Former Google CEO Eric Schmidt sounds the alarm over a ‘Manhattan Project' for superintelligent AI
Former Google CEO Eric Schmidt, Scale AI CEO Alexandr Wang, and Center for AI Safety Director Dan Hendrycks are sounding the alarm about the global race to build superintelligent AI.
In a new paper titled Superintelligence Strategy, Schmidt and his co-authors argue that the U.S. should not pursue the development of artificial general intelligence (AGI) through a government-backed, Manhattan Project-style push.
The fear is that a high-stakes race to build superintelligent AI could lead to dangerous global conflicts between the superpowers, much like the nuclear arms race.
"The Manhattan Project assumes that rivals will acquiesce to an enduring imbalance or omnicide rather than move to prevent it," the co-authors wrote. "What begins as a push for a superweapon and global control risks prompting hostile countermeasures and escalating tensions, thereby undermining the very stability the strategy purports to secure."
The paper comes as U.S. policymakers consider a large-scale, state-funded AI project to compete with China's AI efforts.
Last year, a U.S. congressional commission proposed a 'Manhattan Project-style' effort to fund the development of AI systems with superhuman intelligence, modeled after America's atomic bomb program in the 1940s.
Since then, the Trump administration has announced a $500 billion investment in AI infrastructure, called the "Stargate Project," and rolled back AI regulations brought in by the previous administration.
Earlier this month, U.S. Secretary of Energy Chris Wright also appeared to promote the idea by saying the country was "at the start of a new Manhattan Project" and that, with President Trump's leadership, "the United States will win the global AI race."
The authors argue that AI development should be handled with extreme caution, not in a race to out-compete global rivals.
The paper lays out the risks of approaching AI development as an all-or-nothing battle for dominance.
Schmidt and his co-authors argue that instead of a high-stakes race, AI should be developed through broadly distributed research with collaboration across governments, private companies, and academia. They emphasize that transparency and international cooperation are critical to ensuring that AI benefits humanity rather than becoming an uncontrollable force.
Schmidt has addressed the threats posed by a global AI race before. In a January Washington Post op-ed, Schmidt called for the US to invest in open source AI efforts to combat China's DeepSeek.
The authors suggest a new concept—Mutual Assured AI Malfunction (MAIM)—modeled on the nuclear arms race's Mutually Assured Destruction (MAD).
"Just as nations once developed nuclear strategies to secure their survival, we now need a coherent superintelligence strategy to navigate a new period of transformative change," the authors wrote.
"We introduce the concept of Mutual Assured AI Malfunction (MAIM): a deterrence regime resembling nuclear mutual assured destruction (MAD) where any state's aggressive bid for unilateral AI dominance is met with preventive sabotage by rivals," they said.
The paper also suggests countries engage in nonproliferation and deterrence, much like they do with nuclear weapons.
"Taken together, the three-part framework of deterrence, nonproliferation, and competitiveness outlines a robust strategy to superintelligence in the years ahead," they said.
This story was originally featured on Fortune.com
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Yahoo
2 hours ago
- Yahoo
Google Finds Workaround for Lobbying Rules That Omits Big Bosses
(Bloomberg) — It was the end of 2018, and Google's leaders were tired of being Number One. For the second year in a row, federal records showed the search giant had spent more than any other individual company on lobbying in Washington. Executives in Mountain View were sick of seeing that mentioned in the press. Sunseeking Germans Face Swiss Backlash Over Alpine Holiday Congestion The US-Canadian Road Safety Gap Is Getting Wider To Head Off Severe Storm Surges, Nova Scotia Invests in 'Living Shorelines' Five Years After Black Lives Matter, Brussels' Colonial Statues Remain For Homeless Cyclists, Bikes Bring an Escape From the Streets Then Google apparently found a workaround. A new analysis of federal lobbying data by the nonprofit Tech Transparency Project shows that Google and its parent company, Alphabet Inc. used an internal reorganization to exclude the value of lobbying by its senior executives from disclosures. The move helped keep Google off the top of the lobbying charts even as it maintained a robust network of advocates pushing its interests in the capital, during federal challenges to its dominance in search and advertising and the beginnings of artificial intelligence regulation. The findings, which were confirmed by a Bloomberg analysis of lobbying records, show that the effect of the accounting change was to lower the amount that Google reported spending to influence the federal government, likely by millions of dollars. The reorganization 'has allowed the company to shield a significant portion of its lobbying expenditures from public view,' the Tech Transparency Project said in its report. A Google spokesperson, José Castañeda, disputed the report and said the company has followed all relevant disclosure laws. 'These are inaccurate claims about a technical change that simply brought us in line with how many other companies report their lobbying activities,' he said. 