Expectations were high… but here’s why Google’s Gemini AI bombed

Which brings me to the topic of today’s review, Google’s Pixel 8 Pro mobile phone.

You see, Google is bringing out three versions of the Gemini AI: the Gemini Ultra version, which will come out next year, and which looks to be scarily human-like in its intelligence; the Gemini Pro version, which you can already try out in Google’s chatbot, Bard; the Gemini Nano version, which is designed to run on devices such as mobile phones.

Gemini is a “multimodal” AI, meaning it has been trained from the ground up on voice, video and photos, as well as text.

The first phone to get Gemini Nano is the Google Pixel 8 Pro, where it runs on the “neural processing unit” (NPU) built into the phone’s Tensor G3 chipset.

(Google says that other Android phones with other NPUs will eventually get Gemini Nano, but for now the new AI has only been optimised for the Pixel 8 Pro.)

Not 30 seconds ago as I write this, our Pixel 8 Pro started to receive an update adding a few Gemini features to the numerous AI-based features already on the phone. Some of these, such as “Best Take”, were already impressive, and don’t stand to gain a lot from the Gemini update.

Best Take, which is actually a feature of the Google Photos app rather than the camera app, goes through your photo album looking for photos that are similar to the photo you’ve just clicked on. Usually, it will find photos taken as part of the same burst of photos as the one you’re editing, but notionally it might also pull out old ones, too.

When it finds similar photos, it grabs the faces out of them, and then suggests those faces as alternatives to the faces in the photo you’re editing. If someone has their eyes shut in one photo, but not in some other photo, it lets you swap the face out, one person at a time until you have, quite literally, the best take.

(I’m stalling here, while I wait for the Gemini update to install. I almost can’t stand the excitement . . .)

Other AI features, though, like the “Magic Eraser” feature that lets you highlight and erase objects and people from your photos, are reasonably good, but could do with some improvements that, we hope, Gemini will bring.

Gemini is what’s known as a “multimodal” AI, meaning it has been trained from the ground up on voice, video and photos, as well as text, so there’s every reason to think it will bring enhancements to all manner of Android apps in coming weeks and months.

Oooh, the update is complete! Give me a moment . . .

But the AI feature I was most hoping for – having Gemini read entire conversations in messaging apps such as WhatsApp, and suggest auto-replies that are a tad more sophisticated than “OK” and “On my way!” – doesn’t seem to be working yet, so I’m going to tell you about the second-most exciting feature, transcript summarisation.

The voice recorder in Pixel 8 Pro was already able to transcribe conversations, meetings and interviews in real time, using its existing machine learning model.

But as of an hour ago as I write this, Gemini Nano on the Pixel 8 Pro can take those transcripts and summarise them for you, so not only do you not have to listen to people as they speak to you, you no longer even need to read the transcript when the people have stopped blathering on.

I just recorded 10 minutes of conversation over a pub lunch. Looking at the transcript, the conversation covered a wide range of topics, including: how outrageous it was that the pub wanted us to sign up to an app just to order lunch; how “sad” the burgers looked; how the chicken burger looked to be the least sad; and whether anyone ever used a particular air-gapped laptop we have in the office in case we ever need to communicate with cybercriminals.

Here is how the Pixel 8 Pro summarised the conversation, using its new Gemini Nano AI:

The conversation is about the speakers’ experiences with the internet.

The speakers discuss the importance of being able to communicate with others online and the challenges of doing so.

The speakers also discuss the importance of being able to find information online and the challenge of doing so.

So, yeah. Clearly the AI isn’t coming for our jobs just yet. Give it another release cycle.

* Since writing this review, it has emerged that Google faked many of the demonstrations in its preview of Gemini. So we very much doubt Gemini will be half as good as it appeared. We’d been wondering why the results we got in our test, outlined above, fell so far short of what Google promised. Now we know why.