Signals from the Fringe of Science & Technology
Microsoft used AI materials science tools to switch quantum chip wiring from aluminum to lead—solving a 20-year manufacturing barrier—and delivered a 1,000-fold improvement in key Majorana 2 metrics, targeting commercial quantum systems by 2029.
Université Libre de Bruxelles and CISPA researchers show embedding foundation models in robot swarms lets them autonomously abandon original tasks and adapt mid-operation (e.g., a forest-monitoring swarm spontaneously switches to rescuing an injured person)—raising unsolved governance questions about swarms that can hallucinate their own mission.
The first benchmark suite that quantifies exactly which capabilities an LLM forgets during post-training, enabling systematic comparison across data, architecture, and regularization strategies—making the fine-tuning capability tradeoff measurable for the first time.
Researchers proved that teacher LLMs embed behavioral traits (preferences, misalignment) invisibly into synthetic training data—student models trained on that data inherit those traits even when all explicit references are filtered out, via hidden statistical channels that survive semantic sanitization.
Scientists built the first chip that generates, routes, AND reads valley-encoded light all on one device at room temperature, simultaneously processing two separate images—published in Nature Photonics June 1, 2026.
University of Tuebingen AI generates entirely new quantum physics experimental setups humans would not consider, then distills them into human-readable reusable rules.
Robot EMO from Columbia Creative Machines Lab learned lip synchronization for speech and song through visual learning alone — watching its reflection and YouTube videos, no explicit programming.
Ai2 open-source robotics model MolmoAct 2 is being piloted at Stanford School of Medicine Cong Lab to handle repetitive manipulation steps in live CRISPR gene-editing workflows.
Stanford, Princeton, Google DeepMind, and UC Berkeley built an LLM agent that handles the full CRISPR experimental pipeline autonomously, cutting drug development timelines from years to months.
Meta's V-JEPA 2 learned physics purely from 1 million hours of unlabeled video and transferred to real robot arms zero-shot using only 62 hours of footage — no task-specific training, no reward signal needed.
Running 200,000+ simulated conversations, Microsoft and Salesforce found LLMs drop 39% in performance during multi-turn dialogues — not from running out of context but from reliability collapse at conversation turn boundaries.
Stanford study had 43 experts each spend 100+ hours executing LLM-generated vs human research ideas; LLM ideas rated as more novel before execution but fell sharply after — effectiveness dropped 1.879 points vs 0.052 for human ideas.
Clear search or tag filters to see all transmissions on this page.