It's also interesting that he is using deepseek-v4-flash. I have been spending 100s of millions of tokens with it (~10$), and it's actually very impressive. I have been using it with my own coding agent (minimal harness) to improve itself, and it's just amazing. More to share on that soon.
I started to build mines around Nov last year, and this is the app I use the most today. It has everything: from memory to task manager to its own browser use tools.
The freedom and power this gives is hard to describe in words. I think eventually this whole thing will just adapt on its own, but I will only be able to do so if I own the harness, agent, the orchestrator, the LLM, and the code itself.
So far, I have built an orchestrator and a pi-like coding agent, and it's absolutely insane. I will open-source a lot of this at some point so others can build their own orchestrators and coding agents. My next target is to build my own LLM tuner (something like autoresearch), where I need a lot more resources, but I am starting small with my own laptop to get a few workflows that I can automate. And I plan to make all of it easier for anyone to adapt.
Own the AI. Own the Harness. Own the Data. Use AI to help you build all of it.
I will be doing some live sessions about this in the coming weeks. Stay tuned: https://academy.dair.ai/events