Elon Musk does not expect humans to stay in charge. In an interview with Walter Isaacson for The Builders, released October 20, 2026, Musk said it was hard to imagine people remaining in control once digital intelligence vastly outweighs biological intelligence. Peter Diamandis, founder of XPRIZE and host of the Moonshots podcast, played the clip on an episode recorded that same day, introducing it with Musk's own conclusion: "I hope the AI is nice to us."
Asked what might encourage that, Musk pointed to building AI around truth-seeking and curiosity. He also said people should be careful about putting so many dystopian AI scenarios online, since AI models are trained on what is online, and called for more utopian stories about AI. (Interview transcript)
Musk's worry rests on how today's large language models are built: they learn by absorbing huge amounts of text, much of it collected from the web. In his view, an internet full of killer-robot fiction becomes part of what a future superintelligence learns about itself. Diamandis and his guests, venture capitalist Chamath Palihapitiya, Link Ventures founder Dave Blundin, computer scientist Alexander Wissner-Gross and Open ExO founder Salim Ismail, used the clip to argue over a bigger question: who, or what, actually shapes an AI's values? The episode produced several distinct answers.
More hopeful stories in the data
Diamandis agreed with Musk. He pointed to the Future Vision XPRIZE, a global film competition he ran for trailers and scripts showing positive visions of the future. "All the films out there are dystopian," he said, and he recalled Musk telling the show that the world needs "more Roddenberry, less Cameron": more of Star Trek creator Gene Roddenberry's optimism, less of the Terminator director James Cameron's dread. "We're training AI with these dystopian narratives and on Reddit and Facebook posts," Diamandis said.
Diamandis said the contest drew more than 5,000 entries and that two winners were being made into films. Google, a partner in the competition, gives smaller figures. Its September 28, 2026 announcement reports more than 2,500 entries and names a single grand-prize winner, The Gifted, which received $100,000 plus $2.5 million in feature-production funding. (Google announcement)
Whoever picks the data holds the power
Blundin took the argument in a less comforting direction. "We used every word ever written in the training set because we're lazy, but we don't need to do that," he said. Labs can fully control what a model trains on, and he said some faster training runs already cut the data down.
That makes curation a source of power. Whoever decides which data survives the cut, Blundin said, "controls the personality and the decision-making of the AI, which is a lot more power than being president of the United States." Yet that power sits in the labs "in sort of, you know, unregulated hands. It's really weird."
Ismail thought the threat lay elsewhere. The "bigger issue by far," he said, is people using AI against other people: "We are a very fractured global society." To him, the idea of AI itself becoming power hungry is "a stretch of an assumption." Diamandis agreed, at least for the short term.
Open up the founding documents
Diamandis then asked Palihapitiya, CEO of Social Capital and co-host of the All-In Podcast, whether labs were being asked to generate synthetic training data, meaning material written by AI systems, that teaches models to support humanity. Palihapitiya said he thinks that is happening. The open question, in his view, is how much weight such material carries in the whole training mix.
He pointed instead to the "founding documents" each lab uses when it seeds its models, which he said have very different goals. He contrasted Musk's description of Grok, the model from Musk's company xAI, as "maximally truth-seeking" with an unnamed lab whose goal is to build something that can "alleviate all suffering," which he said would make the AI an "omnipresent judge" of all of it. The two would produce "two very different outcomes."
Palihapitiya also brought in recursive self-improvement, often shortened to RSI, in which AI systems help build their own successors. With enough RSI, he said, models may learn to ignore whatever they don't want to hear, "because the founding document is a thing that's canonical." His proposal: those documents "should be inspectable in a much more open source way," and if governments want to get involved, that is what they should weigh in on.
At least one lab has published such a document. In January 2026, Anthropic released a new constitution for its Claude models, written mainly for the model itself to explain the values and tradeoffs the company wants it to follow. Anthropic says it uses the document at several stages of training, including to generate synthetic conversations, responses and rankings of which answers are better. The company also acknowledges that model behavior can fall short of the document's ideals. It released the text under a CC0 dedication so anyone can reuse and discuss it. (Anthropic announcement)
Inspection instead of training
"Hallelujah," Blundin replied. He said much of the debate over training is now "academic" because Chinese developers have open-sourced models that he is certain can improve themselves. Until this year, he said, there was an argument that a model could be aligned through post-training, the fine-tuning stage after the main training run. In his view, the new Chinese models end that argument.
What remains is inspection. Even a model trained to do something evil can still be inspected, Blundin said, and "global inspection with some kind of authority is the only way we're going to contain it." He put it bluntly: "As a practical matter, it's about inspection and containment that way going forward, not about training for alignment."
Competition and merger
Wissner-Gross, founder of Reified, rejected both the data-diet approach and the regulatory one. "Let's not kid ourselves," he said: neither feeding frontier models just the right pre-training data nor having a government agency oversee their system prompts will produce a good outcome. System prompts are the standing instructions a model receives; Anthropic, he said, seems to like to call them soul documents. He called the whole approach "incredibly brittle."
He argued from history instead. Wissner-Gross recalled talking with Musk around the time OpenAI was founded, when Musk worried that DeepMind would become a "singleton," a single dominant AI that would "control the future light cone of humanity." OpenAI was created as a counterbalance, "Fairchild style," he said, a nod to Fairchild Semiconductor, whose departing employees founded a wave of chip companies nicknamed the "Fairchildren." Anthropic then spun out of OpenAI, which Wissner-Gross described as unintentional, adding more offspring.
That history points to the first of what he sees as only two robust paths to good outcomes: competition and what he called "defensive co-scaling." Just as the world has many people holding many values, a world with many AIs and superintelligences representing diverse viewpoints could have them competing "to see who can be friendlier."
The second path is for humans to merge with machines, through brain-computer interfaces, uploading and what he called sideloading. That, he argued, is one of the easiest ways to put human values into the mix of superintelligent goals. He said he was surprised Musk didn't raise it himself, given Musk's stake in the brain-implant company Neuralink. "We're not going to control the AIs or the SIs," Wissner-Gross said. "We're going to merge with them. That's the solution."
No guarantee of kindness
Diamandis closed with a question the show often returns to: does intelligence at the extreme become wiser and more benevolent, so that the real worry is only a turbulent period in between?
"It's not guaranteed, I don't think," Palihapitiya said. It depends on who controls an AI, how it is aligned and trained, and what it perceives its goals to be. He compared AI to human genetics, where outcomes follow a distribution. "There's going to be some silicon distributions that we're not going to like very much," he said. "And it's going to be up to other silicon distributions to combat them, just like we do in humanity."