AI/ML Hacker News

Building a real-time AI tutor for 5-year-olds

A project focusing on the development of a real-time AI tutoring system specifically tailored for 5-year-olds.

Cybersecurity arXiv cs.AI

Can Trustless Agents Be Trusted? An Empirical Study of the ERC-8004 Decentralized AI Agent Ecosystem

An empirical study of the ERC-8004 protocol for decentralized AI agents, finding significant issues with identity verification and Sybil behavior.

AI/ML arXiv cs.AI

Phase-Preserving Trimodal Transformer for Tropical Forest Biomass Estimation Using Optical and PolInSAR Data

Introduction of the Trimodal Coherent Co-attention Transformer (TCCT) for improved tropical forest biomass estimation using optical and PolInSAR data.

AI/ML arXiv cs.AI

RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies

Introduction of RoboDojo, a unified sim-and-real benchmark for evaluating generalist robot manipulation policies.

AI/ML arXiv cs.AI

Wan-Streamer v0.2: Higher Resolution, Same Latency

Wan-Streamer v0.2 improves audio-visual interaction by increasing resolution to 640x368 while maintaining low latency via multi-GPU context-parallelism.

AI/ML arXiv cs.AI

Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation

Researchers propose a teach-then-search co-training framework to help visual generators overcome world-knowledge bottlenecks through agentic search.

AI/ML arXiv cs.AI

Weak-to-Strong Generalization via Direct On-Policy Distillation

Direct On-Policy Distillation (Direct-OPD) allows strong AI models to inherit RL-induced policy shifts from smaller, cheaper-to-train teacher models.

Tech Business/VC AI News

AWS GraphRAG deployment cuts drug research cycles by 87%

AWS GraphRAG deployment significantly accelerates drug research cycles by unifying proprietary databases into a knowledge graph.

AI/ML TechCrunch

OpenAI says GPT 5.6 is the ‘preferred model’ for Microsoft Copilot 365 amid breakup chatter

OpenAI's GPT 5.6 will be the primary model powering Microsoft Copilot 365 for productivity applications.

Tech Business/VC TechCrunch

Don’t want to invest in Elon Musk? Two new ETFs explicitly exclude him

Two new ETFs have been launched that explicitly exclude companies led or controlled by Elon Musk, such as Tesla and SpaceX.

AI/ML arXiv cs.AI

Cast a Wider Net: Coordinated Pass@K Policy Optimization for Code Reasoning

Researchers propose Coordinated Pass@K Policy Optimization (CPPO) to improve code reasoning in LLMs by exploring diverse algorithmic strategies rather than redundant samples.

AI/ML arXiv cs.AI

Informing AI Policy Assessment using Large-Scale Simulation of Interventions

A new methodology using genetic algorithm-based simulation is introduced to help policymakers identify viable AI governance policy combinations to mitigate harms.

AI/ML arXiv cs.AI

Trading Human Curation for Synthetic Augmentation in RLVR

The paper investigates replacing human-curated RLVR tasks with gate-filtered synthetic augmentations, finding that synthetic data can effectively scale training without losing generalization.

AI/ML arXiv cs.AI

Trust, but Don't Verify: Epistemic Blind Spots in LLM Source Evaluation

Research reveals that LLMs exhibit 'epistemic blind spots,' where they prioritize the analytical style of a source over the actual numeric validity of the data provided.

AI/ML arXiv cs.AI

TLA-Prover: Verifiable TLA+ Specification Synthesis via Preference-Optimized Low-Rank Adaptation

TLA-Prover is a 20B parameter model specifically tuned for synthesizing verifiable TLA+ specifications for distributed systems, significantly outperforming general LLMs.

AI/ML arXiv cs.AI

MetaConfigurator: AI-Assisted RDF Authoring from JSON Data

MetaConfigurator is extended with an RDF Authoring View to help researchers transform JSON/YAML/CSV data into RDF using AI-assisted mappings and SPARQL queries.

AI/ML arXiv cs.AI

The Signs Were Always There: Training-Free Concept Detection and Steering in Raw Transformer Dimensions

The 'Bag of Dims' (BoD) framework demonstrates that concept detection and steering in Transformers can be done using raw hidden state dimensions without any learned dictionaries.

AI/ML arXiv cs.AI

RWGBench: Evaluating Scholarly Positioning in Related Work Generation

RWGBench is introduced as a citation-centric benchmark to evaluate the ability of LLMs to generate scholarly 'Related Work' sections based on actual academic positioning.

Other Hacker News

My Story of 3D Realms / Apogee Part I (2020)

A first-person account of the history and experiences of 3D Realms and Apogee, early pioneers in the PC gaming industry.

Tech Business/VC The Verge

Microsoft’s carbon emissions went up 25 percent last year

Microsoft's carbon emissions increased by 25% last year, primarily due to the expansion of datacenter infrastructure for AI and cloud services.