Daily Digest — 2026-05-24

Posted on Edited on In Symbols count in article: 5.4k Reading time ≈ 10 mins.

Overview

Today’s pattern is grounding: frontier AI is getting tied to real software, real provenance systems, real environments, real schools, and more operational network stacks.

Top Stories

Anthropic says AI-assisted vulnerability hunting is already outrunning patch capacity

Anthropic’s initial Project Glasswing update says Claude Mythos Preview and roughly 50 partners have already surfaced more than 10,000 high- or critical-severity vulnerabilities in important software. The interesting shift is that cyberdefense may no longer be bottlenecked on finding vulnerabilities first, but on verifying, disclosing, and patching them fast enough.

Source: https://www.anthropic.com/research/glasswing-initial-update

Google expands SynthID and content-verification tools across its products

Google says it is broadening its provenance stack across Search, Gemini, Chrome, Pixel, and Cloud, while expanding partner use of SynthID watermarking. That matters because AI-origin detection is moving from a specialized trust-and-safety feature toward something that looks like mainstream web infrastructure.

Source: https://blog.google/innovation-and-ai/products/identifying-ai-generated-media-online/

Google connects Project Genie to Street View

Google DeepMind says Project Genie can now turn real U.S. places from Street View into interactive simulated worlds for eligible AI Ultra users. This is notable because world models are becoming less toy-like and more anchored to real geospatial context, which could matter for simulation, training, and embodied-agent workflows.

Source: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/project-genie-expands/

Starlink says its collaboration with the Brazilian non-profit Redes do Futuro has connected 140 schools across the remote Amazon region, serving more than 14,000 students. For the LEO narrative, this is one of the clearest examples of satellite connectivity showing up as basic civic infrastructure rather than just a launch or direct-to-cell headline.

Source: https://x.com/Starlink/status/2058267430829670801

Qualcomm spotlights a new ONNX Runtime plugin path for edge AI deployment

Qualcomm’s weekly AI roundup says it debuted the first ONNX Runtime Plugin Execution Provider and highlighted new AI Hub deployments including CenterPoint for real-time LiDAR object detection at the edge. The practical signal is that model deployment across embedded and automotive systems is getting more productized and less bespoke.

Source: https://x.com/Qualcomm/status/2057952187850895810

Andrej Karpathy joins Anthropic

Karpathy said he has joined Anthropic and expects the next few years at the LLM frontier to be especially formative. Talent moves at this level are worth tracking because they often precede shifts in research culture, technical direction, and the educational layer around frontier-model development.

Source: https://x.com/karpathy/status/2056753169888334312

DeepSeek makes its V4-Pro discount permanent

DeepSeek said the discounted pricing for DeepSeek-V4-Pro is now permanent rather than a temporary promotion. That is a meaningful market signal for builders, because the next phase of model competition increasingly depends on durable cost-performance and long-context practicality rather than only leaderboard prestige.

Source: https://x.com/deepseek_ai/status/2057854261699195173

Research Radar

Beyond Spherical Wavefront: Near-Field Channel Estimation Under Wavefront Anisotropy

Authors: Heling Zhang, Xiujun Zhang, Xiaofeng Zhong, Shidong Zhou
Venue: arXiv
This paper targets a very practical 6G/mmWave problem: near-field channel estimation breaks down when reflected wavefronts stop behaving like nice clean spheres. The authors propose an anisotropic wavefront model and corresponding estimation method, which could matter for precise beamforming as arrays get larger and environments get messier.

Source: http://arxiv.org/abs/2605.22117v1

A Camera-Cooperative ISAC Framework for Multimodal Non-Cooperative UAVs Sensing

Authors: Wenfeng Wu, Luping Xiang, Kun Yang
Venue: arXiv
This one is interesting because it blends cameras with integrated sensing-and-communications (ISAC): vision handles coarse airspace awareness while wireless sensing handles precise tracking. That multimodal division of labor feels realistic for UAV monitoring systems that need both efficiency and accuracy.

Source: http://arxiv.org/abs/2605.22090v1

Demystifying Deep Reinforcement Learning: A Neuro-Symbolic Framework for Interpretable Open RAN Automation

Authors: Jie Lu, Peihao Yan, Pang-Ning Tan, Y. Thomas Hou, Huacheng Zeng
Venue: arXiv
Open RAN automation keeps running into the trust problem: DRL can work, but operators do not want opaque control logic inside carrier-grade systems. This paper proposes a neuro-symbolic distillation approach to turn black-box DRL policies into more interpretable symbolic rules, which is exactly the kind of bridge real deployments need.

Source: http://arxiv.org/abs/2605.10648v2

MIT/Harvard Events This Week

Source Issues

  • TNT calendar still returned stale February–April listings, so event picks were cross-checked on direct Harvard/MIT pages.
  • @ASTSpaceMobile returned no fresh usable posts, and @NextGAlliance was effectively stale.
  • IEEE Xplore and ACM did not surface strong fresh matches quickly, so today’s Research Radar leans on arXiv.

Bottom Line

Frontier AI is becoming more operational and more grounded: in security workflows, provenance systems, real-world simulation, edge deployment stacks, and connectivity infrastructure that reaches actual communities.