The value is in the LLMs

September 4, 2025 (2mo ago, 21.73 y.o.)

MCPs are trivial

⚠️ Disclaimer at the end.

---
"The real value is in the LLMs themselves" — Why?
One of the products I had developed and tried to sell with Gianlu was Deeply: an AI psychologist. I strongly believe that psychology/therapy is one of the fields where AI shines the most when used well — I use it very frequently and I’m always exploring new ways to use it for this purpose, architecting ways to make it less biased toward what I say and to help me discover new points of view and recognize science-based patterns.

If I believe that AI is amazing for therapy, why do I think Deeply is not? The way AI is working for me in this field is through constantly trying new experiments, trial & error strategies, and prompting — the system prompt of Deeply + the adaptive follow-up system are very easy to replicate just by copy-pasting that prompt into ChatGPT, and in fact they would be limiting, since they wouldn’t allow me to do the additional experiments that make my AI therapy experience awesome. By the way, if you’re interested I made a nice guide on how to use AI for “therapy” effectively (for the average person who has nothing clinically diagnosed and just wants to challenge their perspective in a self-growth mindset etc.), maybe it would be cool to publish it and put it on Hacker News.

What about the application layer? Isn't there a ton of value to be added in UX/workflow integration?
One way in which I agree is that, for example, it would be nice if Photoshop or various video editors had a chat interface — there are even nice examples (e.g. https://github.com/loonghao/photoshop-python-api-mcp-server), but as you can see from the triviality of the MCP server that exposes the functionalities and how well it works out of the box, I believe the value is in the LLM itself. And the MCP servers wouldn’t even be necessary except to wrap ways of performing actions programmatically. And for applications like Photoshop/DaVinci, the real work left to do to make them usable in natural language is not building *an MCP*, but rather exposing programmatic actions that can currently only be done through the GUI (and then wrapping them in an MCP, but that’s the easy part).

Special thanks to my friend Manfredi as this was a snippet from one of our conversations.

Disclaimer: when I talk about “therapy” in this post I mean therapy-like behaviors for self-growth intentions for healthy people with no diagnosed mental illness etc. — Think about it as “enhanced journaling” or “self cognitive-behavioral analysis”. You know what I mean. Take this as it is. Contact a real therapist if needed. Do not sue me.