The next year of AI target analysis
The premier nonprofit partner of US satellite imagery analysis and military targeting hosted the US Gov. product manager of Anthropic on Wednesday for a discussion on the future of generative AI.
Sitting next to the Chief AI Officer of Marine Corps Intelligence, Aaron Tooke, on Wednesday evening at the event center of the Association of the U.S. Army, Steven Sloss threw out a hypothetical: “I wonder if we can’t start to move from seeing a tank three times a year with an exquisite system to being able to track the tank from factory to field because we’ve seen it continuously, we’ve watched it continuously.”
Rather than being stuck with a small number of expensive U.S. Government spy satellites — “exquisite” systems in intelligence lingo — Anthropic’s U.S. Government product manager was suggesting the usage of artificial intelligence to fuse together a stream of imagery from low-cost-to-launch commercial satellites. (This is largely the domain of the sponsor of the event, Planet Labs, whose business development executive moderated the panel giving rise to the statement, and whose director of intelligence sales earlier demoed integration with the maritime intelligence platform SynMax.)
Mr. Tooke similarly noted the use of Automatic Target Recognition (ATR) for identifying whether an object was actually an adversary’s amphibious assault vehicle and proposed training AI “red cells” on top of Chinese and Russian military doctrine to produce more accurate adversaries in war games. The third panelist, Joe Beieler of the University of Maryland’s Applied Research Laboratory for Intelligence and Security (ARLIS), rhetorically asked if generative AI could identify different variants of F-16 and F-35 fighter jets.
Just hours before the start of the event, Google was confirmed to have invested its second billion dollar tranche into Anthropic, with a predicted valuation of $60 billion, versus its competitor OpenAI’s reported $157 billion value.
Despite OpenAI’s pretense of preventing military work through its terms of service, the company has turned a blind eye to such uses when resold through the Azure OpenAI product of its largest investor, Microsoft, including through a contract with U.S. Africa Command (AFRICOM). Anthropic largely differentiated itself from OpenAI through stronger safety branding, though the company announced in November that it would begin reselling its artificial intelligence models to government agencies through Palantir, more than a year after announcing an “enterprise” partnership with Scale AI. (According to his LinkedIn profile, Anthropic’s U.S. Government product manager joined the company in October, after roughly four years at Microsoft and roughtly five-and-a-half years at Palantir.)
The opening joke for the conference, directed at the associate director for capabilities of the U.S. National Geospatial-Intelligence Agency (NGA), Trey Treadwell, was on whether his agency would be applying for funding from OpenAI’s newly announced $500 billion Stargate infrastructure project, following a requisite reference to President Trump rescinding President Biden’s executive order on AI safety hours after taking office on Monday.
Beyond a lengthy explanation of NGA’s ongoing bureaucratic reorganization into Program Executive Offices (PEOs) and a desire to cut between $100 million and $200 million in yearly costs through discovering redundancies using Large Language Models, Treadwell noted that his agency had begun significantly integrating the widely reported Project Maven artificial intelligence effort into its own systems just months earlier, despite significant adoption by the regional combatant commands.
When asked by the author why the NGA had yet to integrate Maven — which transferred to the agency from the Pentagon in January 2023 — Treadwell provided a long list of reasons, including that the original system ran on the Secret Internet Protocol Router Network (SIPRNet), while NGA primarily operates at the Top Secret-level Joint Worldwide Intelligence Communication System (JWICS). Treadwell also stated that NGA had preexisting tools overlapping with some of Maven’s capabilities.
Treadwell also emphasized that Maven has overshadowed the NGA’s automated system for global indications and warnings, going by the mouthful of Analysis Services Production Environment for the National System for Geospatial Intelligence (ASPEN).
The audience was also reminded of the existence of the NGA’s unprecedented $708 million “Sequoia” data labelling request for proposals which went out on September 30. According to the official press release, Sequoia “will support the NGA Maven Program,” as well as “natural language processing, analytic models and AI/ML models that support business process automation for the GEOINT mission.”
Treadwell also noted the NGA’s intent to partner with the Army, Navy, Central Intelligence Agency (CIA), and National Reconnaissance Office (NRO) on data labelling, calling it a “fool’s errand” not to share. And as the author reported last April, Scale AI, an attendee at Wednesday’s NGA event, has referred to its St. Louis, Missouri data labelling office as an “ammo factory” for the “AI wars.”
This genie is not going back in the bottle...Scary people with scary software; imagine all the things that can go wrong, and then multiply by 10. Great reporting.