The author argues that extreme hardware specialization, where a dataset is compiled directly into custom silicon, is closer than expected. As a proof point, a single-cycle Connect-4 chip was learned end-to-end from simulated games for under $10 of compute, running at ~25 MHz and consuming a few hundred picojoules per move, matching a CPU engine's speed at two to three orders of magnitude lower energy.
End-to-End Learning
端到端学习
Training a model or system as a whole from raw input to final output without hand-engineered stages.
Curated Intelligence
Baidu unveiled Apollo World, a full-modal world model for physical AI, introducing '2+3' criteria that prioritize causal fidelity over visual rendering. By modeling traffic participants as decision-making agents, it provides a high-fidelity virtual training environment for Luobo Kuaipao (Apollo Go), aiming to resolve the 'impossible triangle' of safety, real-time performance, and interaction in autonomous driving to accelerate large-scale L4 deployment.
Cartesia Research introduces 'Hierarchical Networks' to address the reliance on hand-crafted encoding in current LLMs. The study advocates returning to the 'bitter lesson' by enabling end-to-end learning directly from raw bytes, aiming to improve generalization in data-scarce domains like science and robotics.
Microsoft Research introduces GridSFM, an end-to-end ML pipeline that combines solar wind physics with local geography to forecast space weather risks for 66,935 US substations. The system detected nearly 80% of major events during evaluation and provides grid operators with 30-60 minute advance warnings.
The BAIR team proposes an information-theoretic framework for evaluating imaging systems at NeurIPS 2025, directly quantifying useful information in noisy measurements. This method replaces traditional metrics like resolution or SNR with an estimator that assesses object distinguishability from raw data, targeting AI-driven non-visual imaging scenarios such as autonomous driving and MRI. Code is open-sourced.
The Hugging Face LeRobot framework now supports L2D, the world's largest open-source multimodal self-driving dataset. Comprising over 90 TB of data from 30 German cities (5000+ hours), it includes HD camera feeds, vehicle states, and natural language instructions, specifically designed for training end-to-end autonomous driving models.