Tools for sale
Tools for Sale
If models become abundant, the scarce thing will be reliable ways to let them act.
LLMs do more each year. Not linearly. It feels exponential.
If they can do almost anything soon, they will need tools: prebuilt tools, maybe MCPs. [1]
Marketplaces for these tools might emerge. Tools as a service. On-demand tools for LLMs, with MCPs as the standard. Open-source alternatives will probably exist too. [2]
Maybe most software engineers will build tools in the future. Not web apps. Tools for AI.
Blacksmiths forged swords for soldiers. Engineers may forge tools for AI. [3]
Why not let LLMs build their own tools? Control, consistency, and predictability. [4] Imagine one wrong prompt or a bad guardrail letting a query like this get past:
DROP TABLE *;
Maybe highly aligned "auditor" AIs could enforce safe boundaries. If LLMs can cause harm, limit the blast radius. Maybe the answer is not letting LLMs do things by themselves. Maybe it is telling LLMs what they are allowed to do.
Notes
[1] MCPs are hot right now. Model Context Protocol makes APIs for LLMs more consistent, enabling plug-and-play capabilities.
[2] Tools could scrape YouTube transcripts or access specific faculty RAG systems. Maybe there will be networks of RAGs, one per department, all accessible via MCPs. Instead of just RAG, it will probably be a hybrid of search and query.
[3] You can't move a pencil across the desk no matter how long you scale test time compute.
[4] Consistency and control feel right. Humans prefer certainty, especially when tools can cause harm. Mistakes help you learn, but not when they cannot be remembered in the context window or internalized inexpensively through retraining.