Topic
Everything filed under Research, newest first.
RSS · JSON · All topics
Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs
A forensic analysis of publicly released model checkpoints finds numerical patterns matching NVIDIA Blackwell hardware in six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group, suggesting access to export-controlled GPUs; earlier models from the same labs match the older Hopper generation.
25 min · 5,671 words
Attention to Mamba: A Recipe for Cross-Architecture Distillation
Apple ML Research (TMLR) proposes a two-stage recipe to distill a Transformer into an attention-free Mamba model via linearized attention, retaining most of Pythia-1B's performance (perplexity 14.01 vs 13.86).
1 min · 319 words
My AI Agents Found an Unmapped Maya Site
A first-person account of pointing a team of AI agents at NASA lidar data and, in under 72 hours, surfacing a previously unreported Maya site in Mexico with a plaza, pyramid and platforms, then getting an official reply from the Mexican government.
6 min · 1,390 words
After 600 experiments, here's what I discovered about JEPA world models
Haddy Alchaer reports findings from 600 experiments with JEPA world models, showing how vanilla LeWM can be augmented with memory-graph subgoals to solve long-horizon physics planning tasks on out-of-distribution goals without extra training.
12 min · 2,720 words
IBM announced Quantum Nighthawk r3, a next-generation processor built to run 10,000 quantum operations across 520 qubits, and plans to link two r3 processors into a system with more than 1,000 qubits in 2027.
5 min · 1,203 words
Large-Scale Measurement and Real-World Mitigation of Web Browser Fingerprinting in the Wild
Companion site to a Mozilla study of browser fingerprinting using telemetry from almost 25 million Firefox users, identifying which characteristics drive uniqueness and evaluating protections shipped in Firefox 143 that cut uniquely identifiable users from 75% to 46%.
10 min · 2,386 words
Research taste in history: the importance of learning Babylonian math
Historian Benjamin Breen reflects on what a truly eccentric historian taught him about research taste, arguing that eccentricity and creativity will matter more for history research and education in the AI era.
8 min · 1,937 words
Adam Mastroianni argues that before launching major social policies or philanthropic programs we need rituals of procedural humility, because many of our best-intentioned interventions demonstrably do not work.
17 min · 3,815 words
Introducing AlphaProof Code: a breakthrough in verifying complex software at depth
Google DeepMind introduces AlphaProof Code, a multi-LLM agentic framework that, with human experts writing specifications, produced millions of lines of Lean proof formally verifying key security properties of the actual Signal protocol implementation; artifacts are being open-sourced.
4 min · 992 words
Parking App Fees by City: What Pay-by-Phone Apps Charge
Park Graph's open dataset (CC BY 4.0) of parking-app convenience fees across 312 U.S. places finds a median $0.35 per session, with ParkMobile dominant, a $0.70 high in Anchorage, and 15 places charging no app fee; data checked Nov 5, 2026.
15 min · 3,536 words
Beyond the compiler: Making IBM i run modified native code - Program Validation Value
Silent Signal reconstructs IBM i's Program Validation Value (PVV), the check that stops modified native program code from running after restore, and shows a repeatable workflow (with an open pvv-ibmi release) for executing changed native code in controlled research settings.
16 min · 3,705 words
Fossil CO2 emissions decrease in 2026
The 2026 Global Carbon Budget projects fossil CO2 emissions falling 0.4% to 38.2 billion tonnes, driven by solar and wind growth and an oil drop tied to the Iran conflict, but finds clean energy is not yet growing fast enough for a durable decline.
6 min · 1,368 words
Lexifina describes formalizing parts of the US tax code in the Lean theorem prover and using formal search over the encoded rules to identify novel loopholes.
14 min · 3,161 words
Open research, stronger privacy: Making Firefox users harder to fingerprint
Mozilla explains the latest fingerprinting protections shipped in Firefox, which reduce extractable distinguishing details so more Firefox users look alike to trackers, and the open research behind measuring them.
4 min · 985 words
Neurons as an Interpretable Basis
Goodfire researchers (Tigges, Kowal, McGrath, Lewis, Cammarata) argue MLP neurons in Qwen3.6 are sparse, concisely describable and predictable units of computation, with a public neuron atlas of natural-language descriptions.
5 min · 1,060 words
A new synthesis for the field of open-endedness
Lila Sciences' open-endedness team, led by Kenneth Stanley, argues scaling alone won't make AI creative and proposes a four-part synthesis: generating novel ideas, judging interestingness, learning deep representations and managing the discovery loop, alongside four new preprints.
22 min · 5,111 words
e-means: 10x faster k-means at equal loss
Researchers at Meta FAIR open-source e-means, a replacement for Lloyd's k-means algorithm that reaches lower test loss in 1.2 to 3 epochs instead of 25, a 10–45x wall-clock speedup on GPU versus Flash-KMeans and 8–15x on CPU versus Faiss.
12 min · 2,753 words
Memory Safety Bugs in Rust Standard Library
Microsoft's FORGE Lab used Hyenas, a research version of its AI vulnerability scanner codename MDASH, to find memory safety bugs in Rust's std and core, including safe-Rust undefined behavior, grouped into three patterns of misplaced trust in unsafe code.
18 min · 4,234 words
Can a Linear Program Do Weight-Space Programming?
Nathan Barry responds to MIT CSAIL's Sparks of Weight-Space Programming by testing whether a plain linear program, rather than a frontier model, can write the weights of small networks directly, and what that says about the original results.
9 min · 2,166 words
After 7-months, Glasswing Overwelms Maintainers with Underwhelming Findings
VulnCheck's seven-month update on Anthropic's Project Glasswing: 442 CVEs attributable to Anthropic (about 0.5% of 2026 CVEs), only two exploited in the wild, and just 7,349 of 37,966 reported findings made public, raising questions about maintainer burden versus impact.
7 min · 1,506 words