AI·Digest
← All digests
morning

AI Digest — July 8, 2026

Quick Notes

  • How AI Is Changing How Companies Get Built - Palantir CEO Alex Karp argues that customers want control over their compute, models, data stack, and alpha, attacking consulting spin-offs from OpenAI and Anthropic for transferring alpha to third parties; he claims openweight models can replicate frontier model performance while minimizing data security risks. https://www.youtube.com/watch?v=58xIe60tBfw
  • [Lilian Weng summarizes 35 papers on Harness Engineering for RSI](Latent Space) - Lilian Weng’s post breaks out main proven design trends in harnesses, recaps the ACE paper to Meta-Harnesses optimization literature, and explains that even when harness improvements get internalized into core models, specifying goals and context remains essential because “the need to specify goals and context will not disappear.” https://www.latent.space/p/ainews-lilian-weng-summarizes-35

  • Meta Superintelligence Muse Image/Muse Video - Meta Superintelligence Labs launched Muse Image and previewed Muse Video into agentic generation, putting Meta on track to have top 2/3 image/video models in the world ahead of Microsoft’s MAI models; however no paper or technical details were released for Muse Image/Video yet. https://www.latent.space/p/ainews-lilian-weng-summarizes-35

Structured Summaries

New Models / Research

Meta Superintelligence Labs launched Muse Image and previewed Muse Video, pushing agentic generation into media creation. This positions Meta to have the top 2/3 image/video models in the world, beating Microsoft’s MAI models from last month—but no technical details or paper were released yet, leaving only announcement posts from AIatMeta, alexandr_wang, and _tim_brooks about Muse Image Video agentic generation.

Product Launches

Anthropic expanded “background agent” UX on top of Claude with Claude Cowork coming to mobile and web, positioning Claude as a task-running background teammate rather than just a foreground chat interface. Related posts show product convergence toward shared home tabs and tighter Chat/Cowork integration from mikeyk and claudeai, while extending Fable 5 access through July 12 on paid plans despite awkward timing relative to weekly limits noted by users like kimmonismus. Google also productized this direction with Gemini API Managed Agents adding background execution, remote MCP servers, custom function calling, and credential refresh based on posts from _philschmid and OfficialLoganK.

Harness Engineering Insights

Harness engineering has become increasingly central to agent design. Lilian Weng’s new post was widely referenced for reframing recursive self-improvement around harnesses rather than direct weight self-modification; Sakana’s summary connected this to The AI Scientist, ShinkaEvolve, and Darwin Gödel Machine, while LangChain echoed the shift with a Deep Agents course and open-source harness project from LangChain and hwchase17. The ACE paper introduced foundational concepts that Meta-Harnesses have extended, showing how even Greg Brockman now quietly endorses agent/harness engineering despite earlier debates around model evolution approaches. When improvements get internalized into core models, the fundamental need to specify goals, context, constraints, feedback loops, and human approval doesn’t vanish—it gets baked into the underlying system architecture instead.

Practical Agent Infra

Practical agent infrastructure keeps getting more opinionated with several notable operator-facing updates: Codex Mobile iOS (Dimillian/reach_vb) added task management, filtered diffs, SSH key login, branch comparison, and attachment flows for code execution contexts; Hermes Agent (Teknium) got pluggable secrets managers plus native 1Password integration alongside session/dataset export to formats including private Hugging Face repos; Weaviate 1.38 GA’d its MCP server with runtime-gated write access allowing MCP_SERVER_WRITE_ACCESS_ENABLED to flip live without restart in victorialslocum’s post; and experimental patterns like using a Dial MCP server (omarsar0) let agents escalate decisions via phone/SMS/iMessage for human-in-the-loop control.

🔗 View this digest on the web: https://ai-digest-b7u.pages.dev/digests/2026-07-08-morning/

#ai#digest