F26 PI Deliverable + Interview
Background
The T4SG Research Initiative runs in parallel to our client-project model. Instead of building a product for a partner org, a research team of SWEs/researchers and 2 leads pursues original social good work aimed at a peer-reviewed workshop or conference submission in the winter window.
As a Research Lead, you are responsible for turning an open research question into work that a team of part-time undergraduates can actually execute. There are two main steps: defining a problem that is both genuinely open and genuinely finishable in a semester, and decomposing it into experiments and weekly tasks your team can run. This deliverable will let you showcase both.
Leads submit their own proposals, which are ranked alongside client projects during project selection.
Deliverable
This is not a task that’s intended to take unreasonable amounts of time — no need to spend more than 4-5 hours completing the prompt.
Context
A major part of a Research Lead's job is deciding what to work on and then compressing it into something achievable. To do this you need a real grasp of where the field currently is, what's genuinely unresolved, and how a question breaks into experiments that six people can run in parallel.
Instructions
Part 1: Research Proposal
< 2 hours
Write a proposal, 1-3 pages, following the section structure of the sample proposal:
- Problem statement & research area overview: frame the open question and make clear what is not currently known
- Overview of the project: Basically what the team builds or runs, concrete enough that the research direction and possible work is clear to project members
- Research question(s)
- Impact: why this matters beyond the CS, how this ties to Social Good
- MVP0 / MVP1 / MVP2 — certainly achievable / realistically achievable / ambitious
- End deliverable and publication goals — name a specific venue and its approximate deadline
Some possible themes:
- AI Safety
- Actionable interpretability
- Evals, monitorability, robustness, red teaming, etc.
- Data filtering in pretraining
- Computational neuroscience
- Bio/med, computational biology
- Computational social science
- Climate
- Cybersecurity/cryptography
Part 2: Short Literature Review
< 1.5 hours
Write a short annotated literature review (~1-1.5 pages) that establishes the gap your project sits in.
For each paper, one or two sentences on what it established and one on why it matters for your proposal. We care much more about depth on a few genuinely relevant papers than breadth across many.
Part 3: First Three Weeks
< 1 hour
Now that you have a proposal and know where it sits in the literature, how does your team start?
What are the first two or three experiments you'd run, and what would each one tell you?
What has to happen early to unblock later work (infrastructure, baselines, data)?
What can run in parallel across 8 people, and what can't?
Describe at a high level what weeks 1-3 look like. Bullets or a short paragraph is fine.
Then, write one concrete weekly deliverable for one of your SWEs. Feel free to use the weekly deliverable template you used in your past T4SG teams. Pick any piece of the work and write the assignment you'd actually send. Assume this one is going to someone who has never trained a model or read a paper end to end.
Submission
Send an email to engineering@t4sg.dev subjected:
[F26 Lead Application] - ROLE - First and Last Name - Submission
During the Interview
How will I present my deliverable?
During the interview, you'll be asked for:
- a walk-through of your research proposal (5-10 mins)
- a walk-through of your literature review (5 mins)
- a walk-through of your first three weeks and the SWE deliverable (DOEs will act as the SWE you assigned it to) (5 mins)
Take this time to explain:
- justification for your decisions (problem choice, why its significant)
- feasibility of each piece: how long do you think it takes? how many people on it?
- anything else you'd like to point out!
What else will the interview involve?
- Follow-up questions about your deliverable (5-10 minutes)
- Expect these to be adversarial about scope, novelty, and impact
- Behavioral questions (5-10 mins)
- Logistics (time commitment, next steps)
- Questions from you!