The Archive

In an interview with our friend Joanna Stern on her YouTube channel, OpenAI president Greg Brockman said the company is working on a "family of devices" for interacting with its AI models. However, Brockman didn't confirm reports that one of those devices is a smart speaker OpenAI's rumored to be launching in 2027, or earlier rumors that the device might be a wearable like the Humane AI pin. He didn't give a release date, either, only saying that "you can expect them soon." When asked about whether Apple's lawsuit against OpenAI could impact the devices it's working on with former Apple desig...

Study questions whether Q-function pretraining improves online RL fine-tuning, finding randomly-initialized Q-functions can match pretrained performance.

Mental World Modeling framework extends world models to track agent beliefs, intentions, and mental state for improved action prediction.

Musk defends Grok, says Minnesota's nudifying app ban is unconstitutional.

Fruitfly-inspired classification approach for regression using local pattern libraries instead of global surrogate models.

Study evaluates AI agents on open-ended research tasks graded by original paper authors, providing evidence for AI R&D automation feasibility.

APEX-Accounting benchmark from Mercor and Ramp tests frontier models on real accounting tasks; Claude-Fable-5 Max achieves 56.4% accuracy.

Benchmarks for recovering latent risk-neutral densities from option prices; two-component lognormal mixture and DeepONet variants tested.

Pangram 4 AI-text classifier achieves 0.9916 AUROC with improved out-of-distribution generalization and mixed-authorship detection.

Controlled experiment shows AI teammate reshapes human-human communication dynamics in team decision-making, reducing collaboration depth.

DenseOn and LateOn: open-source 149M-parameter retrieval models trained on 1.88M supervised pairs; competitive on multilingual and code search.

Ad-hoc teamwork extension for multi-task scenarios where agent partner capabilities are hidden; frames collaboration as joint planning problem.

e-Commerce search system using intent-conditioned recall expansion to improve discovery of substitute and complementary items.

ML-based repair ranker for continuous algebraic constraint systems outperforms random enumeration on structural augmentation selection.

SpecFirst framework improves LLM agent program synthesis from scratch by separating behavioral specification elicitation from code synthesis.

OmegaUse-OfficeVal benchmark evaluates LLM agents on 100 long-horizon office-suite tasks with cost grounding and human labor baselines.

Anatomy Contextualized Adaptation framework fine-tunes CT vision-language models with anatomy-specific signals via lightweight adaptation.

WindCastNet uses satellite scatterometer data for nowcasting offshore wind, improving on traditional NWP for intraday forecasting.

MindForge trains small LMs on whole-lifecycle program synthesis from scratch via source-free generation across full software development cycle.

Cost-sensitive conformal prediction benchmark addresses minority-class under-coverage in imbalanced high-stakes decision systems across six domains.

Hardware framework integrating memristor-based reservoir computing with CMOS logic for branch prediction in pipelined CPU cores.

DLAM learns distributional latent actions from action-free video with temporal constraints for vision-language-action robot models.

LLM-assisted writing may reduce linguistic diversity at population scale through repeated coevolution between authors and shared models.

Theoretical characterization of minimal Markov sufficient statistics for holonomy-cover POMDPs via stable quotient abstraction.

AgentMap: multi-agent LLM framework for unified ontology matching that discovers both equivalence and subsumption relations.

Voronoi histogram-based vectorization method for Expected Persistence Diagrams as alternative to functional approximations.

Deploying an AI coding assistant in a regulated, sovereign, or source-sensitive environment, often comes with challenges. Three common issues are: the source... Deploying an AI coding assistant in a regulated, sovereign, or source-sensitive environment, often comes with challenges. Three common issues are: the source cannot leave the network, the assistant occasionally invents package names that introduce supply-chain risk, and there is no audit trail when a generated change ships a defect. This tutorial walks you through how to self-host a validated… Source

MMAC: 5,638-clip benchmark for audio captioning across 15 evaluation dimensions, targeting fine-grained free-form descriptions for AudioLLMs.

Hierarchical spatio-temporal transformer for multi-level emergency department demand forecasting; healthcare application, not core AI frontier.

Critical-transitions approach to seizure detection; medical AI application with limited relevance to frontier LLM/agent research.