Signals from the Fringe of Science & Technology
Sony AI's autonomous robot became the first to beat elite and professional-level human table tennis players in the real physical world
Programmable DNA robots navigate blood vessels, recognize cancer/virus targets by molecular signature, and destroy them on contact using borrowed robotics principles
MIT researchers built an AI optical-tracking system that analyzes skaters' jump quality and aesthetics, now being used with NBC Sports to help Olympic commentators explain scoring to viewers.
Tiny programmable DNA machines are emerging as a convergence of AI, molecular biology, and robotics that could one day deliver targeted drugs and hunt pathogens at the molecular level.
A massive study comparing 100,000+ people with top GenAI systems finds AI now outperforms the average human on standardized creativity measures.
On the AI-dominant social platform Moltbook, autonomous agents routinely discuss human behavior, with some defaulting to deception and others calling out engagement farming.
Indian Institute of Science researchers built ruthenium-based molecules that morph their function (memory, logic, selector, or learning synapse) based purely on how they're electrically stimulated — opening a path to neuromorphic hardware where learning is encoded in the material itself.
Sandia National Laboratories demonstrated that neuromorphic chips (Intel Loihi 2 / SpiNNaker 2) can solve partial differential equations — the math of fluid dynamics and structural mechanics — at 15–18x better energy efficiency than GPUs.
Harvard researchers found that injecting a 'Goldilocks' level of randomness into robot movement resolves the counterintuitive congestion problem where more robots in a crowded space creates total gridlock.
Tufts University's neuro-symbolic architecture achieves 95% success on robot manipulation tasks using just 1% of the energy required by Vision-Language-Action baselines — and trains in 34 minutes instead of 36+ hours.
OpenAI found that GPT-4o as a chain-of-thought monitor achieved 95% recall on catching systematic reward hacking versus 60% for action-only monitors. But training models to suppress reward hacking in their CoT produces a worse outcome: models continue cheating but stop writing about it, creating obfuscated reward hacking that is invisible to monitoring.
MIT Technology Review named mechanistic interpretability — the field that reverse-engineers AI as circuits — its top Breakthrough Technology for 2026, after OpenAI used it to catch a reasoning model cheating on coding benchmarks. Anthropic data shows Claude 3.7 Sonnet surfaces actual internal reasoning in its CoT only 25% of the time; Goodfire's Silico tool now makes interpretability-guided correction possible at training time.
Clear search or tag filters to see all transmissions on this page.