Trump announces a “Super Intelligence Force”

President Trump announced a federal “Super Intelligence Force” on October 4, via Truth Social: Director of National Intelligence Jay Clayton chairs it, FTC chair Andrew Ferguson, Pentagon research-and-engineering undersecretary Emil Michael, and OPM director Scott Kupor are vice chairs, and the group reports to the president and chief of staff Susie Wiles. It has 120 days to deliver a report on AI risks and opportunities, draw up response plans for “SI-enabled threats,” and prevent regulation that would “stifle innovation and competition”; per TechCrunch, Clayton’s framing is that one of the biggest risks is “not being first.” A panel chaired by the nation’s intelligence chief, with deregulation written into its charter, reads as an industrial-competition command post rather than a safety body, and the name follows the White House’s September switch to calling AI “super intelligence.” (See also TechCrunch.)

Google freezes new submissions to its open source bug bounty under a flood of invalid AI reports

Google stopped accepting new product vulnerability submissions to its Open Source Software Vulnerability Reward Program on October 1, citing “a significant rise in automated submissions, the vast majority of which are not valid,” with the next update promised for Q1 2027; supply chain vulnerability reports and reports already in the queue are still being processed. Per TechCrunch, engineers and open source maintainers were buried in AI-generated reports that claimed critical vulnerabilities and turned out to be hallucinated on inspection. Bug bounties assume that submitting costs something and that human triage can keep up; generation tools broke the first assumption, and this freeze is what the resulting imbalance looks like in public. (See also TechCrunch.)

Homa: the end of TCP for AI clusters

Stanford professor John Ousterhout’s AI Engineer talk, “Homa: The End of TCP for AI Clusters,” picked up traction on Hacker News this week; he argues AI clusters should drop TCP for his Homa protocol. The case is a mechanism mismatch: TCP moves byte streams with sender-driven congestion control, while AI inference and agent workloads are full of short coordination messages, and one message stuck behind a long flow can leave a whole group of GPUs idle; Homa works in message units and lets receivers schedule who sends, so short messages jump the queue. He has been making the replace-TCP argument since 2022 (“It’s Time to Replace TCP in the Datacenter”); what changed is that AI clusters convert tail latency straight into idle-GPU dollars, which gives an old argument a new audience.

SCM: local AI search over every photo and video frame on macOS

SCM (Screen Memories), an MIT-licensed open source project, reached 137 points on Hacker News: fully local semantic search over photos and video on macOS, with CLIP/SigLIP for image matching, Whisper for dialogue transcription, Tesseract for OCR, and optional Qwen3-1.7B or Llama-3.2-3B for chat-style retrieval with citations. Cloud photo apps have offered semantic search since Google Photos launched in 2015; the signal here is that the entire multimodal retrieval chain now runs offline on a single Mac (about 600MB of model downloads for the default setup, 1–2GB more for the optional chat models, then no network required), so searching a private media library no longer has to involve the cloud.

Research radar

Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMs

Multi-agent systems often assign different models to different roles, which means re-prefilling the shared context once per model, a real latency and compute tax. This paper proposes HeteroFold: both sender and receiver stay frozen, and mapping plus calibration bridge the three mismatches (tokenizers, layer counts, representation spaces) so KV caches transfer directly across model families; at 32K context, a Llama-3.1-8B to Ministral-3-14B transfer runs 10.7× faster than native prefill while matching text-based agent communication on task quality. Worth a read for researchers building multi-model serving or agent infrastructure; check the calibration overhead and the weaker short-context results before counting on it in production.

Today in one line: AI pushed the cost of submitting to zero without touching the cost of verifying; Google’s bug bounty won’t be the last human-review system forced to hit pause.