From:Internet Info Agency 2026-03-22 20:56:00
On March 22, Andrej Karpathy, former head of Artificial Intelligence and Autopilot at Tesla, revealed that while training the GPT-2 model, after months of unsuccessful manual tuning, he handed over the optimization process to an autonomous agent. The AI found—overnight—a set of parameter combinations that humans had easily overlooked but which significantly impacted performance. These parameters exhibited complex interactions, demonstrating that systematic search outperformed human intuition. Karpathy urged researchers to deliberately "step back" in domains with objective evaluation metrics and leverage automated tools effectively. He noted that although AI has advanced rapidly in verifiable tasks like programming, its application remains challenging in areas that are difficult to quantify.

Geely Galaxy Warship 700 Global Debut: AI-Powered All-Terrain SUV Unveiled
NIO CEO Li Bin: Over 550,000 In-House Developed Chips Mass-Produced
BYD Launches OTA Update Tracker for Dynasty & Ocean Series Owners (2025–2026 Models)
NIO ES9 to Fill the Void in Premium Electric SUVs Above RMB 600,000
U.S. Regulators Escalate Probe into Tesla's FSD System, Covering 3.2 Million Vehicles
Tesla to Spend $2.9B on Chinese Solar Equipment to Accelerate U.S. 100GW Capacity Buildout
NIO Unveils Gen 5 Battery Swap Station Timeline: Testing Starts End-March, Large-Scale Rollout in H2
Bentley Posts Seventh Consecutive Year of Profit, Accelerates Electrification Shift