'Our lobbying expenditures began decreasing in 2018, after we restructured our government affairs team and cut spending on consultants.' Internal Reshuffle Starting in 2019, Google began cutting ties with some of its external lobbying firms, a move it acknowledged publicly as part of an overhaul of its Washington operations. But the shuffling of external lobbying firms doesn't explain the whole of the decline in Google's reported lobbying expenses, which fell from more than $22 million in 2018 to $8.9 million in the Covid-disrupted year of 2020, and have subsequently remained well below pre-pandemic levels. There's been another, quieter change: in early 2020, Google moved its in-house lobbyists into a new subsidiary, called Google Client Services LLC. It's that unit which now files spending disclosures for Google's lobbying activities. The reorganization meant that the parent companies Google and Alphabet no longer directly employed any lobbyists – defined under federal disclosure law as people spending at least 20% of their time on influencing Congress or the executive branch. Companies that file lobbying disclosure reports are supposed to also account for the time that other senior executives — those who don't meet the 20% threshold – devote to lobbying, according to legal experts and the compliance guide for the Lobbying Disclosure Act published by Congressional leaders. That generally involves prorating their annual compensation to account for the days they spend influencing the government. But since Google moved lobbyists into the Google Client Services subsidiary, the parent company no longer meets the threshold for filing disclosures under the Lobbying Disclosure Act, according to the TTP analysis. That means Google no longer reports the lobbying expenses of high-ranking managers who aren't part of the Client Services unit — like Chief Executive Officer Sundar Pichai and chief legal officer Kent Walker — to the public, as it once did. As a result, in 2020 Google dropped out of the top 20 in corporate lobbying expenses for the first time in nearly a decade, the TTP analysis found. While Google's reported annual spending has since edged back up again, it hasn't come close to the No.1 slot in the company lobbying rankings that it used to occupy. For the past five years, that position has alternated between two other tech giants: Meta Platforms Inc. and Inc. Antitrust Challenge There's been plenty going in in Washington over the period that was crucial for Google's business. For one thing, the company — like many peers — is betting heavily on AI, a field where decisions in the US capital will shape the commercial landscape. Google has also been under assault from antitrust authorities over its dominance in search and digital advertising. The company has maintained in those lawsuits that its success is down to consumer choice and superior innovation, rather than a result of its power to shape laws and regulations. Publicity around its lobbying spending has the potential to undercut such arguments and alienate regulators. When executives are as highly paid as many in Silicon Valley, the prorated amounts can add up to millions — even for just a few days' worth of lobbying. Google reported total compensation for Pichai of more than $225 million in 2022, thanks to grants of stock. His total compensation was $10.7 million in 2024. Walker's total compensation was more than $30 million last year, the company reported. Some say the new structure Google is employing flouts the spirit of the federal disclosure law – if not the letter itself. 'This is just too cute by half,' said William Luneburg, a professor emeritus at the University of Pittsburgh School of Law, and the co-editor of the manual for lobbying compliance published by the American Bar Association. 'On the face of it, it's wrong,' he said. 'They have to report all of their expenses, which would include the time of officers and directors and other employees that spend their time engaging in lobbying activity.' 'We always comply with disclosure laws and any suggestion of improper reporting is false,' said Castañeda, the Google spokesperson. TTP said it examined lobbying disclosures of several other companies that filed reports via a similar subsidiary model, but didn't find any that had used the structure to remove executive lobbying from their disclosures. —With assistance from Davey Alba and Sarah Frier. Americans Are Getting Priced Out of Homeownership at Record Rates Dubai's Housing Boom Is Stoking Fears of Another Crash Bessent on Tariffs, Deficits and Embracing Trump's Economic Plan Why It's Actually a Good Time to Buy a House, According to a Zillow Economist A $340 Million New York Office Makeover Is Converting Boardrooms to Bedrooms ©2025 Bloomberg L.P. Error in retrieving data Sign in to access your portfolio Error in retrieving data Error in retrieving data Error in retrieving data Error in retrieving data
CNET
2 hours ago
- CNET
ChatGPT-5 Gives You Real Choices After All. Here's a Quick Breakdown
