Project summary

BASE Research seeks funding to provide technical training in AI safety to Black scientists. Currently, BASE provides a 12-week online ARENA training program to 20 participants. In February, BASE will introduce a 16-week hybrid training program for another 20 participants, including 12 weeks of technical learning followed by 4 weeks of capstone work.

What are this project's goals? How will you achieve them?

The purpose of the project is to increase accessibility to technical training in the field of AI safety for people with the relevant skills and interest who face various obstacles connected to financial difficulties, limited access to computing power, or insufficient technical education. BASE plans to create technical capacity for AI safety among the Black community of AI safety practitioners. This goal will be accomplished through the introduction of the ARENA curriculum through two separate tracks:

• Community of Practice track: 12-week, part-time course held remotely on weekends for up to 20 people. The ARENA curriculum includes weekly modules and is designed so participants can learn each lesson while balancing other professional responsibilities.

• Hybrid technical training track: 16-week course that starts in February with 20 participants in the group. It involves 12 weeks of technical training and 4 weeks of capstone project work. In addition to remote lessons, there will be in-person lessons at FAR Labs.

Both courses require GPU credits to perform some technical tasks and projects. Additional expenses connected with the in-person part of the hybrid curriculum for the February group include instructor fees, venue costs, meals, and technical resources needed to complete the curriculum and capstone.

How will this funding be used?

Funding will be used for:

• GPU credits for 40 participants: $11,000

• Facilities: $6,400

• Light refreshments for in-person sessions: $800

• Teaching Assistant 1: $7,500

• Teaching Assistant 2: $7,500

• Program coordination and technical support: $8,000

• Participant stipends: $8,800

• Total request: $50,000

Given the nature of the training sessions and participants' capstone projects, GPU credits will provide computing power for compute-intensive activities such as interpretability, evaluation, supervision, and other aspects of technical AI safety that cannot practically be achieved on regular PCs.

The participant stipends will offset some of the considerable time costs incurred in training and the capstone projects. The facilities budget will help offset the cost of renting facilities to run the hybrid program, while the food budget will cover snacks during the in-person sessions.

Who is on your team? What's your track record on similar projects?

Lawrence Wagner is the Executive Director and founder of BASE Research and manages the project's execution, coordination, and delivery. He previously served as a Research Manager at Machine Learning Alignment & Transparency and Security (MATS). In that role, he planned activities, managed research, coordinated mentors, scaffolded literature, and provided weekly research reviews for AI safety researchers. Lawrence completed BlueDot Impact’s Technical AI Safety and AGI Strategy courses. He also contributed to software development projects for the U.S. Missile Defense Agency and raised $1 million for the cybersecurity apprenticeship program.

Robert Amanfu is BASE's Technical Director. He currently works as a Principal Data Scientist at At-Bay and has held machine learning and data science roles at Plaid, Stripe, Fast, and Capital One. He contributes to AI safety evaluation through the AI Safety Engineering Taskforce and Terminal Bench. So far, Robert has trained eight members of the Community of Practice through the ARENA curriculum and twenty fellows from Cohort 1, and he will oversee the program's technical delivery and support participants throughout training.

What are the most likely causes and outcomes if this project fails?

Key risks include participant dropout and a high need for support. The ARENA curriculum is not easy, and some participants may face problems with pacing, coding, and even setting up their computing environment.

To guard against these risks, BASE will ensure that:

• The curriculum is broken down into small, manageable segments

• Office hours are provided on a consistent basis

• There will be two teaching assistants to help participants

If the project does not go as planned, some participants may not finish the curriculum or may not become technically competent enough to move into AI safety research. In any case, the participants will learn the basics covered by this program.

How much money have you raised in the last 12 months, and from where?

Over the past 12 months, BASE has raised approximately $15,000. This includes:

• $11,000 through Manifund, including a $10,000 contribution from Ryan Kidd and $1,000 from Jim Chapman

• Approximately $4,000 in previous campaign donations