Google’s AI architect lived rent-free in Elon Musk’s head

About a week into the Musk v. Altman trial, we’ve heard from some of the most powerful people in tech — including OpenAI president Greg Brockman, Elon Musk’s fixer Jared Birchall, and Musk himself. But one of the most prominent characters is hovering around the margins: Demis Hassabis, CEO of Google DeepMind.

Hassabis is the architect of Google’s in-house AI lab. He founded DeepMind as an independent startup in 2010 and sold it to Google four years later, reportedly for between $400-650 million. Since then, he’s been at the helm of many of Google’s largest AI research breakthroughs, like AlphaFold — and he’s climbed the ladder from there, now leading Google Gemini, the team formerly known as Google Brain, and even for-profit DeepMind spinoff Isomorphic Labs.

From the start, OpenAI was designed to oppose Google. Musk testified that he was inspired to found it by a conversation with Google’s Larry Page, where Page — in his telling — shrugged at the notion of AI wiping out humanity. It’s not surprising Musk and Altman’s circle would be wary of its AI team. But court documents and testimony reveal just how much Google and Hassabis specifically struck fear into their hearts.

During Brockman’s testimony this week, he said Musk talked about Hassabis “many, many times” throughout the early years of OpenAI, calling Musk “very consistent and fixated” on the man. When he attended an AI-focused dinner with Altman and Musk, Brockman added, the first thing he recalled Musk asking was, “Is Demis Hassabis evil?”

A dinner with Demis before OpenAI’s founding was “extremely alarming,” Musk wrote in one email to Brockman and Ilya Sutskever, a fellow OpenAI co-founder. “I feel like they are playing the Super Bowl and we are playing the Puppy Bowl. Unless we want to have our ass handed to us, we need to step up our game dramatically,” he said in 2016.

Hassabis makes one of his first appearances in court documents soon after OpenAI’s founding, as Musk was talking up the new lab’s “open” nature in the press. In January of 2016, Musk forwarded Altman and Sutskever, who was poached from Google, a message Hassabis had sent him. The Google head disagreed with Musk and his cofounders “extolling the virtues of open sourcing AI.” Hassabis wrote that it was “actually very dangerous,” adding, “I presume you realise that this is not some sort of panacea that will somehow magically solve the AI problem?” A handful of months later, OpenAI co-founder (and current company president) Greg Brockman told Musk that Google’s “policy people” wanted to speak to him, for fear OpenAI would “build a public narrative that it’s wrong to have any closed-source AI.” Musk was especially interested in who, specifically, called Brockman from Google.

It was the start of years of competition, and the stakes would only rise from there.

About six months later, Musk started to relay his concerns about beating Google DeepMind in the AI race. He wrote to his Neuralink associates, “Deepmind is moving very fast. I am concerned that OpenAI is not on a path to catch up. Setting it up as a non-profit might, in hindsight, have been the wrong move. Sense of urgency is not as high.”

In September of 2017, Brockman and Sutskever wrote to Musk expressing concern about his control over OpenAI, using Google as an example of precisely what not to do. “You are concerned that Demis could create an AGI dictatorship. So do we. So it is a bad idea to create a structure where you could become a dictator if you chose to.”

By the start of 2018, Musk was seemingly in a full-fledged spiral over Google’s AI influence and the need for OpenAI to overtake the tech giant — and the relative panic had spread to others as well. Musk wrote in a January email exchange that OpenAl was “on a path of certain failure relative to Google. There obviously needs to be immediate and dramatic action or everyone except for Google will be consigned to irrelevance.” He and OpenAI cofounder Andrej Karpathy were so concerned that they suggested folding OpenAI into Tesla so it would be better-resourced.

“It’s unclear if a company could ‘catch up’ to Google scale” without a merger, Karpathy later wrote: “I cannot see anything else that has the potential to reach sustainable Google-scale capital within a decade.”

Shivon Zilis, an OpenAI board member at the time, suggested a direct intervention. Zilis, who now shares four children with Musk, wrote a personal plea asking him to “slow down” Hassabis. “There is a very low probability of a good future if someone doesn’t slow Demis down. Slowing him down is the only nonnegotiable net good action I can see,” Zilis wrote. “I think you know I’m not a malicious person but in this case it feels fundamentally irresponsible to not find a way to slow or alter his path.” Musk responded that they could discuss it that evening over the phone but for the first time seemed dejected about his prospects in his battle against Hassabis, writing, “I doubt I could do so in a meaningful way.”

Zilis continued her personal pleas to Musk to overtake Hassabis, relaying rumors from Altman and others. “On top of the folks that secretly converse on Twitter DM because they don’t trust Demis not to spy on their email and gchat, a part of the inner group also meets in a London coffee shop without cell phones to have in person discussions away from him,” she wrote.

By November of that year, Musk wrote in an email that he had fully “lost confidence” that OpenAI could “serve as an effective counterweight” to beat Hassabis and DeepMind, and that he was planning to do so via Tesla instead. “We have cash flow on the order of billions of dollars per year to build hardware that hopefully has at least a dark horse chance to keep Google honest,” he wrote. “My probability assessment of OpenAl being relevant to DeepMind/Google without a dramatic change in execution and resources is 0%,” Musk wrote a few weeks later. “Unfortunately, humanity’s future is in the hands of Demis … And they are doing a lot more than this.”

Three months later, in March 2019, the last mention of Hassabis in the trial exhibits so far comes from a mysterious message sent by Altman to Musk with no further details available.

“Have some mild Demis updates to share,” Altman wrote. Musk agreed to talk it over on the phone.