The biggest pushback after OpenAI announced its new GPT-5 model for ChatGPT came from devotees of older models who felt the new generative AI chatbot lacked the panache of its predecessors. Now you have more choices of pre-GPT-5 models (although you'll have to hunt for some of them) and better control over which components of GPT-5 handle your questions. OpenAI is still sorting through a somewhat rocky launch of GPT-5, led by complaints about the lack of model choices. The model has been anticipated for more than two years and comes as competitors like Anthropic and Google have released powerful new versions of their AI models this year. (Disclosure: Ziff Davis, CNET's parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.) OpenAI planned for one model that could handle everything: GPT-5 includes two different modes, one fast and lean for simple tasks and one aimed at reasoning for complicated ones. A routing program would decide which model handled a given prompt. That's still the default in ChatGPT, but it's not your only option. Here's a look at the menu: Watch this: ChatGPT Users Want the Old Models Back, Intel CEO Goes to the White House & More | Tech Today 03:03 Choices of GPT-5 models There are a few different modes of GPT-5 you can select between if you want to use OpenAI's newest technology. Here's a quick rundown: Enlarge Image This is what your choices will look like if you don't enable legacy models in settings. Screenshot by Jon Reed/CNET Auto: This mode allows the switching program built into GPT-5 to decide whether your query is handled by a lighter, faster large language model or a bigger, slower reasoning model. OpenAI CEO Sam Altman posted on X that this will be the best fit for most people. This mode allows the switching program built into GPT-5 to decide whether your query is handled by a lighter, faster large language model or a bigger, slower reasoning model. OpenAI CEO Sam Altman posted on X that this will be the best fit for most people. Fast: Your query will go straight to the fastest, lightest model. Expect quick, basic answers, but not the in-depth research you'd get from a reasoning model. Your query will go straight to the fastest, lightest model. Expect quick, basic answers, but not the in-depth research you'd get from a reasoning model. Thinking: This is a reasoning model, meaning it'll try to answer your question over several steps. It might use web searches and other tools, or it might go back and redo its past steps to try to get the right answer. There are some limits on how much you can use this model (3,000 messages per week right now). This is a reasoning model, meaning it'll try to answer your question over several steps. It might use web searches and other tools, or it might go back and redo its past steps to try to get the right answer. There are some limits on how much you can use this model (3,000 messages per week right now). Thinking mini: This is a lighter, faster version of the Thinking model. If you hit the usage limit on Thinking, you'll be stuck with Thinking mini. You can access it on your own if you enable more model options (more on that below). This is a lighter, faster version of the Thinking model. If you hit the usage limit on Thinking, you'll be stuck with Thinking mini. You can access it on your own if you enable more model options (more on that below). Pro: The most powerful reasoning model in the fleet, this is only available at the moment for Pro users, who pay $200 per month. (Although Altman did tease giving Plus users, who pay $20 a month, a very limited number of queries as a taste.) Again, sticking with Auto is probably easiest for most users. Think of it like driving a car with an automatic transmission instead of having to change gears manually. Sure, people who are really into cars might prefer the stick shift, but most people should probably just let the machine handle it. How to get the older OpenAI models Everyone with a paid ChatGPT subscription can access the older GPT-4o model directly in the same menu where you can choose your flavor of GPT-5. This model received the most clamor from users after it was removed, and Altman said if OpenAI ever decides to take it away permanently, "we will give plenty of notice." But 4o isn't your only choice (if you're a paid user). You just have to know where to look. To access GPT-4.1, 4o-mini and 3o, along with GPT-5 Thinking mini, you'll have to go into your Settings and toggle on "Show additional models."
CNBC
2 hours ago
- CNBC
Elon Musk's xAI loses co-founder Igor Babuschkin, who's leaving to start venture firm
Igor Babuschkin, a founding member of Elon Musk's xAI, said Wednesday that he's leaving the artificial intelligence startup to launch his own venture firm. "Today was my last day at xAI, the company that I helped start with Elon Musk in 2023," Babuschkin wrote on X, which is owned by xAI. "I still remember the day I first met Elon, we talked for hours about AI and what the future might hold. We both felt that a new AI company with a different kind of mission was needed. Building AI that advances humanity has been my lifelong dream." A former research engineer for Google's DeepMind and ex-member of OpenAI's technical staff, Babuschkin recounted some of xAI's major operational achievements during his tenure, including building out engineering teams at the company. "Through blood sweat and tears, our team's blistering velocity built the Memphis supercluster, and shipped frontier models faster than any company in history," he wrote. The facility in Memphis processes data and trains the models that power xAI's Grok chatbot. Musk wrote, in response, "Thanks for helping build @xAI! We wouldn't be here without you." Babuschkin said he's starting Babuschkin Ventures to support AI safety research and invest in startups in "AI and agentic systems that advance humanity and unlock the mysteries of our universe."
