$Taiwan Semiconductor(TSM.US) $NVIDIA(NVDA.US) I don't understand macroeconomics, mainly relying on narratives and individual feelings to judge. The market's Beta looks very volatile. But from an industrial and technological dynamics perspective, here are some updated judgments:

1. MCP is the industrial protocol for function call/tool use. It has now become the de facto early standard.

2. Current "agents," whether they are workflows with deep human involvement or model-based ReAct, share the commonality that the models have not seen these tools/functions during training (Pre/Post train), nor have they been trained on datasets based on <tools, actions, rewards> (except in the coding domain, which is a special case).

3. During the training phase (especially the pre-train phase), allowing LLMs to fully learn to plan/call various possible tools is crystallizing into a new industrial consensus. In other words, beyond workflow/ReAct, agents should become an intrinsic capability of LLMs. Will today's application-layer innovations be swallowed up once again by the progress of large models themselves?

4. With agents as the goal, there is still much to be done in pre-train/post-train.