Palmer Luckey on why Chinese open-source models are a non-starter for weapons systems
Palmer Luckey explains that Anduril's AI systems for weapons are fundamentally different from open-source consumer models, focusing on processing raw sensor data rather than text-based language models. He argues that specialized defense AI requires deterministic, auditable systems that have little commercial application, positioning Anduril as a decade ahead in this domain.
Summary
Palmer Luckey addresses concerns about Chinese open-source AI models being used in weapons systems by clarifying that Anduril has been developing specialized AI for defense applications since 2017. He distinguishes between general-purpose foundation models used by consumers and the specialized systems required for military applications. The key difference is that commercial large language models are non-deterministic, non-auditable, and designed for text input/output processing, whereas Anduril's systems process raw sensor data including radar feeds and sonar data from submarines. Luckey emphasizes that these specialized defense models have very few consumer applications and vice versa, making them fundamentally different categories of technology. He credits Anduril's early focus on AI for defense (noting the company's AI acronym was intentional) and positions the company as significantly ahead of recent entrants to the AI-for-defense space who are primarily recent converts to the idea that AI can change everything.
Key Insights
- Anduril has been developing specialized AI for defense applications since 2017, placing them a decade ahead of current discussions about AI in weapons systems
- Consumer foundation models are unsuitable for weapons guidance because they are largely non-deterministic and non-auditable
- Defense AI systems that process raw sensor data like radar feeds and sonar have very few commercial applications and operate in a fundamentally different category from consumer LLMs
- Anduril's systems cannot ingest text and output text, but instead must process raw data from multiple sensor types including radar and sonar from submarines
- Most recent advocates for AI changing everything are newcomers to the field, while Anduril embedded AI focus into its identity from inception with its name acronym
Topics
Transcript
[0:00] Right now, the open source models are all Chinese. So, I'm pretty sure you're not going to be putting Chinese models into remote weapons. Are you going to build Anduril foundation models or how how are you going to get AI in those devices? >> I mean, we've been doing it since 2017. So, we're way ahead of everybody. We're a we're a we're a decade beyond that question. The foundation models that people are building are not the types of models you want to be doing this. They're non they're largely non-deterministic. They're non-auditable. It it's a very different category. The thing that we are building, unfortunately, has very few [0:30] consumer applications and vice versa. There of…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Peter H. Diamandis
Dr. Alex Wissner-Gross's roadmap for how digital twins of your own cells.
Dr. Wissner-Gross proposes that digital twins of cells, trained on large biological datasets similar to how large language models are trained, will enable exhaustive computational searches to discover cures for diseases by finding intervention pathways from diseased to healthy cell states.
Emad Mostaque, Peter Diamandis and DB2 admit even they can't keep up with the model releases anymore
Emad Mostaque and Peter Diamandis express frustration with the rapid pace of AI model releases, stating they cannot keep up with the continuous stream of new models like 5.5, 5, and Astra. Despite acknowledging progress is good, they humorously convey exhaustion from the relentless pace of AI advancement.
Why All The Spaceports Are In Red States
The transcript discusses how federal land availability in red states, particularly in the American Southwest, will likely become locations for new federal spaceports following a White House executive order. The speaker explains that western states like Nevada, Utah, New Mexico, and Arizona, which contain substantial federally-owned land, are leading candidates for space launch facilities, while private spaceports are expected to develop on the Gulf Coast.
If We're in a Simulation, Do This
The speakers discuss the simulation hypothesis and propose that if we're living in a multi-scale simulation where different areas have varying computational fidelity, the optimal strategy would be to spend time around interesting people who would be simulated at higher detail. They extend this to suggest that society-scale simulation could represent a novel form of governance Earth hasn't historically experienced.