What You Need To Know From Today's Google IO: PaLM 2, Pixel Fold, AI Everywhere

At its downsized developer conference on Wednesday, Google showed off present and planned Pixel hardware – a foldable Pixel among them – and PaLM 2, a large language model that follows in the footsteps of last year's initial Pathways Language Model (PaLM) and now whispers to various Google products.

There was more, too, crammed into a single day. It was the first time since the event debuted in 2008 that the show hasn't run to two or three days, barring its 2020 absence. Coincidentally, the tech giant's staff roster is shorter by 12,000 this year, although Google still has plenty left around.

The pre-show could have been shorter still. A DJ, backed by inelegant generated graphics, subjected the audience to unpleasant electronic tunes smothered by grating vocoder lyrics. And yet people applauded in the chilly open-air stadium.

Alphabet CEO Sundar Pichai arrived on stage, here at the Shoreline Amphitheater in Silicon Valley, to say that Google is reimagining all of its core products, including search, to take advantage of advances in generative AI. Which rather set the tone for the morning.

You can sit through the three-and-a-half-hour PR blitz yourself using the video below, check out our summary here, or both. Or neither. We're not a cop.

Youtube Video

Google has been researching and developing artificially intelligent systems for years, coming up with the transformer architecture powering today's hyped-up chatbots, and inserting this sort of tech into its products, services, and internal pipelines during that time. It now finds itself playing catch up, at least in terms of public perception.

Rival OpenAI, now largely tied to Microsoft, has made waves over the past year by making its machine learning models GPT-3, ChatGPT, GPT-4, and DALL-E publicly available and Microsoft has put these stochastic parrots into products like GitHub Copilot, Bing, and its 365 cloud.

Simultaneously, open source projects including Stable Diffusion in August and reproductions of Facebook's LLaMA that appeared in March, have prompted concern among Googlers about being left behind.

Google's immediate answer is PaLM 2, which can handle writing, coding, and calculations with the occasional competency people have come to expect from LLMs, Pichai said. The foundational model has been trained on more than 100 languages, on scientific data sets, and code. It comes in four sizes, known as: Gecko, Otter, Bison, and Unicorn.

PaLM 2 has medical and security flavors, known as Med-PaLM 2 and sec-PaLM. There's a followup model in the works called Gemini, which is multimodal, much like the Meta's ImageBind.

Pichai said Google is introducing 25 products and features based on PaLM 2. Several of these not-quite-baked capabilities can be tried prior to general release through Search Labs.

A bard by any other name

Bard, the AI chatbot, is now running on PaLM 2, and is being used among other things to help developers code. The previously waitlisted service is now open to the public, and 180 countries are included.

Sissie Hsiao, VP and general manager of Google Assistant, said Bard now knows 20 programming languages, and next week the AI service is scheduled to gain the ability to cite the source of its suggestions.

Bard is being connected to various Google apps and third-party services, like Adobe Firefly, Instacart, and others. Hsiao demonstrated by asking Bard to return a list of colleges with animation programs, to show them on a map, then to convert the list to a table, to add a column, and then move the data to Google Sheets.

• Google's here to boost your cloud security and the magic ingredient? AI, of course
• Pixies keep switching off my morning alarm, says Google Pixel owner
• It is now safe to turn off your brain: Google CEO asked Bard to plan his dad's 80th birthday
• Now Google to shove its answer to ChatGPT into Gmail, Docs, apps via APIs, more

Cathy Edwards, VP of engineering, showed off how generative AI is being shoehorned into the Google Search interface. Presented with the question "what's better for a family with kids under 3 and a dog, Bryce Canyon or Arches?" – a query that in the past would be likely to be carried out as a series of keyword submissions, Edwards observed – Google Search returned AI generated text atop the more familiar stack of sponsored links and search result links.

Essentially, with AI assistance, multiple queries can be handled at once, to deliver a single, recommended answer. Presenting this result may require further design adjustment as there's only so much information that can be presented on-screen in a coherent way. The evolving Search experience is available to those who sign up for Search Labs.

"AI is not only a market-enabler, it is also a big platform shift," said Pichai.

Google is also, in the coming months we're told, going to add features to its image search system to help people understand where pictures probably first appeared, when those snaps were first indexed, and where else you can find them. That info, plus support for an in-file metadata label for AI-generated images, is expected to help netizens tell the difference between legit photos and machine-made or faked ones.

A Duet in the cloud

Later this year, the various PaLM-powered AI capabilities under development will be wrapped into a single suite called Duet AI for Workspace. Google Slides will let you generate slide images with a few words. Google Sheets will be able to organize rows and columns on demand. In Google Meet, you'll be able to generate unique backgrounds. And in Docs, you'll be able to get AI writing help, for what it's worth.

Google Cloud customers will have access to Duet AI for Google Cloud, with features like code assistance, chat assistance, and Duet AI for AppSheet – a no-code way to create business apps. These features are being made available to those in the company's trusted tester program.

That code assistant tool, dubbed Codey, is pitched as a rival to Microsoft's GitHub Copilot.

Thomas Kurian, head of Google Cloud, chimed in with: "Google is using the power of AI to transform the way you work."

Google Cloud customers can also now get on the waiting list for A3 GPU VMs, which include 8 H100 Nvidia GPUs, NVIDIA NVSwitch and NVLink 4.0 bandwidth of 3.6 TB/s, 4th Gen Intel Xeon Scalable processors, 2TB of memory via 4800 MHz DDR5 DIMMs, and 10x greater networking bandwidth.

GCP is branding this set of technologies as an A3 supercomputer, designed for training, testing, and running ML models.

And then there are the Pixels

Rick Osterloh, SVP at Google, talked up the 6.1-inch Pixel 7a, now available for $499 in the US at least, featuring a 72 percent larger camera sensor than its predecessor and the same chipset as last year's model.

"From the beginning, Pixel was conceived as an AI-first mobile computer," he said, pointing to the phone's Tensor G2 chip for AI computation.

The handset also includes AI-powered Call Assist, with functions like Direct My Call, Call Screen, Hold for Me, Clear Calling (noise reduction), and Wait Times (hold time estimation).

It's a phone that folds ... It's the Pixel Fold

But Google does expensive phones too, as it showed with the long-anticipated foldable version of its handset, the Pixel Fold, shipping next month.

Starting at $1,799, the handset folds out to become a 7.6-inch tablet and can perform tricks like live translation between two languages on panels facing different directions. Pre-orders come with a Pixel Watch, too.

The Pixel Fold played a role in demonstrating Google's Universal Translator project – with a screen on both sides of the device, it looks like an ideal way to translate speech to text in real-time conversation. Then again, Google has been pushing this concept for years so we've heard this before.

And finally there's the upcoming Pixel Tablet, an 11-inch fondleslab that can be integrated into Google's home automation systems. It's your for $499 and Google made much of the charging and speaker dock that comes free when it ships next month, possibly in a sly dig at those companies that charge a lot for kit and then even more for accessories. ®