You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
AI can write a proof, but the person reading it may still have no idea why it works. Even after checking that it is correct, they may struggle to change an assumption or use the same idea elsewhere. I want to work on that gap between having an answer and knowing enough to use it.
I am building ShareMind AI, a company focused on research into how people can learn and use knowledge from AI. For this project, I want to build and test a small system that studies a piece of work, finds out where a reader is stuck, and tries different ways to help. Does this help the reader understand more, or learn faster, than using an LLM in the usual way?
I would start with proof understanding. I have begun exploratory conversations with an applied mathematician, who has not yet committed to a pilot. If we go ahead, we would choose public proofs and work out what the reader wants to understand. A proof is a useful first case because checking each step can still leave you without a sense of the argument as a whole.
The system would first investigate the proof: what it depends on, how the argument fits together, and what remains uncertain. It would then ask about the reader's background and difficulties. From there it could choose an example, explain a missing concept, or ask a question to check whether an explanation helped. I want these steps to work together, so what we learn about the reader changes what the system does next.
Before comparing tools, we would decide what the reader should be able to do afterward. For a proof, that might mean explaining why a step is needed, working through a changed assumption, or using the idea on another problem. We would compare the system with an ordinary LLM explanation or the reader's usual way of working with an LLM. Where useful, we would also measure time spent and performance on a related task. Saying an explanation felt clear would be a separate measure.
With the minimum funding, I would focus on one case and a simple prototype. With the full budget, I would add another kind of expert work and try to reuse what we built in the first case. A third case is possible if time and recruitment allow. These would be small exploratory studies; the participants, tasks, and sample sizes are not settled yet.
I would keep data collection light. Participants could mark where they got stuck, keep brief agreed records, and talk through their experience afterward. I am not planning continuous observation or video recording.
I plan to write up the results, including comparisons that do not favor our approach, and share the evaluation methods so others can use them. Decisions about releasing code or data would come later, with the relevant permissions. I hope this work will help researchers and engineers learn difficult technical material and judge what they can actually do with it.
The minimum budget is $30,000 for about 3–4 months: $18,000 for my research time, $5,000 for expert-user pilots and evaluation, $5,000 for compute and APIs, and $2,000 for limited help building the prototype. That would cover the smaller, one-case project described above.
The proposed $75,000 budget would support about six months: $36,000 for my research time, $12,000 for expert-user pilots and evaluation, $9,000 for compute and APIs, and $18,000 for part-time research or engineering help. This would give me more time to test the approach in a second setting and build tools we can reuse across cases.
I plan to finish my PhD in Fall 2026. This work would start after graduation and once the necessary work authorization takes effect. I do not yet have a confirmed start date.
If other funding comes through, I will disclose it and adjust the budgets to avoid paying for the same work twice.
I am Hanbo Xie, a PhD researcher in Psychology / Cognitive & Neural Systems at Georgia Tech, advised by Robert C. Wilson. I study human decision-making and cognitive models. My work has also used LLMs and think-aloud data to study how people think, and explored automated cognitive model discovery and AI interpretability.
You can find my papers and background on my website.
One starting point for this project is my research on AIS, where agents investigate trained neural networks and collect evidence about them. This also raises the question of how to pass what they find on to a person. It gives me a place to start, but we have not yet shown that the full process proposed here improves human understanding.
I would lead the work. Robert C. Wilson has agreed to join ShareMind's advisor board. The larger budget would let me bring in part-time research or engineering help.
An ordinary LLM may already do the job well enough. If our system adds little, I would reconsider building a separate tool for that kind of work.
Each expert may need something so different that little carries over between cases. If that happens, I would narrow the work to one setting rather than keep building a general system.
People may like the explanations without getting better at the tasks or saving time. I would treat that as a failure to show improved understanding, and change the approach before expanding it.
The system may take too much time or money to investigate the material and adapt to each reader. I would count those costs when judging the results. If they outweigh the benefit, we would need to simplify the approach or stop working on that use case.
I have not received cash funding for this project or ShareMind in the past 12 months.
I have applied to Emergent Ventures for $100,000 over about nine months. The application is pending. I also have Google Cloud research credits, which are not cash and are not counted in this budget. I have not assumed they can be used for the company.
There are no bids on this project.