The Collection

Marcus agrees with Senator Josh Hawley that AI companies should be liable for damage they cause. He wrote as much in Taming Silicon Valley. The quarrel is with what comes next. On the technolibertarian right, liability is being sold as if it cancelled the need for rules.

Joe Lonsdale, amplified by David Sacks, says companies should pay for harm and that regulation is the wrong answer. The first half is true. The second does not follow. Cuban’s reply to Sacks asked who writes the laws that make those suits work: new AI-specific statutes, or old statutes stretched until they tear. Without that writing, the courthouse stays a stage set.

Without the right regulation, liability is pointless.

Marcus has been here before. In May 2023, Hawley asked him in Senate testimony whether a private right of action alone might suffice: open the courthouse doors, let the harmed sue, let class actions do the rest. Marcus said no. Existing law was written before this stack. Copyright was unclear. Wholesale misinformation was not the same as a single false claim. Section 230’s reach was speculative. Litigation can take a decade. The threat of a suit is a tool. It is not a coverage map.

Hawley has since moved. His bipartisan AI framework, paired with this week’s tweet, champions both liability and regulation. Investors may prefer a world where their portfolio faces only plaintiffs. The rest of us should not confuse an open door with a finished statute.

Jacob Coxon resigned from Anthropic after three years of pretraining work at OpenAI and Anthropic. “Neither company is acting responsibly,” he wrote. “The people building AI earnestly believe that it could kill us all by the end of the decade.” He was not the first. William Saunders told a Senate committee in 2024 he had lost faith that OpenAI alone would make responsible AGI decisions. Mrinank Sharma left Anthropic in February saying the world was in peril. Coxon’s message was the one that broke through.

Views passed a hundred and seventy million. CNN, CBS, and Fox booked him. Sheryl Crow mentioned the resignation on Instagram. TMZ asked Matt Damon about human extinction. Joe Rogan interviewed Daniel Kokotajlo. PauseAI’s Maxime Fournes titled a LessWrong post “Crunch Time is Now. Drop Everything.”

The House started a seven-week recess.

What channels the attention is still unclear. A global pause treaty is unlikely soon. Trump is due to meet Chinese officials with AI safety on the agenda; recent comments from both sides do not read like a deal. At home, AI risk is suddenly a national political topic. The House nonetheless began a seven-week recess. Understanding AI judges it unlikely that any AI legislation passes before the new year. No one will force the labs to change how they do business in the next few months. Some leaders signal willingness to slow if rivals do. That is a voluntary bargain, not a statute. The tweet did the work of a headline. The calendar still has to do the work of a bill.

Early this year a Cloudflare salesperson asked the CIO for API keys. The seller had used AI to build what they called a SuperApp for go-to-market. They wanted production access to about a dozen systems of record and admin rights on a deployment pipeline. They were the first of many. The company had to equip them and keep customer data safe.

Cloudflare’s answer was one internal workspace, Cloudflare OS, for engineers and sellers alike: a shared library of skill files with company context, a portal that connects agents to internal systems scoped to the permissions the person already has, and a gateway that logs every model call and gates models by role. In one month the sales team saved an estimated ten thousand hours on territory planning and proposals. Staff built more than four thousand small apps.

Own the harness. Rent the model.

Jax’s frame: the harness is where speed, reliability, and cost get decided for the whole company. Skills, context, tools, deterministic gates, and evals. Done well, they compound. Vercel runs more than a hundred agents in one monorepo, observed the same way regardless of owning team. Anthropic field marketer Adam Ward’s Monday digest invented event URLs when the sheet was blank; he wrote a hard rule never to invent a URL, and by week’s end the prompt held nine rules, each from seller feedback. Writer’s controlled study of twenty-two tasks across six models with an orchestration layer cut cost per task by forty-one percent. The scarce good used to be intelligence. Now it is reliability, speed, and cost, held in the workspace you own while the weights stay on someone else’s meter.

Semafor’s business desk opened with Detroit in March 2020: GM, Ford, and Fiat Chrysler wanted a coordinated plant halt over Covid. Lawyers warned that talking about it could violate antitrust. The plants went dark anyway. Today the proposed conference call is among frontier labs, and Amodei has asked for a narrow antitrust waiver so they can discuss a coordinated slowdown. Critics call it doom as drawbridge. Semafor calls an AI OPEC unlikely. Zuckerberg wants no part of one. Competition among rich executives is hard-coded. Any cartel-like behavior, the desk writes, will look more like airlines competing on reliability and bundles while the underlying product commoditizes. Airlines do not compete on safety; crashes are bad for everyone.

AI Governance Lead tracked how Amodei’s “pace the frontier” essay actually landed. Sam Altman and Elon Musk endorsed slower capability advances and independent evaluators. The White House dismissed the warnings as overblown and doubled down on beating China. Beijing called the framing fearmongering and kept accelerating. Jensen Huang and Mark Zuckerberg rejected coordinated slowdowns outright. The highest-profile safety intervention of the year clarified the dominant stance in both capitals: competition first.

Buyer power is becoming a primary governance mechanism.

High-stakes buyers are voting with workloads. The Pentagon had already moved roughly ninety percent of classified workloads off Anthropic after an earlier fallout. Nvidia, Palantir, and Booz Allen are reported diverting from Anthropic, and in some cases OpenAI, over data-retention and IP concerns while promoting sovereign and open-weight alternatives. Salesforce deepened Claude ties. Novo Nordisk signed a major Claude drug-discovery deal. European firms keep expanding use. OpenAI launched a misalignment disclosure framework and released six new incidents; Semafor notes each disclosure confirms how little control they have, which invites more regulation and pushes some users toward open weights. Palantir’s Alex Karp told CNBC he questions whether either company will IPO given the liability, and floated nationalization. Canada and Germany funded Yoshua Bengio’s LawZero. The pause talk did not converge the tracks. It advertised how far apart they already were.