正文 Markdown
From Karpathy’s latest post: “LLM apps will organize, finetune and actually animate teams of them into deployed professionals in specific verticals by supplying private data, sensors and actuators and feedback loops.” This is exactly right. In a world of AI agents, there is a much thicker layer above the LLM than was initially perceived. The thin wrapper criticism somewhat worked in a world where people were repackaging tokens with a lightly customized interface or system prompt, which was largely all that was possible 2 years ago. But AI agents will combine tools, proprietary data, highly domain specific system prompts, specialized interfaces for the human-in-the-loop parts of the workflows, advanced context engineering to deal with context window limits, and more. The vast majority of these will perform better when tailored to a specific vertical, job function, or type of task. Further, to get real adoption in the enterprise, a heavy degree of system integration and change management is generally needed to drive the workflow changes and adoption. This is why companies (or at least products) focused on specific opportunities will often be required to actually power these workflows. Tons of opportunity here in the year to come.