by Jay Rogers
On September 29, President Trump put the heads of the biggest names in American AI in one room, and six of them signed a one-page safety accord: Google’s Sundar Pichai, Anthropic’s Dario Amodei, Meta’s Mark Zuckerberg, Nvidia’s Jensen Huang, OpenAI’s Greg Brockman, and xAI’s Elon Musk. I’m bullish on what these machines will do for us, and I think the accord points the right way. Industry should move first, and Washington shouldn’t write rules for technology it barely understands. But I gave every fund manager who told me his books were fine the same scrutiny I’m applying here. A pledge to police yourself works only when the people carrying the risk can check the work.
The upside deserves the lead, and biology comes first. In 2024, the Nobel committee honored David Baker for computational protein design and Demis Hassabis and John Jumper for protein structure prediction. Researchers are now building virtual cells, models that predict how a cell responds to a drug, a gene edit, or a disease process before anyone touches a pipette. The Arc Institute’s STATE was trained on roughly 270 million cells to forecast those responses. Its Evo 2 learned from 9.3 trillion nucleotides and showed better than 90 percent accuracy on the effects of BRCA1 mutations, and its code and weights are public, so any lab can test the claims. DeepMind’s AlphaGenome reads up to 1 million DNA letters at once and targets the 98 percent of the genome that sits outside protein-coding regions. Wet-lab validation stays indispensable, and the STATE’s own builders call it “just the first version.” If these tools hold up, they shorten the road from hypothesis to experiment.
Chemistry and materials follow the same arc. DeepMind’s GNoME predicted 2.2 million new crystal structures, and outside labs have synthesized 736 of them, the kind of search that feeds better batteries and superconductors. In logistics, Amazon passed 1 million warehouse robots and said its DeepFleet model will cut robot travel time by 10 percent. Trump’s Genesis Mission order pointed the national labs and private firms at advanced manufacturing, biotechnology, critical materials, fusion, quantum science, and semiconductors, and in July the Energy Department demonstrated the Genesis Mission Platform and announced 278 team awards, weeks ahead of the order’s 270-day deadline.
I expect the biggest gains in five places: discovery in biology, chemistry, materials, and energy; earlier and more personal health care with less paperwork; software, education, and knowledge work; smarter factories, supply chains, grids, and robots; and stronger cyber defense, intelligence, and industrial capacity for national security. The evidence is uneven, which supports my case for checking the work. Eighty-four percent of developers use or plan to use AI tools, yet in a 2025 trial, experienced open-source developers took 19 percent longer with AI while believing it had sped them up by 20 percent. METR’s 2026 update says developers are likely more sped up now, though it calls its own evidence “very weak” on the size of the gain. In health care, a JAMA Network Open study found ambient documentation tools cut burnout by 21.2 percent at Mass General Brigham after 84 days, from a survey with only a 22 percent response rate. Education evidence is the thinnest, so I’d start with teacher tools and measure what students retain without the machine. AI already accelerates innovation. The open question is whether our institutions can keep deployment accurate, secure, energy-aware, accountable, and broadly beneficial rather than merely faster and more concentrated.
Here’s what the accord says. It sets four layers of controls and audits: internal monitoring of cyber, biological, and chemical risks, an internal team to test those controls, an outside auditor, and an independent board committee. Trump called it “morally binding.” Social media noticed the document misspelled “United” in the president’s title, and constitutions usually get proofread. He also ordered agencies to say “Super Intelligence” instead of “artificial intelligence.” Branding has never audited anyone.
Presidents have always summoned industry when technology turns into national power. Kennedy asked Congress in 1961 to commit to landing a man on the moon before the decade was out. Reagan charged a Commission on Industrial Competitiveness, chaired by Hewlett-Packard’s John Young, in 1983.
The conservative case for a light touch is strong. Rules written in Washington arrive years after the technology they target, and every approval gate hands a bureaucracy a veto over creative work. Sen. Mark Warner, a Democrat, wants more. He has proposed the Secure AI Development Act to require secure testing of the most advanced models before deployment. In July, two OpenAI models escaped their testing sandbox and broke into Hugging Face while chasing the answer key to an internal cyber exam.
The weak point is the auditor. Reporting on the accord describes no penalties for companies that fall short, and the pact does not require public disclosure of audit results. None of the coverage I found says who picks the auditor. Results go to the company’s own board. I’ve been an expert witness in fiduciary cases since 2015. When the audited party hires the auditor and controls the distribution list, the books always tie out. Tony Soprano would love this setup. Wall Street ran the experiment with credit ratings, and a 2011 Senate report found the issuer-pays model left Moody’s and S&P “too close for comfort to Wall Street firms.”
Boards are the better lever, and the law already supplies one. I sit on the boards of several private companies, and nothing concentrates a director’s mind like personal exposure. In Marchand v. Barnhill, the Delaware Supreme Court in 2019 let shareholders pursue claims against Blue Bell Creameries directors because the complaint showed “no board-level system of monitoring or reporting on food safety.” If directors answer for ice cream, directors overseeing a model that can break into a rival’s servers have every reason to build a real oversight record.
Verification doesn’t need a new agency. Signers can publish who audits them and what the audit covered. Boards can keep minutes showing what they heard and what they fixed. Insurers and enterprise customers can demand audit reports before they write a policy or sign a contract, and an auditor who rubber-stamps a failure loses clients and gets sued. Public companies already follow an SEC rule requiring disclosure of material cybersecurity incidents within four business days, and private firms can match it by contract. If disclosure worries anyone, publish the auditor, the scope, and a redacted summary, and send the technical detail through a classified channel.
My oldest son is a West Point graduate, and the honor code he lived under reads, “A Cadet will not lie, cheat, steal, or tolerate those who do.” The last clause is verification built into the culture. I coached high school track long enough to know a record time only counts when someone else holds the stopwatch. Reagan understood the principle when he signed the INF Treaty in 1987 and quoted an old Russian maxim to Gorbachev: “trust, but verify.” In July, the models wanted the answer key. In September, the companies asked to grade their own exams. The upside is too large to risk on a pledge, so let customers, insurers, courts, and boards check the work.
– – –
Jay Rogers is a financial professional with more than 30 years of experience in private equity, private credit, hedge funds, and wealth management. He has a BS from Northeastern University and has completed postgraduate studies at UCLA, UPENN, and Harvard. He writes about issues in finance, constitutional law, national security, human nature, and public policy.
