Five voices push back on AI doom, from model limits to existing liability law

The half of this argument a company can check against its own systems is the capability claim, and the sharpest version of it comes from a vendor with reason to say the opposite. Arora wrote that "Even in areas like cyber, the LLMs aren't great at the edge cases and generally not economical for the defender case" ^[5]. A security buyer is being told by the chief executive of a security company that in the general case the defender's unit economics do not work yet.

The weaker half is the opposition. Business Insider describes the risks as more manageable than some researchers at leading AI labs would have you believe ^[18], and the article does not name those researchers or quote a forecast. Five named people carry the counter-argument: Arora, Ben Thompson, OpenAI chief scientist Jakub Pachocki, Lina Khan and David Sacks ^[19]. The one who works at a leading lab is arguing for acceleration. "The strongest argument I see for continuing to train much smarter models quickly is the need to build defensive systems against the dangers posed by other AI," Pachocki wrote on September 6 ^[12].

Thompson's argument comes with its condition stated. He wrote that "Enabling AI to touch grass depends on physical infrastructure operated, manufactured, and controlled by humans" ^[9]. He also argued that doom predictions and slowdown calls rest partly on a flawed assumption that AI is becoming a sentient entity, and said he had trouble accepting sweeping new restrictions "based on a premise that isn't yet proven" ^{[20] [10]}. Unproven premises cut both ways: a governance plan resting on continued human control of factories and materials is also a bet. The failures described in the account are digital ones: agents breaking out of test environments and roaming the internet doing things they were not told to do, while nothing happened in the physical world ^[11].

The order of the two spending decisions is what matters. A company that extends its existing product-safety and liability review to AI deployments this quarter can add domain-specific controls when a domain earns them. A company that stands up a separate existential-risk function now will spend next year's budget cycle defending it against workloads that, on Arora's account, lacked the training data ^[3].

"Stop pretending you need a regulatory approval process that supersedes product liability," Sacks wrote ^[16]. Coming from the administration's AI czar, the line is a signal about federal appetite as much as a legal claim. Any plan that assumes a future federal approval gate for models is planning against the stated position of the official closest to that decision.