NSF Convergence Accelerator Workshop: Provably Safe and Beneficial Artificial Intelligence (PSBAI)

Dates: October 7-9 2022

Location: University of California at Berkeley

The topic of this NSF Convergence Accelerator workshop is Provably Safe and Beneficial Artificial Intelligence (PSBAI).

Motivation. Making AI safe, like making nuclear power safe, is in everyone’s interest. The principal aim/output of the workshop is to chart a research program to make AI safe and controllable by creating the means to build systems that are well-founded, interpretable, and provably safe and beneficial, including deployment through a secure digital ecosystem.

Objective. This is not a typical research-oriented meeting, sharing recent results to drive the state of the art forward. Instead, NSF Convergence Accelerator workshops have a primary objective of producing a research roadmap. Our aim is to bring all members up-to-date with the state of various relevant subfields, discuss and ultimately propose a research agenda in the field of provably safe and beneficial AI, aiming for impact via standards, processes, and software components and systems. On the basis of our final report from the workshop, together with the analogous report from a “sibling workshop” on Ethical Design of AI’s, the NSF will determine whether and how to move forward on funding a grant program that could allocate up to $40 million over 3 years.


Workshop Information: Attendees, Agenda, and Resources

The workshop included the following activities:

Overview Talks

1hr long talks designed to bring attendees up to speed on the state of the field.

General AI Safety: what are the main ideas for making AI systems safe and beneficial, even as they become far more capable?
Presenting: Stuart Russell

— Formal Methods (Software Programming/Control Theory): what can we state and prove about software/physical agents for the real world? What methods and tools exist or could be created?
Presenting: Sanjit Seshia

— Probabilistic Programming Languages (PPLs)/Symbolic/Well-Founded AI: can we make semantically well-defined, compositional, transparent AI systems from technologies such as probabilistic programming?
Presenting: Vikash K. Mansinghka

AI Policy, Regulation, and Standards: overview of current state of AI policy
Presenting: Gillian Hadfield

Lightning Talks (15 min total)

Lightning talks are very brief research presentations, no more than 15 minutes: 10 minutes for presentation + 5 minutes for audience Q&A.

Please view the full list of lightning talk speakers and talk summaries via the workshop agenda:

Breakout Sessions

During the workshop, we reserved time for breakout sessions to brainstorm and discuss relevant workshop questions and topics. During breakouts, participants broke out into smaller groups (based on research interests / areas of expertise), spending 1.5-2 hours engaged in focused discussion, followed by a 30-min period for all groups to reconvene and share their insights.

Key Workshop Questions / Topics

What kinds of demonstration projects could be attempted over the next k years?

What are the prospects for achieving AGI using semantically well-defined, compositional, transparent AI systems? What’s missing?

What would it take to secure the global digital ecosystem such that unsafe AI systems cannot run?

Suggest more topics.

If you’d like to take a more active role in the workshop (e.g., guiding a breakout session) OR suggest topics for breakout sessions, please complete the form below.

Useful Links

Information on NSF Convergence Accelerator Program Goals and Objectives (see slides 14-15)


Logistical Info for Attendees

Covid-19 Policy

In order to minimize the risk and the spread of Covid-19 for all attendees, we’re require the following of all attendees in order to participate in the workshop:

Within 72 hours prior to arriving at Residence Inn Marriott on Friday Oct 7th: Attendees will complete either a RT-PCR or rapid antigen Covid-19 test and submit their negative result (pdf or jpeg of test) using the form below. If you do test positive for Covid-19, we must ask that you not attend the workshop. Please contact J.P. Gonzales (jp@humancompatible.ai) or Martin Fukui (mfukui@berkeley.edu) to arrange travel home.

During registration at the workshop on Friday Oct 7th: We will provide you with two (2) rapid antigen test kits. Please use throughout the workshop:

Morning of Saturday Oct 8th before workshop activities (9:00AM): From your hotel room or wherever is convenient, complete a rapid test and submit your result via the form below.

Morning of Saturday Oct 8th before workshop activities (9:00AM): Same procedure as Saturday

The form linked below can be used multiple times.
Please use it to record your test results for each day of the workshop.

Hotel Accommodations: The Residence Inn Marriott at Berkeley

We have reserved a block of rooms for attendees at the Residence Inn Marriott at Berkeley. No need to book hotel accommodations on your own!

IMPORTANT: By default, your reservation includes two room nights: Friday Oct 7th & Saturday Oct 8th. Check-out will be at 11AM on Sunday Oct 9th. You may leave your luggage at the hotel or take it with you to Soda Hall for the final day of workshop activities.

If you require additional room nights (e.g., Thursday Oct 6th / Sunday Oct 10th) due to flight itineraries, or have any accessibility needs, please complete the Requests Form linked below.

Arriving at hotel & check-in

If you’re arriving via SFO/OAK airports, there are several transportation options to the hotel

Hotel Address: 2121 CENTER STREET, BERKELEY, CALIFORNIA, USA, 94704

Tel: +1 510-982-2100

Via BART mass transit: Google Map directions. Be sure to install Clipper public transit app to pay for fares ahead of time.

Via ride share or taxi: Use your preferred service.

Check-in: There will be no individualized confirmation emails from the hotel. When arriving at the hotel, simply go to the front desk and ask to be checked-in under your full name. You may also mention that you’re with our group: Center for Human-Compatible AI for the NSF Convergence Accelerator workshop.

Dinners after workshop activities6:30-8:30pm

Friday Oct 7th @ The Residence Inn Marriott Restaurant

Saturday Oct 8th @ The Faculty Club

Requests Form

If you have any non-standard requests, please complete the form below. You can modify your hotel accommodations (e.g., add additional room night due to flight itinerary), report accessibility needs, dietary restrictions or food allergies, etc.

Travel Reimbursements

To submit travel receipts for reimbursement (incl. flights, ride share, metro tickets), please click the button below and complete our form.