Riemann Hypothesis Reportedly Proved Twice Independently
Riemann Hypothesis Reportedly Proved Twice Independently
On November 20, 2026, reports emerged claiming the Riemann Hypothesis has been proved twice independently. The conjecture, open for over 160 years, concerns the distribution of prime numbers and is one of mathematics’ most famous unsolved problems. If verified, it would mark a historic breakthrough and could signal real progress in AI-driven formal mathematical reasoning. Details remain unconfirmed; the claim originated from the Latent Space newsletter and awaits formal peer review.
New Method Separates Interpretability from Reconstruction in Sparse Autoencoders
A new arXiv paper argues that traditional sparse autoencoders (SAEs) use a single sparse representation to both reconstruct inputs and expose interpretable features, forcing the two goals to compete. The authors propose an architecture that separates interpretability from reconstruction, so SAEs no longer trade one for the other when decomposing neural representations. This directly benefits mechanistic interpretability research: cleaner features without sacrificing reconstruction quality.
New Model Tackles Diverse Sampling in Language Models
A new arXiv paper addresses a common language model problem: individual samples are high quality, but as a set they cluster around the same semantic modes and lack diversity. The authors formalize diverse generation as sampling a size-K subset of complete strings from a determinantal point process and build a corresponding model. This directly helps scenarios requiring batch candidate generation, such as content creation and synthetic data, by reducing repetitive outputs.
HapBO Speeds Up Bayesian Optimization with Expert Language Priors
A new arXiv paper introduces HapBO, which feeds expert knowledge expressed in natural language—such as intuitions or documentation—into Bayesian optimization. BO is far more sample-efficient with prior knowledge, but in practice such knowledge often exists only as natural language. HapBO uses an alpha-pi mechanism to turn language priors into optimization signals, cutting the number of trials needed. It suits tuning and materials design where experiments are costly and expert experience is rich.
Game-Theoretic Data Allocation Predicts Expert Model Specialization
A new arXiv paper asks whether the data allocation used to train expert models can also predict which expert performs best on a given prompt. The authors propose a game-theoretic framework where the allocation itself becomes an inspectable blueprint of the division of labor and guides routing at inference. This means expert specialization is visible before training, and routing no longer relies solely on post-hoc evaluation. It has reference value for designing and debugging MoE-style multi-expert systems.
Predicting and Correcting Model-Merging Interference Before Merging
A new arXiv paper offers a better way to handle interference in model merging. Merging combines separately fine-tuned models by adding their task vectors to a shared backbone, but interference is typically assessed only after building and evaluating a candidate merge, making coefficient search expensive. The authors show that first-order linearization can predict interference before merging and guide corrections. This cuts compute for coefficient search and is practical for teams that merge models frequently.
Auditing What GraphCast Learned with Sparse Autoencoder Features
A new arXiv paper uses sparse autoencoder features to audit GraphCast, DeepMind’s autoregressive weather forecasting model. Learned weather models are experimentally accessible computational systems whose internal representations can be inspected and manipulated. The researchers extract SAE features from GraphCast and test their behavior. This provides a tool for understanding the internals of learned weather models and extends mechanistic interpretability from language models to scientific computing models.
ClawDOS Brings an AI Coding Agent to DOS
A developer released ClawDOS, an AI coding agent that runs on DOS. The project drew 156 points and 50 comments on Hacker News. DOS lacks modern networking and toolchains, so fitting an AI agent into this retro environment is largely an exercise in technical curiosity and extreme porting. Practical value for most developers is limited, but it demonstrates the feasibility of AI agents under severely constrained conditions.