduced a new onboarding process tailored to teens, along with an “AI literacy guide” to, as Google puts it, “help teens use AI responsibly.”
There's been a lot of debate about whether kids are using generative AI tools for their intended purposes or for the wrong reasons, and Google is surely keen to avoid headlines suggesting that Gemini is a plagiarised essay generator or that it could be giving teenagers inappropriate advice on personal matters, so it's taking all the steps it can to prevent the worst from happening.