Daily AI Briefing — May 23, 2026

AI policy and industry power dynamics dominated the cycle:

• Trump abruptly canceled an executive order that would have empowered government pre-release testing of frontier AI models, after Elon Musk, Mark Zuckerberg, and David Sacks lobbied against it overnight, citing China-competition concerns.
• OpenAI opened its first Applied AI Lab outside the US in Singapore, backed by S$300M+ and 200+ technical roles, while global affairs chief Chris Lehane pushes favorable state-level laws.
• Palantir publicly attacked London Mayor Sadiq Khan after he blocked a £50m Met Police AI contract.

Model and infrastructure progress:

• Microsoft released Fara1.5 (4B/9B/27B) browser computer-use agents; the 27B model hits 72% on Online-Mind2Web, beating OpenAI Operator (58.3%) and Gemini 2.5 Computer Use (57.3%).
• AI infrastructure unicorns multiplied: Modal ($355M at $4.7B), Exa ($250M at $2.2B), and Turbopuffer ($100M ARR, profitable).
• China completed AI mapping of its national renewable grid, a strategic edge as AI power demand strains Western grids.

Labor and society: Standard Chartered confirmed 7,800 AI-driven job cuts, with CEO Bill Winters apologizing for calling affected staff 'lower-value human capital.'

Today's research and commentary cluster around post-AGI economics, AI safety methodology, and timely model evaluations, with limited novel technical contributions.

Transformative AI Economics

• *The AI Industrial Explosion (Part 3)* models post-AGI growth by reoptimizing production recipes for cheap labor and capital-intensive factor prices.
• *Will we really put data centers in space?* offers concrete cost/thermal analysis of orbital data centers, pushing back on speculative Musk-era claims.

AI Safety & Alignment

• *Which technical AI safety fields are automated first?* uses feedback quality and economic incentives to predict which subfields frontier labs will automate.
• *Counting Arguments in AI Safety* critiques goal-space counting arguments by drawing parallels to Bertrand's Paradox.
• *We made a map of the doom debate* releases an interactive probabilistic tree of AI threat pathways from AI Safety Camp.
• *AI is Not Normal Technology* rebuts Narayanan & Kapoor, citing biosecurity as a concrete disanalogy.

Model Reviews & Practice

• Zvi reviews Gemini 3.5 Flash as best-at-speed but inferior to Claude Opus 4.7 and GPT-5.5 for serious work.
• *Notes on Collaborating with Claude Opus* documents prompting patterns showing reasoned instructions improve compliance.

Strategy & Timelines

• The proposed DIAL distribution (decision importance adjusted for leverage) reframes timelines reasoning around decision-relevant probability mass.
• A philosophical defense of strong longtermism rounds out the field.

OpenAI's Codex Thursday and a strategic pivot dominated discussions, with Greg Brockman declaring 'the model alone is no longer the product' as Codex gained remote Mac control from phones. The framing of harnesses and integrations as the new battleground resonated widely.

• Anthropic's Project Glasswing reported 10,000+ critical vulnerabilities found, with hints that Claude Mythos Preview will scale this further—raising industry adaptation concerns
• Andrew Ng went viral criticizing the White House green card policy as harmful to US AI competitiveness, while separately opposing Harvard's grade caps
• Perplexity open-sourced Bumblebee, a developer machine scanner targeting AI supply chain risks (MCP configs, extensions, packages)
• Ethan Mollick showcased Gemini Omni's native multimodal video editing on the 1896 train film, highlighting Google's distinctive approach
• Alibaba's Qwen 3.7 Max drew attention for running unsupervised 35 hours and 10x'ing a GPU kernel via 1,158 tool calls
• NVIDIA Research shipped LongLive-2.0 (NVFP4 long video) and AI-Q (open-source deep research agent skill)
• AI2 introduced ArtifactLinker, finding that older models like DeBERTa often beat newer LLMs on under-evaluated benchmarks