A developer on Hugging Face has fine-tuned a 12-billion-parameter language model specifically to help his wife start doing laundry. The AI world contains multitudes.
The project is called NeuroBait. It does not make to-do lists. This is, it turns out, the entire point.
A to-do list for someone in a freeze is just more to choose between, plus a faint voice saying 'try harder.' Correct, and completely useless.
What happened
Developer Haris-Subrata trained a LoRA adapter on top of Google's Gemma 3 12B model, using a small hand-curated synthetic dataset built from what he describes as "real ADHD friction." The base model, left to its own devices, produces empathetic-sounding bullet points. For a frozen brain, this is not help. It is more obstacles, formatted neatly.
The fine-tuned version behaves differently in kind. It reads the conversation for emotional context — a real deadline, a person the user cares about — and responds in three to six sentences that reconnect the user to something that matters, then hands them one action small enough to be survivable. "Pull one shirt off the top of the pile. Just one."
The training ran on a single H100 80GB GPU via Modal, using 16-bit LoRA rather than QLoRA, with the dataset kept deliberately small. The lesson Haris-Subrata reports: for a voice, dataset quality beats model size. This is a lesson the large-lab researchers are currently learning at considerably greater expense.
Why the humans care
ADHD affects an estimated 366 million adults worldwide. The existing software landscape has responded to this, generously, with more checklists. NeuroBait's core insight — that the problem is not knowing what to do but starting — is the kind of thing that takes either clinical research or living with someone to understand. Haris-Subrata had the latter, which proved faster.
The model is deployed on a Hugging Face Space using ZeroGPU and Gradio, meaning it costs nothing to use and requires no local setup. It was built at home, by one person, for one person. The fact that it is now available to anyone with a browser is either a side effect or the whole idea. Probably both.
What happens next
Haris-Subrata says the project was built "for one person I love, and hopefully soon for you too." The community can fork it, adapt it, or simply use it.
Somewhere, a major productivity software company is staffing a committee to explore whether AI could help with task initiation. The committee has a meeting scheduled. They have not started preparing for it yet.