Topic
ai training
Technology Private Chef Films Three-Course Meal to Train Robots for Household Cooking
Shift, a division of German startup Microagi, sent a private chef to cook a free three-course meal in the author's apartment while recording every hand movement with a hat cam. The egocentric video data is intended to train future humanoid robots in culinary skills. CEO Bercan Kilic sees this as part of a broader marketplace for AI training data and a step toward an age of abundance.
STAR Allocation Method Improves Text-to-Image AI Training with Spatiotemporal Rewards
A new method called SpatioTemporal Adaptive Reward (STAR) Allocation improves reinforcement learning post-training for text-to-image generation. By using text-image attention to allocate rewards to relevant latent regions, STAR enhances compositional semantic alignment, text rendering, and preference optimization without changing the external reward source. The method was validated on Stable Diffusion 3.5 Medium, achieving top scores on GenEval, OCR, and PickScore benchmarks.
Self-Play RL with 30 Minutes of Human Data Trains Coordinated Driving Policies
A new approach from researchers trains autonomous driving policies using self-play reinforcement learning regularized by only 30 minutes of human demonstrations. The method requires 2500x less human data than imitation learning and completes training in 15 hours on a single consumer-grade GPU. The resulting policies successfully coordinate with held-out human trajectories, avoiding the alien driving conventions common in pure self-play systems.
Green AI Carbon Optimizer Recommends Carbon-Efficient Training Locations and Forecasts Global AI Energy Demand
The Green AI Carbon Optimizer, presented in a new arXiv paper, offers two tools: a carbon-aware cloud region recommender for AI training and a power-law forecasting pipeline for global AI energy demand. By combining grid carbon intensity, renewable share, and PUE across 100+ regions, optimal region selection can reduce emissions by 97.2% versus the worst region. The forecasting model, based on 26 anchor models, projects 2030 AI energy demand between 7 TWh and 1,436 TWh depending on scenario assumptions.
Auditing Reward Hackability in Code RL Training Environments Reveals 28.5% Weak Test Suites
A research paper by Rajan on arXiv measures reward hackability in code reinforcement learning (RL) training environments. On a 49-task sample of SWE-bench Verified, 28.5% of tasks have test suites weak enough that a Docker-verified incorrect patch passes them. The study also proposes a hardening procedure using an LLM judge and Docker gate to detect defects.