Our internships require a background in mathematics and computer science. Existing research experience in machine learning is strongly advantageous but not required. We are interested in people who can demonstrate technical excellence and wish to transition to technical AI safety research. Examples include undergraduate or Master's students in computer science or adjacent fields, PhD students/researchers, professional software or ML engineers, etc.
This internship is designed for individuals who are interested in technical AI safety research. All applicants should aim to understand CHAI’s research interests before applying. See here and here for papers from CHAI, though note that the interests of mentors shift over time. To give a more up-to-date sense of potential projects, here are examples of topics that interns worked on in the previous cohort (papers for most of these are still under preparation or review):
- Attacking vision-language models, making them output arbitrary text and other bad behaviors by feeding them adversarial input images (Luke Bailey, Euan Ong, Scott Emmons)
- Developing tensortrust.ai, an online game for collecting a large dataset of LLM prompt attacks and defenses (Sam Toyer, Olivia Watkins, Ethan Mendes, Justin Svegliato, Luke Bailey, Tiffany Wang, Isaac Ong, Karim Elmaaroufi)
- Generalizing RLHF theory to the partially observable setting, and showing how ignoring partial observability can lead to bad reward inference (Leon Lang, Davis Foote, Erik Jenner, Scott Emmons)
- An LLM interpretability benchmark that scores explanations based on how well they help predict the LLM’s behavior (Edmund Mills, Shiye Su, Scott Emmons)
- Studying multi-agent contracts and negotiations between LLM agents in Minecraft (using Voyager) (Julian Yocum, Justin Svegliato)
- Analyzing how AIs can and should deal with the fact that human preferences are changing (and can be affected by the AI) (Davis Foote, Micah Carroll, Anand Siththaranjan)
- Developing deep RL methods that enable an agent to learn how to comply with ethical theories (Mason Nakamura, Justin Svegliato)
- Developing confidence measures for LLMs to better calibrate and mitigate hallucinations (Yijin Hua, Justin Svegliato)
- Probing systemic weaknesses in LLM-based ethics models (Ethan Mendes, Micah Carroll, Sam Toyer)
General Information
- Location: In-person (at UC Berkeley) is preferred but remote is possible.
- Deadline: November 13th, 2023
- Start Date: Flexible
- Duration: Internships are typically 12 to 16 weeks
- Compensation: $3,500 per month for remote interns. $5,000 per month for in-person interns.
- International Applicants: We accept international applicants
- Requirements:
- Cover Letter or Research Proposal (choose one and see instructions below)
- Resume
- Academic Transcript
Cover Letter or Research Proposal (Choose One)
The primary purpose of the Cover Letter or Research Proposal is for us to match you to a project that interests you.
Most of our interns are generally interested in technical AI safety research but do not have a specific project in mind when they start the internship. Throughout the interview process, we learn more about each intern's interests and match them with a mentor who has an existing project idea that fits the intern's skills and interests. If you do not have a particular project in mind, then we ask you to please write a Cover Letter answering the following questions:
- Why do you want to work at CHAI as opposed to other research labs?
- What are you hoping to achieve from the internship? For example, are you seeking to improve certain research skills, contribute to a publication, test out whether AI research is a good fit for your career, or something else?
- What are your research interests in AI? For example, are you interested in RL, NLP, theory, etc?
Alternatively, some of our interns apply to the program with a specific project or detailed research interests in mind. If this applies to you, then please write a Research Proposal describing your project and what kind of mentorship you would like to receive.
Internship Application Process Overview
The internship application process has four phases. Please note: while we will do our best to adhere to them, all dates in the Internship Application Process Overview are subject to change.
- Initial Review (Phase 1)
- We will examine your application based on motivation, research potential, grades, experience, programming ability, and other criteria.
- Applicants will likely receive a response by late November.
- Programming Assessment (Phase 2)
- If you pass the Initial Review Phase, then you will be given an online programming test.
- Applicants will receive a response by late December.
- Interviews (Phase 3)
- If you pass the Programming Assessment, then you will be interviewed starting in early to mid January.
- CHAI has several mentors who are willing to take on interns. Each mentor that is interested in working with you will contact you to schedule an interview. It's possible that you will speak to more than one mentor during this phase if multiple mentors are interested in working with you.
- Offer (Phase 4)
- Applicants will receive offers by early to mid February.
- If you are given an offer by one mentor, then you will work with that mentor if you choose to take the internship.
- If you are given multiple offers from different mentors, then you will get to choose which mentor you want to work with.
- Typically, the internship will begin around April or May but the start date will ultimately depend on you and your mentor(s). You will have to coordinate with your mentor(s) on when to begin your internship.
Other Information
- For any questions, please contact chai-admin@berkeley.edu.
- In the event that your situation changes (e.g. you receive a competing offer) and you need us to respond to you sooner than you had initially thought, then please let us know.