@Michaelzsguo: Silicon Valley has no shortage of parties. Not the eat-and-drink kind, but gatherings for technical exchange and sharing. Just now, Andrej Karpathy shared his years of experience with AI Agents at another small meetup. Karpathy mainly covered four points: 1. In 2016, OpenAI was too early in building agents…

There are quite a few parties in Silicon Valley. Not the kind for eating, drinking, and having fun, but gatherings for sharing technical insights. Just recently, Andrej Karpathy shared his experiences with agents over the past few years at another small meetup.

Karpathy mainly covered four points:

1. In 2016, OpenAI was too early for agents At that time, they tried to use reinforcement learning to build agents that could operate computers, click on web pages, and complete tasks, in a project called World of Bits. But the tech stack was immature, and they basically had to rely on RL to “bump into rewards,” so the results were poor.

2. The right approach back then was to build language models first He believes that later developments proved it was correct to set aside agents and focus on building language models. Today, agents are heating up again because the underlying tools are completely different: instead of hard-training with RL, we now combine language models, tool calling, memory, planning, and other capabilities.

3. Agents are easy to demo but hard to productize He warned people not to be fooled by demos. Autonomous driving and VR are similar examples: it’s easy to imagine, easy to make a demo, but turning it into a reliable product may take a decade. The same goes for agents — they are suitable for long-term investment, not short-term fantasies.

4. Agents need to draw inspiration from human cognitive structures He mentioned that we can find inspiration again from neuroscience — for example, the hippocampus corresponds to memory and retrieval, the prefrontal cortex corresponds to planning and reflection, and the thalamus/basal ganglia may correspond to multi-module coordination. Language models are just one part; a true agent requires a full set of cognitive tools.

Finally, his encouragement is: entrepreneurs and hackers working on agents now are really at the frontier of capability. Large labs have thoroughly explored large model training, but agents are still new, and many things have yet to be systematically explored.