OpenAI, Google Accused of New Safety Gaps

Apply to our Digital Media Accelerator!

Updates from FLI

• FLI President Max Tegmark spoke to the Daily Caller about the global ‘AI arms race’ that’s intensifying: “Whether this ends in a global AI renaissance or in disaster depends on choices we make now. I’d much rather see cooperation - even just basic communication - than a reckless contest to see who can deploy an uncontrollable AI first."

• We’re thrilled to have supported the development of this new (free!) online course about shaping positive futures with AI, from the Foresight Institute’s Existential Hope program. Check it out now, and keep an eye out for FLI Executive Director Anthony Aguirre in the course content!

• FLI’s AI & National Security Lead Hamza Chaudhry spoke to Reuters about AI disruption, especially regarding human labor displacement:

• Hamza also spoke to Politico about the self-improving AI models tech companies are racing to build, stating “the product being sold is the lack of human supervision — and that’s the most alarming development here.”

• We’re proud to have supported the latest video from Pindex outlining experts’ predictions about AI:

• We’ve published the latest edition of Anna Hehir and Maggie Munro’s Autonomous Weapons Newsletter, in advance of upcoming UN AWS talks.

• On the FLI podcast, Astera Institute AGI Safety Researcher Steven Byrnes joined host Gus Docker to discuss brain-like AGI, and its dangers.

• Also on the podcast, Foresight Institute’s Allison Duettman joined for an episode on the pros and cons of centralized vs. decentralized AI, and how to build human-empowering AI.

• Finally, George Washington University Assistant Professor of Political Science Jeffrey Ding joined to discuss China’s AI strategy.

What We’re Reading

• AI Frontiers: From the Center for AI Safety and advised by Lawrence Lessig, Yoshua Bengio, and Stuart Russell, AI Frontiers is a new digital platform featuring expert dialogue and debate about hot AI topics. Check out their fast-growing collection here.

• AI 2027: Former OpenAI researcher Daniel Kokotajlo, along with Scott Alexander, Eli Lifland, and Thomas Larsen, published a chilling look at how AI could realistically take over by 2027. Explore it in “AI 2027”, here.

• Demonstrating dual-use AI: SecureBio and the Center for AI Safety recently released the Virology Capabilities Test, a benchmark measuring LLMs’ ability to problem-solve in wet labs handling biological and chemical materials. For example, the test found that OpenAI’s new o3 model outperformed 90%+ of human experts - highlighting the potential for widely-available AI models to aid in both helpful lab work and harmful bioweapon production. You can read more in their paper, in TIME’s coverage, and in AI Frontiers.

• There’s room for every concern: A new study published in PNAS finds that "existential risk narratives increase concerns for catastrophic risks without diminishing the significant worries respondents express for immediate harms" - in contrast to claims that talking about AI's catastrophic risks distract from the harms it's already causing.

• Scaling laws for scalable oversight: FLI President Max Tegmark, along with Joshua Engels, David D. Baek, and Subhash Kantamneni, published a new paper trying to “quantify how smarter AI can be controlled by dumber AI and humans via nested ‘scalable oversight’". You can read it now, here.