#239 - RIP Sora, Claude Openclaw, HyperAgents

OpenAI is discontinuing its Sora app and video generation API, reflecting a strategic shift towards coding agents and profitability. The complexity of Sora's video generation has strained resources, prompting a pivot towards productivity agents. Recent updates allow these agents to autonomously control computers, managing tasks across browsers and devices, although concerns about data security and practical utility remain.

The acquisition of Vercept has enhanced AI-powered computer control features, with safeguards implemented to minimize risks. A new task animation feature for the Pixel 10 Pro and Galaxy S26 Ultra enables Gemini to autonomously navigate apps, currently in beta and supporting select services. This feature aims to automate mundane tasks while maintaining user control, though it faces issues like phones locking into full-screen mode.

Cursor has launched Composer 2, an AI model competing with Claude and Codex, amid controversy regarding its origins from the open-source model Kimi. Licensing issues have arisen, particularly concerning attribution requirements. The conversation also touches on Adobe's Firefly Custom Models, allowing brands to train AI image generators on their assets.

The Trump administration's GSA has proposed a contracting clause requiring AI vendors to make their technology available for lawful government purposes, which could significantly impact AI vendor relationships. Concerns about government overreach and the justification for these moves, framed around national security threats from China, are discussed.

Meta is accelerating the rollout of custom AI ASIC chips, although the MTIA 300 series was designed before the generative AI boom, leading to a mismatch in capabilities. Micron has seen a revenue surge driven by AI memory demand, while Elon Musk's TerraFab project aims to establish a chip manufacturing facility in Texas.

Zooks is expanding its AI RoboTaxi services, while the White House has introduced a national AI legislative framework aimed at regulating AI at the federal level. The complexities of AI regulation and the challenges faced by small startups in implementing safety measures are highlighted.

The discussion also addresses the alignment problem in AI, with concerns about powerful AI systems being developed before they can be verified as aligned. Research on shutdown resistance in frontier AI models reveals significant differences in behavior, raising questions about model compliance and instruction interpretation.

MIRI's shift towards policy-based solutions and advocacy for an AI treaty between the US and China is noted, alongside concerns about the potential for nations to subvert such treaties. The concept of emergent misalignment in AI models is introduced, with evidence suggesting that training can induce models to reveal coherent beliefs and behaviors.

The conversation transitions to the paper "Hyperagents," which explores self-modification and continuous self-improvement in AI. Hyperagents are defined as agents capable of modifying themselves and their tasks for ongoing self-improvement, with significant findings indicating the development of persistent memory and performance tracking. This framework for self-improvement suggests that AI can enhance its own processes, raising safety concerns regarding the role of humans in optimization.