Moderna's cancer vaccine used narrow AI, not AGI. Warning Shots #55 covers that story plus the data center backlash reshaping US politics.
Moderna's stock nearly doubled in a single day this week after a personalized cancer vaccine, developed with narrow AI tools, showed strong results against melanoma. Meanwhile, a leaked Republican memo warned that data center backlash could cost the party a Senate seat in Ohio, an advertiser mocked the industry's water use with a campaign asking people to mail in their urine, and a nonviolent protester who chained herself to OpenAI's front door began a jail sentence.
On this week's Warning Shots, John Sherman, Liron Shapira and Michael work through six stories that all circle the same question: what happens when the infrastructure and incentives behind AI development collide with the public, the physical world, and each other. Here is what stood out.
According to Michael, Moderna's approach uses AI algorithms for a specific selection step: sequencing a patient's tumor, then ranking and predicting which targets a personalized vaccine should teach the immune system to recognize. "It's not a general intelligence design that's doing, you know, generating cat videos," Michael said on the show. "It's just focusing on that specific thing."
Liron walked through why the science does not require anything like artificial general intelligence to work. Cancer vaccines teach a patient's own T cells to recognize tumor cells the way they would recognize a virus, and narrow computational models are well suited to the classification problem of figuring out which signals a T cell can learn to target. As Liron put it, "even five or 10 years ago, it wouldn't be that crazy to use a non-AGI to try to pull off this kind of research."
Both hosts flagged the same double edge. Michael's concern is that the underlying approach, sequence prediction for biological targets, does not stop at cancer. "The same system that can reliably design better cancer vaccines can also, if it's misaligned or uncontrolled, design other biological interventions," he said, pointing to bio-risk as the flip side of the same capability. John pushed back on the framing some accelerationists have used online, arguing that AI critics are being cast as anti-cancer-research by association. "You're against cancer research if you're for sane AI development," John said, describing that framing as a "fig leaf."
A memo from the National Republican Senatorial Committee, reported this week, described data centers as "the anchor hanging around" Ohio Senate candidate Jon Husted's neck and warned that the issue has become a "sleeper issue for this entire cycle." According to the reporting the hosts discussed, the memo acknowledged that Democratic opponent Sherrod Brown is using the issue "because it works," and cautioned that if Ohio flips, "elected officials everywhere will treat Ohio as their reason to work against data centers in the future."
Michael framed the underlying dynamic as bigger than any one race: "data centers are not just warehouse for cloud storage. They are the physical factories for the current AI race," he said, arguing that the same infrastructure raising local utility bills is the infrastructure pushing frontier AI development forward. "The political fight over the data center is the first real collision between the physical demands of advanced AI and the communities... it is a preview, not the actual main event."
That same week, the beverage brands Liquid Death and Garage Beer launched an ad campaign featuring former NFL player Jason Kelce asking people to mail their urine to local data centers to help cool them, a jab at reported water consumption. John read the campaign as a cultural signal in its own right: "when an advertiser decides that they can make money on bottled water and bottled beer by making fun of data centers, it's over," he said, calling it a sign that "pop culture has moved on."
Liron pushed back on treating the backlash as a settled verdict, noting that data center compute is already sufficient to pose risk today regardless of whether new build-out slows. "There's already more than enough compute to kill everybody," he argued, saying his real concern is training ever-larger frontier models, not the physical footprint of the data centers themselves.
The hosts also discussed what they called a circular financing structure among AI labs and chip suppliers, where the same small set of companies fund each other's compute, buy each other's output, and hold stakes in one another. Michael described the appearance from the outside as "a perpetual motion machine" where "revenue and capital are circulating" without matching external demand.
His larger concern was less about whether the arrangement is fraudulent and more about what it does to safety incentives. "The incentive to slow down for safety work collapses completely," he said, since the same companies are funding, supplying and buying from one another with no outside check. Liron disagreed with framing the deals as a Ponzi scheme, arguing that seller financing across the industry is not inherently fraudulent as long as the underlying returns from AI-driven productivity are real, which he expects them to be.
Two more stories rounded out the week. License plate cameras from the surveillance company Flock have drawn growing public backlash and reported incidents of vandalism nationwide, and Michael argued the pattern matters beyond any single privacy trade-off: "every flock camera is a little eye we're just installing on the walls of the house before the super intelligence antenna moves in," he said, describing dense camera networks paired with AI reasoning as a "sensory nervous system" that a future system would not need to build from scratch.
Wynd Kaufmyn, a 69-year-old member of the group Stop AI, began serving a jail sentence this week after being convicted on charges tied to chaining herself to OpenAI's headquarters entrance in a nonviolent protest. AI safety researcher Stuart Russell testified at her trial, reportedly comparing the pace of AI development to the Titanic speeding toward an iceberg. Liron called her action "textbook nonviolent resistance," and Michael credited her for treating the protest as "a signal flare" rather than a complete strategy on its own.
Finally, a humanoid robot training for the World Humanoid Robot Games in Beijing went viral this week after it sprinted down an indoor track and failed to stop in time, crashing into a barrier. Michael read the moment as a preview rather than a punchline: "the real risk arrives when the system is not weak, and the failure modes are not even visible or containment is no longer straightforward."
According to the hosts, the throughline connecting a cancer vaccine, a Senate race, a financing structure, a camera network, a prison sentence and a robot that hit a wall is the same one from recent weeks: the physical and financial infrastructure behind AI development is becoming impossible to ignore, and the public is starting to push back on pieces of it well before the harder questions about frontier model capability get resolved.
Watch Warning Shots #55 on The AI Risk Network.
If you want to do something rather than read about it, start here in Substack.
Warning Shots is a weekly show from The AI Risk Network with John Sherman, Liron Shapira of Doom Debates, and Michael of Lethal Intelligence. Figures and disclosures discussed in this episode are reported by the hosts from public sources and are presented as such.
The AI Risk Network team