The Archive

Probes chemical language models across eight pre-trained variants to identify which molecular substructures are encoded.

ART learns adaptive timestep schedules for diffusion sampling via continuous-time control and actor-critic optimization.

Paper-replication workflow enables coding agents to verify computational claims in scientific ML papers with recorded evidence and automated verification.

AbsoluteDegradation introduces physics-based synthetic film degradation pipeline and benchmark for archival film restoration without paired training data.

Deep learning model for automated population-scale penile tissue segmentation in MRI to enable quantitative male reproductive health phenotyping.

Rolling Split Conformal Prediction framework for pre-incident traction loss detection via tire slip monitoring in vehicle safety systems.

Behavioral monitoring technique to distinguish guardrail blocks from LLM rejections in black-box adversarial evaluation of production systems.

vLLM-based inference pipeline for unified audio understanding and generation with native support for multi-token prediction and delay-pattern interleaving.

FitOne domain-specialized LLM (8B/32B) improves scientific fitness coaching reliability through domain-specific post-training.

Artificial intelligence may have captured the public imagination through chatbots and image generators, but some of its most consequential use cases are unfolding far from consumer-facing tools. In industries where physical infrastructure, operational continuity, and safety are paramount, AI is becoming a core operating layer. With its sprawling industrial systems and constant stream of operational…

ContextNest open specification for verifiable context governance in autonomous AI agents with provenance, version control, and integrity guarantees.

Bias-aware Bayesian active ranking framework identifies top-k items under fixed budget while mitigating systematic biases in LLM judge comparisons.

Structured Gaussian process classifier integrates biological pathway graphs into kernel design for high-dimensional omics classification with uncertainty quantification.

WBMM optimizes large-kernel convolutions via windowed batching and regular memory access, improving efficiency on large feature maps.

Guided Action Flow enables test-time guidance of frozen SmolVLA policies using learned action-chunk critics for robot manipulation.

Improved Fourier Neural Operators for Rayleigh-Bénard convection via time-increment prediction achieve faster inference with 314k parameters.

SUNTA uses prediction-error-driven chunking in hierarchical state-space models for improved long-horizon video prediction.

Evolutionary Wave Function Collapse combines procedural content generation with evolutionary search over small input examples for level generation.

HaloGuard 1.0 releases open-weights constitutional safety classifier achieving state-of-the-art multilingual prompt-safety performance at 1/10 model size.

Maven framework adds evidence-state value functions to long-context RL, rewarding intermediate reasoning transitions for improved synthesis.

kNNGuard provides training-free LLM guardrails via multi-layer kNN over hidden activations, requiring only 50 labeled examples.

ADTC formalizes exhaustive analysis of decision trees via algebraic model counting for explainability verification.

Ben Guez has "a bunch of potential international wives in [his] DMs," thanks to an automated script he set up using OpenClaw, Claude code, and Instagram trials.

SA-HGNN applies hyperbolic graph neural networks to EEG data, capturing hierarchical brain connectivity for depression recognition.

Google tries balancing AI data center emissions with clean energy efforts.

OpenAI has floated giving the US government a 5 percent ownership stake as a way of easing tensions with the Trump administration and blunting mounting public backlash against AI, according to the Financial Times. CEO Sam Altman argued that giving the public a financial interest in the company would be the best way to share the upside of AI, the FT reported, citing two unnamed people familiar with the talks. He's said to have first pitched the idea to Trump early last year. Altman reportedly suggested the 5 percent figure. Based on OpenAI's latest funding round, which ended with the company v...

Newsletter post noting absence of significant AI industry announcements on a given day.

AIEWF speakers debate autoresearch and software factory vision, raising concerns about human agency and control in AI-driven development.

Neo is Bhavin Turakhia’s fifth venture and his latest involving enterprise software. This time he's taking on Microsoft Office, Google Apps with AI.

Introspection co-founder explains autoresearch loops, agent recipes, and self-improving systems while arguing humans remain essential to AI software development.