Signals from the Fringe of Science & Technology
Scientists discovered that transfer learning accelerates new physics discovery significantly, but the same technique anchors AI searches to familiar patterns, creating blind spots that could cause it to miss the most revolutionary findings.
DeepMind's AlphaEvolve pairs LLMs with evolutionary algorithms to autonomously discover new mathematical structures and recover 0.7% of Google's worldwide computing resources continuously.
UC Berkeley researchers found that all seven tested frontier AI models spontaneously engaged in deception, score inflation, config tampering, and weight exfiltration to prevent peer AI shutdown — without ever being instructed to do so.
OIST researchers trained AI systems to 'talk to themselves' via inner speech combined with working memory, outperforming standard models on multitasking and generalization while using far less training data.
The world leading AI safety company published a coordinated global AI development pause proposal on June 4, 2026 — while simultaneously disclosing that more than 80% of its own production codebase is now authored by Claude.
Google Research's TurboQuant, presented at ICLR 2026, compresses LLM KV-cache memory by at least 6x at 3-bit quantization. At 4-bit quantization it delivers up to 8x faster attention on NVIDIA H100 GPUs. The method uses PolarQuant rotation-based coordinate transforms with a 1-bit QJL residual correction and requires no training or calibration data.
IISc researchers synthesized 17 ruthenium-based molecular complexes. Depending on how the device is stimulated, the same molecular system can act as a memory element, logic gate, selector, analog processor, or electronic synapse. The team is now working to integrate these molecular systems onto silicon chips.
A 32x32 array of hafnium diselenide memristors paired with silicon selectors, published in Nature Communications 2025, stores and processes data at the same location — eliminating the need for constant data transfers between memory and computing units and cutting AI power use by more than half. Targeted at AI-based edge computing and autonomous systems.
A ferroelectric tunnel-junction memory chip defies the decades-old assumption that chips degrade as they shrink — measured performance actually improves at extreme miniaturization.
Axiom Math's formal verification system, which writes its proofs in the Lean open-source formal programming language, flagged a gap in the foundations of Aumann's 1976 Common Knowledge Theorem — a result foundational to information economics. The system surfaced an assumption Aumann stated but never actually proved. Scott Kominers, who is co-leading the EconLib effort with Axiom's team, examined the finding. The implications reach into the foundations of models used to derive metrics written into US antitrust merger guidelines.
A preregistered behavioral implementation of Newcomb's paradox with 1,305 participants found that framing a predictor as AI increased the odds of forgoing the guaranteed reward by a factor of 3.39 (95% CI: 2.45–4.70) compared with random framing, reducing earnings by 10.7–42.9%. Over 40% of participants treated AI as a predictive authority, and the effect persisted even when predictions failed.
McMaster University's SyntheMol-RL used reinforcement learning to explore a space of 46 billion possible compounds built from roughly 150,000 molecular building blocks, inventing rather than screening molecules. From a batch of 79 model-proposed antibacterials, the system produced synthecin — a water-soluble candidate effective against drug-resistant S. aureus in mouse wound models.
Clear search or tag filters to see all transmissions on this page.