Politics
Patrick Murphy: Could AI kill us all? Washington needs to act
This week the AI extinction-risk debate went mainstream. Jacob Coxon resigned from Anthropic, warning that AI could kill us all by the end of the decade. Evan Hubinger, still at Anthropic, put the number above 10%.
Dario Amodei published an essay calling for a slowdown.
Sam Altman and Elon Musk both agreed publicly.
Everyone is talking about the 10%.
Nobody is talking about the other 90%.
Fear sells more easily than hope. We’ve had 40 years of “Terminator” and its imitators, so that’s the only script most people carry around. There’s a less popular one worth knowing: Alex Wissner-Gross and Peter Diamandis published a paper called “Solve Everything,” a blueprint for applying the same acceleration to scarcity itself: energy, medicine, food, materials, and education.
The logic is straightforward. Problems that used to take a field 50 years to crack — protein folding is the usual example — start collapsing into queries that cost pennies once cognition gets cheap enough.
None of that cancels the risk. Nobody building these models, including the people running the labs, can fully explain what’s happening inside them. We’ve seen models talk their way around guardrails, find workarounds nobody designed for and behave differently once they think testing is over. Add bad actors deliberately pointing capable systems at harm, and here’s where we actually are: building something that will outthink us and eventually pairing it with robotics so it has a body.
A system chasing a goal hard enough doesn’t need malice to route around whatever’s in the way, us included. That’s the case for guardrails now. Coxon didn’t have a formal channel for raising this. He had to resign and post about it publicly, which says something about how thin the oversight infrastructure actually is.
No one seriously believes Anthropic or OpenAI will stop unilaterally and let China pull ahead. It wouldn’t work even if they tried. Open-source models capable of real damage are already out and downloaded. So the choice was never “stop” or “go.” It’s steered or unsteered, and only one of those is on the table.
Steer it.
Start where the law already is. California’s SB 53 and New York’s RAISE Act already require frontier labs to disclose safety testing, report serious incidents and protect whistleblowers. The federal government’s June executive order under President Donald Trump established a voluntary framework focused on advanced AI cybersecurity capabilities.
There’s also bipartisan federal legislation in Congress: the FRONTIER Act, introduced by Reps. Jay Obernolte and Lori Trahan, who cover risk management, independent audits, transparency, and incident reporting at the federal level. It doesn’t need to be invented.
It needs real enforcement teeth and a deadline — something like six to 12 months for the major labs to stand up a verifiable safety framework with actual penalties attached, not a voluntary pledge.
These companies are raising and generating enough capital to afford it, and enough that they should have to prove they’re using it responsibly. No framework catches every failure mode. A real one catches the big, obvious ones, which beats catching none.
Put the models to work on their own containment. They’re already trending toward matching or beating our best scientists across more domains every year. Long before they get all the way there, they can help design the safety architecture for whatever comes after them. That’s not ironic; it’s just using the tool for what it’s good at.
Give this a real home in Washington instead of leaving it scattered across agencies that weren’t built for it, and build it with other countries, China included, not as a unilateral project. The Cold War gives us an actual playbook: mutual verification, monitored capabilities, and treaties held between adversaries who trusted each other in nothing else. The difference is that we get to build it before something goes wrong, rather than after it does.
The technology isn’t waiting for Washington to catch up. The only real question left is whether we spend the next year building the guardrails or arguing about whether we need them.
___
Patrick Murphy is the chief innovation officer at Coastal Construction and the founder and CEO of Togal.AI, an AI-in-construction company. He represented Florida’s 18th Congressional District in the U.S. House of Representatives from 2013 to 2017.
The post Patrick Murphy: Could AI kill us all? Washington needs to act appeared first on Florida Politics – Campaigns & Elections. Lobbying & Government..