- | 8:42 am
‘Deep Research’ shows how Google can win the AI race
I took Google’s new Gemini research agent for a spin and came away very impressed.
Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week here.
Google flexes with its new “Deep Research” AI tool
Not too long ago, many of us were perplexed by Google’s slow start in the generative AI boom. Well, not any more! With its work on the Gemini LLMs and new Gemini-powered services, Google has caught up—and in some cases, surpassed—the competition. Google’s success is owed in large part to a willingness to play to its own strengths. Gemini Deep Research is a good example of that strategy.
Deep Research, which is available to Gemini Advanced users, is something like an AI version of a human research assistant. It works best for harder, multifaceted questions that require explanation. For instance, I prompted with this question: What are the challenges of creating humanoid robots that can do tasks that they’ve not encountered in their training?
First, Deep Research restated the question as, “Find research papers and articles on the topic of ‘zero-shot learning’ in robotics,” based on a cursory search of relevant research on the web. Then, it generated a multistep research plan, which was a listing of the key aspects of the research question, along with the places the information could be found.
After I agreed to the plan, the agent got busy. “I’m on it. I’ll let you know when your research is done. In the meantime, you can leave this chat.” (There’s something very satisfying about watching an AI agent do my work while I drink coffee.) I watched as it raced over the internet and began compiling a list of sources. About three minutes later it had compiled a 60-item list of source articles and publications, including research papers, journal articles, Medium posts, and Reddit discussions. From all these sources, the agent synthesized a 2,100-word, citation-filled essay that answered my question. Impressive.
Finding all those sources requires a deep web index, and nobody does web indexing better than Google. Selecting and compiling the most relevant chunks of information from each article requires the Gemini 1.5 Pro model that powers Deep Research to hold a lot of information in its memory at once. The Gemini 1.5 Pro has a one-million-token (more than 770,000 words) context window—larger than any other public-facing LLM. Synthesizing all the relevant data into a complete answer requires some planning and reasoning. Gemini 1.5 Pro is one of a small group of LLMs that score higher than 85% on the MMLU benchmark, which evaluates the performance of machine learning models. And you can output Deep Research answers directly to a Google Doc, which is nice.
I still believe Google is in the best position to be the leader in consumer AI over the long term. Using the people, experience, and data it has built up over its years may give it the edge.
Marc Andreessen’s wild story about talking AI with Biden staffers
During a recent episode of the Honestly podcast with Bari Weiss of The Free Press,
billionaire VC Marc Andreessen recounted a meeting he’d had last spring with Biden administration staff to discuss AI. He says the meeting was “horrifying,” and marked the moment he decided to throw his support behind Donald Trump in the presidential election. Andreessen says the Biden staffers told him they intended to exert tight control over the leading AI companies. “They said this is not going to be a startup thing, he recalled. “They actually said flat out to us, ‘Don’t fund AI startups, that’s not something that we’re gonna allow to happen. They’re not gonna be allowed to exist.’”
According to Andreessen, the unnamed Biden staffers said the AI industry would consist of “two or three big companies” that would work closely with the government. “‘We’re gonna basically wrap them in a . . . government cocoon,’” Andreessen told Weiss. “‘We’re gonna protect them from competition. We’re gonna control them, and we’re gonna dictate what they do.’”
That strategy would mark a dramatic shift in the White House’s posture. Biden’s executive order on AI called on AI companies to set and follow safety testing and transparency standards, but was voluntary and nonbinding. Did the Biden staffers know something about researchers’ progress toward superintelligence that the public doesn’t yet know?
Or, could Andreessen have “misinterpreted” the Biden staffers’ words and intentions? It’s a known fact that Andreessen is no fan of tech regulations. In an essay he wrote in 2023, titled, “Why AI will save the world,” he warned against government interference in the budding AI industry. And it’s a known fact that his firm, Andreessen Horowitz, was one of the loudest opponents of California’s ill-fated AI bill last fall.
With a dysfunctional Congress and David Sacks tapped to be the Trump administration’s “AI czar,” Andreessen probably has nothing to worry about. The U.S. likely won’t see anything like the EU’s AI Act for a long time (if ever).
Get another look at our AI 20 series
Our second annual AI 20 series was in full swing, spotlighting the most interesting technologists, entrepreneurs, corporate leaders, and creative thinkers shaping the world of artificial intelligence. Among those profiled is Thrive Capital’s Vince Hankes, who has raised the New York-based VC’s visibility over the past two years by making some large bets on AI companies including OpenAI.
Hankes and Thrive founder Josh Kushner (brother of Jared Kushner) initiated and managed the firm’s relationship with OpenAI. It was a good match. Investing in frontier model developers is a high-stakes game, and the payoff may take up to a decade to mature.
“[We] don’t do many things, but when we do something, we’re getting excited about an opportunity [and] we really double, triple, quadruple down in terms of our time trying to understand it very deeply,” Hankes said.
You can find profiles of this year’s AI 20 honorees here.
