WK15 - MIT How to AI Almost Anything - Interaction 2: Human AI interaction скачать в хорошем качестве

WK15 - MIT How to AI Almost Anything - Interaction 2: Human AI interaction

Human-AI interaction is evolving toward a general schematic for AGI that integrates multimodal inputs, step-by-step reasoning, and constant human-in-the-loop engagement. A key technical advancement in this field is the disentanglement of high-level reasoning (planning in natural language) from low-level actions (specific digital or physical movements like clicking or hovering), which prevents model confusion and improves task success. Case studies like the Interactive Sketchpad demonstrate that providing students with visual hints and code-generated diagrams is more effective for learning than simply revealing an answer. To manage agent reliability, researchers use uncertainty quantification—sampling a model multiple times to check for high entropy—to signal when a model is confused and needs human intervention. Finally, the science of quantification is becoming essential to predict and control model behavior, as multimodal models can often compound social biases or sacrifice robustness for accuracy, necessitating rigorous indirect evaluation methods like human model simulation and debugging. Course website: https://mit-mi.github.io/how2ai-cours... 🔔 Don't get left behind in the AI era. Hit subscribe to get more deep dives into the AI breakthroughs. #humanaiinteraction #multimodalreasoning #aiagents #humanintheloop #uncertaintyquantification