The AI Manhattan Project

AI systems are rapidly being pulled into military and domestic policing workflows—less as “assistants” and more as target-generation engines—while major tech firms quietly loosen restrictions on weapons use. The central concern is that once AI helps compile and prioritize who to harm, human oversight becomes too late or too limited to prevent harm, especially when decisions are automated and scaled.

A shift in Silicon Valley’s priorities frames the change. Early branding around “do no evil” has faded, replaced by a focus on building tools for the American security state and the defense industry. Generative AI may have disappointed in everyday life, but military AI has become the new funding magnet. Venture capital and startup accelerators are backing weapons-related companies, and large platforms—including OpenAI, Meta, and Google—have moved toward defense work while removing or softening policy language that previously barred military use.

The transcript highlights how this pivot shows up in concrete partnerships and personnel flows. Y Combinator backed a weapons manufacturer, and a major VC reportedly promised $500 million for defense technology. In parallel, senior executives and engineers are being integrated into military structures—described as AI leaders being sworn in as lieutenant colonels—signaling that the industry’s talent pipeline is feeding defense priorities.

Meta’s Llama is presented as an example of AI being made available to U.S. government agencies, including defense and national security users, with downstream distribution to top contractors such as Lockheed Martin and Anduril, and to firms providing related services like Oracle and Palantir. OpenAI is described as working on anti-drone AI with Anduril and producing an “AI action plan” while recruiting from military and intelligence circles, including the CIA, NSA, Pentagon, and special operations.

Palantir is treated as the most explicit case. The company’s software is described as data integration and decision automation: taking years of surveillance and disparate records, connecting them into profiles, and producing lists that guide actions. In immigration enforcement, that means assembling dossiers from border entries, visas, addresses, tax records, social media, relationships, and law enforcement history to generate targeting lists for ICE. In war, the same approach is described as fusing spyware, surveillance footage, and facial recognition to produce kill lists that can be executed by drones or other systems.

The transcript argues that accuracy concerns don’t stop deployment because the strategy is to maximize the number of potential targets and then accept collateral damage. It cites reporting about Israeli systems that can wait for specific vehicles to trigger lethal action with minimal human proximity, and it references “Where’s Daddy” software used to identify when suspected fighters return home for bombing during sleep—while noting uncertainty about Palantir’s direct role. A quoted remark from a journalist interview is used to emphasize the operational logic: enter hundreds of targets and wait to see who can be killed.

The broader warning is that AI-driven violence dehumanizes both targets and operators, turning war and policing into spreadsheet-driven automation. With defense budgets, lucrative contracts, and political alignment described as mutually reinforcing, the transcript concludes that resistance is difficult—especially as workers who object are reportedly fired, protesters are arrested, and surveillance infrastructure expands alongside lethal systems.