Open up a language model and see how it works.
Rolling applications. Seats are limited, so apply sooner rather than later.
LENS, the Learning to Explain Neural Systems Bootcamp, is an eight-week, part-time, hands-on interpretability cohort at the Secure AI Hub, Delhi. Weekly sessions, in person, starting 21 November 2026.
Key details
Ask what happens inside the model.
Most of what we know about language models comes from what they output. Interpretability opens the model up and asks which internal states produce an answer, and whether changing those states changes the answer.
Take a concrete question. When a model translates Hindi into Tamil, does it pass through English on the way? You can check by loading a small open model and reading its internal states layer by layer. Questions like this are where the bootcamp starts.
Research from the first week.
LENS is an eight-week, part-time, hands-on course. Every session combines reading with code you run against a live model, so you're doing research from the first week.
You'll work in a team of three. Each team picks one concept, tests whether a small open model represents it and actually uses it, and writes up what they find. By the end, every team has a result, a figure that holds up across seeds and a second model, a five-minute talk and a written report.
Two hours of teaching, an hour of office hours.
- We open with the one idea the session is built on.
- You work in pairs through three timed checkpoints in a Colab notebook. Every exercise comes with a hint and a worked solution.
- We close by setting each team's milestone for the next week.
Office hours are for debugging, questions about your team's project, and anything you got stuck on during the week.
Each session builds on the last one, so missing a week makes the next one much harder. We're looking for people who can attend every week and put in the work between sessions.
- A laptop and a Google account (we run everything in Colab)
- Willingness to build with transformers, PyTorch, scikit-learn and matplotlib
We provide compute as needed.
Your first research project, with support.
The bootcamp is built for people who code but are new to research, and who are curious about what happens inside large language models. We're here to support you toward your first publication.
Who should apply
- Undergraduate students
- Postgraduate students
- Recent graduates
- Early-career scientists, researchers, and engineers
What you need
- Comfort reading and writing Python code
- Experience training at least one neural network. Comfortable with PyTorch or a similar framework.
- Basic linear algebra and probability
Week 1 starts from how a transformer works, so you don't need any background in interpretability or AI safety.
Who this is not for
The pace and content are set for people starting out, so the bootcamp is probably the wrong fit if:
- you've already authored or co-authored an AI safety paper, or completed a research fellowship such as MATS or SPAR. We'd still like to hear from you about helping as a TA or mentor: team [AT] secureaifutureslab [DOT] [COM].
- you're doing or have completed a PhD, or have 3 or more years of full-time work experience
- you haven't written any Python yet. Start with a Python and PyTorch fundamentals course and apply to a later cohort.
- you can't come to the Secure AI Hub in person on most weekends of the program
Program Schedule
The first six weeks teach one method each. The last two turn each team's result into research.
| Week | Topic | What you do |
|---|---|---|
| 1 | How a transformer works | Walk through a forward pass and write your first hooks to read a model's internal states. Teams form and choose a concept. |
| 2 | The logit lens | Watch an answer form layer by layer, and test the Hindi-to-Tamil question yourself. |
| 3 | Steering vectors | Change a model's behavior by adding a single vector, and see what breaks. |
| 4 | Controlled evaluations | Build an evaluation with baselines and controls to measure whether an effect is real. |
| 5 | Activation patching | Find where in the model a behavior happens by swapping activations between runs. |
| 6 | Probes | Train probes on activations, then test whether the model actually uses what they find. |
| 7 | Making the result hold up | Rerun your best figure across seeds and a second model, and review another team's draft. |
| 8 | Demo day | Each team gives a five-minute talk and submits its write-up. |
Dates and topics are tentative and subject to change.
Free of cost.
We cover compute, venue, and program costs. You will still have to cover your travel and accommodation.
Application process
Applications are rolling, and we may close early once seats fill, so apply sooner rather than later.
| Step | What happens |
|---|---|
| 1. Apply (Active Now) | Fill in the application form. We'll ask you to read one of three well-known AI safety pieces and respond to it. We recommend about 45 minutes reading and understanding it, and 15 minutes writing your answer. |
| 2. Interview | A 15-minute call to get to know you more. |
| 3. Cohort Finalization | We will send out acceptance emails by Nov 14th |
| 4. Start | Session 1 at the Secure AI Hub from Nov 21st. |
What we look for
The application asks you to describe a project you're proud of and respond to one of three well-known AI safety readings. Please write those answers yourself. We strongly discourage you to write the application answers using AI. Applications with heavy AI usage will be immediately rejected.
Due to logistical constraints, we may not be able to provide feedback on individual applications.
Apply for the LENS Bootcamp.
Rolling applications, limited seats. Deadline: Nov 6th 2026 11:59 PM IST
Apply now →Questions? Write to team [AT] secureaifutureslab [DOT] [COM]