The Prompt: Google’s New AI Mode

Welcome back to The Prompt.

What it means to “Google” something might soon change dramatically. The tech giant announced a new AI mode for its search engine that includes a chatbot-like feature that enables users to ask questions about anything and get answers through Google’s AI models. The feature, which can be accessed through a new tab in Google, comes after the company’s somewhat problematic launch of AI overviews. Google’s AI mode won’t just serve up information–it will also book flights, buy event tickets and surface products.

Now let’s get into the headlines.

BIG PLAYS

AI’s tendency to hallucinate and cook up nonsensical fallacies is common knowledge at this point. But the latest AI glitch came last week when Elon Musk’s AI chatbot, Grok replied to hundreds of unrelated prompts and even random selfies on X, with assertions of violence against white people in South Africa, Forbes reported. The claims echoed statements made by Musk himself, a white man from South Africa who has in the past made claims that the country is racist against white people. The chatbot’s untriggered responses went viral as leaders like OpenAI’s CEO Sam Altman quipped about it on X. xAI, the company behind Grok, said the responses were a result of an “unauthorized modification,” to the system. The incident underlines that all chatbots— regardless of how their makers market them— are encoded with biases.

ETHICS + LEGAL

In one of the first legal cases of its kind, the government is investigating AI music fraud, Wired reports. Mike Smith, a lesser known musician, allegedly made $10 million in royalties from platforms like Spotify by creating an army of bots that streamed AI-generated songs (tied to artists that didn’t exist) on an indefinite loop.

AI DEAL OF THE WEEK

Buzzy AI-powered notetaking app Granola just raised $43 million in funding at $250 million valuation. The Series B round was led by Nat Friedman and Daniel Gross’ venture outfit NFDG. Granola CEO Chris Pedegral told Forbes the company’s aim is to help meetings be more productive by providing people with all the context they need to get work done. Unlike other AI notetakers that automatically join and record Zoom meetings and email AI-generated notes and transcriptions to attendees after the meeting, Granola doesn’t record audio or video without permission. “It’s super awkward when these meeting bots join and the person isn’t there… We want it to feel less invasive,” says Pedegral, whose previous startup, AI learning app called Socroatic, was acquired by Google in 2018. The app has seen early traction among venture capitalists and founders who tend to frequently hop in and out of meetings.

DEEP DIVE

Rolling out AI tools that provide meaningful returns on investment is the number one priority for enterprises today. But building complex systems like AI agents– systems designed to complete specific jobs on their own— is a daunting challenge, one that several businesses are struggling to do successfully, Glean CEO Arvind Jain told Forbes.

For one, these AI agents should be able to access all the disparate data scattered across applications and locked behind software systems. They also require permissions to save their work in those systems without compromising security. And, as better AI models with more advanced capabilities (from conducting in depth research to writing code) get released, businesses end up having to sign contracts with hundreds of different vendors. But even as the models become more powerful, hallucination rates (a measurement of the times AI gets it wrong) have only gotten worse, the New York Times reported. “It requires a lot of engineering work,” Jain said. “Turns out all you’re doing is low level data movement plumbing work to build these agents.”

Jain said his $4.6 billion-valued company Glean, dubbed as a ChatGPT for the workplace, has a fix. The company is releasing a suite of some 30 pre-programmed AI agents that can be immediately put to work on day-to-today office tasks. These systems can automatically resolve IT tickets, help prepare for meetings and write performance reviews (that can be edited). Glean, which counts companies like Grammarly and Zillow among its customers, has already found success in the strategy, despite rivals like Writer closing in. The company said it crossed $100 million in annualized revenue (signed contracts) in February and has raised over $600 million in total venture capital.

More important than ROI, Jain said, is to help educate workers on how to best use AI as reflexively as any other tool. “AI is not an easy technology. It’s very powerful but it’s also odd,” he said. “It’s going to take time for us as humans to actually really figure out how to get the best out of it.”

WEEKLY DEMO

Los Angeles-based AI video avatar startup HeyGen is adding new capabilities to its software that allow users to specify the tone, pacing and style of the voices for their video avatars. It is also launching a so-called “voice mirroring” feature that lets a user upload a short audio clip of them saying something, that can then be replicated (emotion and tonality intact) by the avatar and inform the characteristics of other speech as well. The aim is to give a person more control over how their avatars speak, like a director explaining to an actor how to carry out a performance, CEO Joshua Xu told Forbes. Xu claims the company now has more than 90,000 customers who largely use the tool to produce marketing and explainer videos as well as training materials.

MODEL BEHAVIOR

Local daily newspaper The Chicago Sun-Times printed an AI-generated summer reading guide that listed books that did not exist alongside real authors, 404 Media reported. The writer admitted that they used AI to compile the list. Other parts of the 64-page section called “Heat Index,” which was supported with ads, also appear to be generated with AI and was published in at least one other newspaper, The Philadelphia Inquirer.