A 2026 Nature Communications paper reports a programmable 3D photonic neural network chip with 8 stacked 8×8 layers — the first photonic NN demo built around vertical compute layers rather than wider planar arrays. It delivers 6,554 TOPS throughput, 93% accuracy on MNIST classification, and 94% fidelity in optical pattern generation, surpassing leading planar photonic systems. The architectural twist matters because it transplants the vertical-stacking strategy that gave GPUs and HBM their density advantage into the photonic regime, suggesting a credible scaling path for optical AI compute.
Aalto University physicists led by Jere Mäkinen made a time crystal out of magnons injected into a Helium-3 superfluid, then for the first time connected it to a tiny mechanical oscillator without breaking its perpetual ticking — turning the crystal into a controllable optomechanical element. Until now, any external hookup would have collapsed the crystal's signature self-sustained motion. Published May 5, 2026 in Nature Communications, the result opens the door to time-crystal-based quantum memory and ultrasensitive sensors that exploit the crystal's unusual longevity.
A 30-author position paper posted May 1, 2026 argues the long-running debate about making LLMs themselves Bayesian is misdirected — what actually needs Bayesian rigor is the *orchestration layer* that decides which tool to call, which expert to consult, and how much compute to spend. They define concrete Bayes-consistency properties for agentic control loops and show how calibrated belief tracking over latent task quantities cuts redundant tool calls and improves resource allocation. It reframes 'agentic AI' from a wrapper-engineering problem into a coherent decision-theoretic one with falsifiable design rules.
A 2026 Advanced Science paper reports soft pneumatic robots — grippers and crawlers with internal air-chamber networks — fabricated directly via 3D digital light printing from a single dynamic-covalent polymer that can be heat-healed after damage. Healing efficiencies reached 94.5% for grippers and 87.5% for crawlers, and the same material is fully reprocessable into new parts. The result is unusual because most self-healing soft-robotics work uses post-hoc material additives; here the healing chemistry is baked into a printable resin, so the whole robot inherits the property at no design cost.
A 2026 npj Computational Materials paper introduces BEE-NET, a bootstrapped ensemble of equivariant graph neural networks that predicts the Eliashberg spectral function and superconducting Tc to within 0.87 K of DFT — fast enough to whittle 1.3 million candidate structures down to 741 stable predicted superconductors. Two of the AI-predicted compounds were then experimentally synthesized and confirmed, including a novel zirconium–indium–nickel alloy with Tc ≈ 9 K — start-to-finish discovery in three months. The pipeline's 99.4% true-negative rate is what makes the screening tractable for a rare property like superconductivity.
BISC is a single-chip brain-computer interface that integrates 65,536 micro-ECoG electrodes, 1,024 recording channels, 16,384 stimulation channels, wireless power, and an ultrawideband radio onto a 3 mm³ CMOS die thinned to 50 microns — under one-thousandth the volume of a standard implant. The wireless link sustains 100 Mbps, at least 100× higher than any other wireless BCI, and the device records stably and chronically from animal cortex. It is the first plausible path to dense, surgically minimal cortical interfaces that don't require tethered electronics or external bone-mounted hardware.
Vladislav Kurilovich and colleagues at Google Quantum AI used the 72-qubit Willow processor to identify a previously unrecognized error mode: ionizing radiation does not just create quasiparticles that tunnel through gap-engineering defenses but also shifts qubit frequencies by up to 3 MHz, desynchronizing them from microwave control pulses and producing correlated phase errors. This explains the stubborn 'logical error rate floor' that has plagued large superconducting arrays despite shielding. The team mitigated it with echo pulses that cancel the unwanted frequency shifts, turning a mystery into a fixable engineering problem.
AI math startup Axiom released Axplorer, a free open-source pattern-discovery tool that solved the Turán four-cycles problem in graph theory in 2.5 hours on a single Mac Pro — a problem that previously required weeks of brute-force computation on supercomputer clusters using Meta's earlier PatternBoost. Crucially Axplorer is not an LLM that derives answers but an iterative generator that produces mathematical patterns the user steers by selecting interesting outputs. The system also matched or improved on the best-known results for two further graph-theory problems, and the company released it specifically to democratize this kind of pattern-search math.
On April 13, 2026, GPT-5.4 Pro produced a Lean-verified proof of Erdős Problem #1196 in roughly 80 minutes from a single prompt, by modeling integer multiplication as a Markov chain via the von Mangoldt function — a route no human had used in the ~90 years researchers had attacked problems of this type, who had instead translated the question into probability theory in 1935. Terence Tao verified the proof within 24 hours and extended it into a broader framework, calling it 'a meaningful contribution to the anatomy of integers that goes well beyond the solution of this particular Erdős problem.' Since January 2026, 15 Erdős problems have moved from 'open' to 'solved,' 11 with AI co-credit.
On April 13, 2026, GPT-5.4 Pro produced a Lean-verified proof of Erdős Problem #1196 in roughly 80 minutes from a single prompt, by modeling integer multiplication as a Markov chain via the von Mangoldt function — a route no human had used in the ~90 years researchers had attacked problems of this type, who had instead translated the question into probability theory in 1935. Terence Tao verified the proof within 24 hours and extended it into a broader framework, calling it 'a meaningful contribution to the anatomy of integers that goes well beyond the solution of this particular Erdős problem.' Since January 2026, 15 Erdős problems have moved from 'open' to 'solved,' 11 with AI co-credit.
Vladislav Kurilovich and colleagues at Google Quantum AI used the 72-qubit Willow processor to identify a previously unrecognized error mode: ionizing radiation does not just create quasiparticles that tunnel through gap-engineering defenses but also shifts qubit frequencies by up to 3 MHz, desynchronizing them from microwave control pulses and producing correlated phase errors. This explains the stubborn 'logical error rate floor' that has plagued large superconducting arrays despite shielding. The team mitigated it with echo pulses that cancel the unwanted frequency shifts, turning a mystery into a fixable engineering problem.
AI math startup Axiom released Axplorer, a free open-source pattern-discovery tool that solved the Turán four-cycles problem in graph theory in 2.5 hours on a single Mac Pro — a problem that previously required weeks of brute-force computation on supercomputer clusters using Meta's earlier PatternBoost. Crucially Axplorer is not an LLM that derives answers but an iterative generator that produces mathematical patterns the user steers by selecting interesting outputs. The system also matched or improved on the best-known results for two further graph-theory problems, and the company released it specifically to democratize this kind of pattern-search math.