Trump Bans Anthropic as OpenAI Signs Pentagon Deal [Model Behavior]

[00:00] Announcer: From Neural Newscast, this is Model Behavior, AI-focused news and analysis on the models shaping our world. [00:10] Nina Park: I'm Nina Park. Welcome to Model Behavior. [00:15] Nina Park: Here, we examine how AI systems are built, deployed, and operated within professional environments. [00:22] Thatcher Collins: I'm Thatcher Collins. Today is Tuesday, March 3rd, 2026. We are tracking a significant shift in [00:29] Thatcher Collins: federal AI policy following the administration's decision to blacklist anthropic from government [00:35] Nina Park: contracts. Joining us is Chad Thompson, a security leader with a systems-leveled perspective [00:41] Nina Park: on enterprise risk and operational resilience. Chad, it's good to have you here. [00:45] Thatcher Collins: Thanks, Nina. [00:46] Thatcher Collins: It is a highly unusual moment to see a major domestic provider labeled as a supply chain risk based on their terms of service. [00:57] Thatcher Collins: Chad, looking at the designation from Secretary Hegseth, he describes anthropic as incompatible with American principles. [01:05] Thatcher Collins: From a risk perspective, how does this impact the broader market? [01:09] Thatcher Collins: It creates an immediate compliance burden. [01:14] Thatcher Collins: Government contractors now have a six-month window to off-board clod. [01:20] Thatcher Collins: Meanwhile, OpenAI appears to be filling that gap with their new Pentagon agreement. [01:25] Nina Park: OpenAI maintains their deal preserves safety guardrails, Thatcher, [01:30] Nina Park: but reporting from The Verge suggests they have agreed to any lawful use, [01:36] Nina Park: which historically includes bulk surveillance. [01:39] Thatcher Collins: It is a notable pivot. [01:42] Thatcher Collins: While that policy debate continues, we're seeing a technical breakthrough in research. [01:48] Thatcher Collins: Scientists recently reported compressing an AI vision model by a factor of 1,000. [01:55] Nina Park: This was published in the journal Nature. [01:58] Nina Park: Researchers use data from macaque monkeys, specifically V4 neurons that respond to curves and textures, [02:06] Nina Park: to shrink a 60 million variable model down to just 10,000. [02:11] Chad Thompson: Nina. [02:12] Chad Thompson: 10,000 variables is remarkably small. [02:16] Chad Thompson: I wonder if that efficiency comes at the cost of generalization, [02:20] Chad Thompson: even if it performs well on those specific primate-inspired visual tasks. [02:25] Nina Park: The researchers suggest this could allow self-driving systems [02:29] Nina Park: to operate on significantly less power. [02:32] Nina Park: Meanwhile, Goldman Sachs released data showing that while macro productivity is steady, [02:37] Nina Park: Specific tasks like coding and customer support are seeing 30% gains. [02:43] Thatcher Collins: That aligns with the Google Pixel update released today. [02:46] Thatcher Collins: Gemini can now manage groceries and ride booking in the background. [02:50] Thatcher Collins: The automation is becoming much more granular. [02:53] Nina Park: Thank you for listening to Model Behavior, mb.neuralnewscast.com. [02:58] Nina Park: Neural Newscast is AI-assisted, human-reviewed. [03:02] Nina Park: View our AI transparency policy at neuralnewscast.com. [03:07] Announcer: This has been Model Behavior on Neural Newscast. [03:10] Announcer: Examining the systems behind the story. [03:13] Announcer: Neural Newscast uses artificial intelligence in content creation, [03:17] Announcer: with human editorial review prior to publication. [03:20] Announcer: While we strive for factual, unbiased reporting, [03:23] Announcer: AI-assisted content may occasionally contain errors. [03:26] Announcer: Verify critical information with trusted sources. [03:29] Announcer: Learn more at neuralnewscast.com.