Focus on data, not models

I regularly attend both offline and online conferences on AI. They are full of brilliant people, interesting projects, and sophisticated technical solutions.

Most conferences follow the same pattern:

• they pick some classification of AI agents;
• then dive deep into one specific level — whichever the author of that particular conference has reached.

As an architect, I always try to think in terms of data. Studying how modern agentic systems are built is useful and interesting. But I am struck by how little attention is paid to the most important thing: data.

We have already been through this with prompts

A few years ago, knowing how to talk to a model properly gave you a real edge. Huge guides on prompt engineering circulated, along with special constructions, chain of thought, few-shot examples, roles, detailed system prompts. But models got smarter, and much of what you used to do externally has moved inside.

I suspect that today's AI harnesses will face the same fate. Not literally: infrastructure, tools, and integrations are not going away. But people will have less and less need to manually tell a smart model in what order to call five tools, when to spin up a sub-agent, what to put in memory, and how to check its own output.

After a few more model generations, most of today's elegant agentic architectures will probably look the way prompts from 2024 look now.

But almost nobody talks about data. Not corporate data, not another vector database of documentation, but a person's own data.

The best ideas come at unexpected moments: on a walk, on a motorcycle, in the shower, at the gym, before sleep. I urge you to save your data in whatever way is convenient for you. It does not matter whether you use Notion, Obsidian, a Telegram channel, a folder of Markdown files, or one endless document. The important thing is that the data stays with you and can be exported in a readable format. This is the only component of a future personal AI system that you cannot download, copy from someone else, or buy retroactively. And the only one whose value only grows over time.

And LLM models are a tool to which you can potentially transfer not just documents, calendars, and task lists, but a significant part of your own context.

My implementation

On a rented server runs Hermes with a Telegram gateway on models from OpenCode Go, transcribing voice messages through Groq Whisper Large and writing notes into its own system.

What this changes

Over time, repeating patterns become visible: which ideas you return to, which mistakes you reproduce, which projects you only consider important in words but never actually come back to months later. Decisions become visible from the outside. But most importantly, a rich context has accumulated for future AI assistants.

The bottom line

A model without your context is like a good doctor without a medical history. It may be very smart, but it cannot help you.

Scaffolding around models will become obsolete in a couple of generations. An archive of your own thoughts will not.

A single file on your computer counts.