Yes, this was a cheeky or sarcastic comment. I wrote it to share with some colleagues unfamiliar with AI safety who were wondering what it looked like to have ‘good’ outcomes in AI policy & governance.
Alexander Saeri
Karma: 634
Good AI governance is pretty easy.
We just have to
solve a bunch of 2000+ year old moral philosophy questions (e.g., ‘what is good’, ‘what is the right action in a given circumstance’, ‘what are good rules for action’), then
figure out how to technically implement them into a non-deterministic software / algorithmic form, then
get international agreement on complex systems of regulation and governance to ensure that technical implementation is done correctly and monitored for compliance without compromising values of democracy, right to privacy, free expression, etc; then
ensure whatever governance arrangements we establish are sufficiently robust or flexible to respond to the transformative impacts of powerful AI on every part of society and industry
within the next ~5-20 years before the technical capacity of these systems outpace our ability to affect them.
Thanks for writing this up, Emily. I think your decision to do this helped me feel more secure about taking a career break of my own—including some time set aside to do no work or career planning!
I’m glad that Australia has signed this statement.
It’s worth noting that until quite recently, the idea of catastrophic misuse or misalignment risks from AI have been dismissed or made fun of in Australian policy discourse. The delegate from Australia, Ed Husic who is Minister for Industry, Science, and Resources, actually wrote an opinion piece in a national newspaper in June 2023 that dismissed concerns about catastrophic risk
In an August 2023 public Town Hall discussion that formed part of the Australian Government’s consultation on ‘Safe and Responsible AI’, a senior advisor to Husic’s department said that trying to regulate risks from advanced AI was like the Wright Brothers trying to plan regulations for a Mars colony, and another key figure dismissed the dual-use risks from AI by likening AI to a ‘kitchen knife’, suggesting that both could be used for good and for harm.
So it was never certain that somewhere like Australia would sign on to a declaration like this, and I’m relieved and happy that we’ve done so.
I’d like to think that the work that’s been happening in the Australian AI Safety community and had an impact on Australia’s decision to agree to the declaration, including
organising Australian experts to call for serious consideration of catastrophic risks from AI and make plans to address those risks,
arranging more than 70 well-researched community submissions to the ‘Safe and Responsible AI’ consultation that called for better institutions to govern risks and concrete action to address them.
A lead long-term focused policy development & advocacy organisation in Australia, Good Ancestors, also created a rigorous submission for the process.
The declaration needs to be followed by action but the combination of this declaration and Australia’s endorsement of the US executive order on AI Safety has led me to feel more hopeful about things going well.
Glad that fasting works for you! I have tried it a couple of times and have found myself too hungry or uncomfortable to sleep at the times I need to (eg, a nap in the middle of the flight).
Great points on equipment; I think they are necessary and think that the bulk of a good neck pillow in carry on luggage is justified because I can’t sleep without it. I also have some comically ugly and oversized sunglasses that fit over my regular glasses and block light from all sides.
Thanks for the post!
I’m familiar with EGMs in the spaces you mentioned. I can see EGMs being quite useful if the basic ideas in an area are settled enough to agree on outcomes (eg the thing that the interventions are trying to create)
Right now I’m unsure what this would be. That said, I support the use of EGMs for synthesising evidence and pointing out research directions. So it could be useful to construct one or some at the level of “has anyone done this yet?”
Thanks for this guide!
One thing that I appreciated when attending a GWWC event was that expectations of responsible conduct were made clear with an explicit announcement at the beginning of the event. I thought this was a good way to create a social agreement among attendees.
I think that some people are reluctant to do this because they think it might bring the mood down, or it feels awkward to call attention to the possibility of harmful behaviour at what is supposed to be a fun or professional event. They might also not be sure exactly what to say. One idea for addressing these barriers would be to provide a basic script that organisers could say, or rewrite in their own words.
Thanks for writing up this work, Zoe. I’m pleased to see a list of explicit recommendations for effective charities to consider in framing their requests for donations.
Selfishly, I’m also pleased that our paper (Saeri et al 2022) turned up in your search!
It’s be interesting to understand your motivations for the literature review and what you might do next with these findings / recommendations.
One thing that our paper necessarily didn’t do was aggregate from individual studies (it only included systematic reviews and meta-anlayses). So it’s interesting to see some of the other effects out there that haven’t yet been subject to a review.
Voted!
I was motivated to write this story for two reasons.
First, I think that there is a lack of clear visual metaphors, stories, or other easy accessible analogies for concepts in AI and its impacts on society. I am often speaking with intelligent non-technical people—including potential users or “micro-regulators” (e.g., organisational policymaker) of AI tools—who have read about AI in the news but don’t have good handles on how to think about these tools and how they interact with existing organisational processes or social understandings.
Second, this specific story was motivated by a discussion with a highly qualified non-technical user of LLMs who expressed skepticism about the capabilities of LLMs (in this case, chatGPT 3.5) because when they prompted the LLM to provide citations for a topic area that the user was an expert in, the research citations provided in the LLM response were wrong or misleading / hallucinations.
One insight that came from our follow-up conversation was that the user were imagining that writing prompts for an LLM to be similar to writing a Google search query. In their understanding, they were requesting a pre-existing record that was stored in the LLM’s database, and so for the LLM to respond with an incorrect list of records indicated that the LLM was fundamentally incapable of a ‘basic’ research task.
Thanks Jacques, I’ll need to check this out. Appreciate the pointer and keen to hear more about an LLM layer on this (e.g., identifying action items or summarising key decision points in a meeting, etc).
I had wondered if it was too hyperbolic to claim that this was an example of proto- or early-PASTA. My earlier draft hedged and said that the next version of these tools would be something like an early PASTA. I would characterise Holden Karnovsky’s post introducing PASTA as describing an agentic system that could improve by making copies of itself and improving itself.
However, when he first introduces the idea of ‘explosive’ scientific and technological advancement, it’s through the thought experiment of creating digital people, which mean that many more minds can be allocated to different research problems.I would argue that using Whisper or GPT-3 in the way I’ve described in this article is applying a kind of information processing system that in a very limited sense, is similar to allocating another mind to the research problem of capturing and analysing speech & text data—because it essentially replaced me or another researcher doing the task. This is especially the case when chaining tools together with (for now) human supervision. This allows Whisper (language processing module) and GPT-3 with prompting (summarisation and analysis module) to combine for more useful ‘mind-replacement’ than either alone.
Thanks for this Jakub! One thing I’ve seen students / participants ask for is more concrete actions that they could take alongside or after AGISF, so I think this will be a useful resource.
Thanks for the update.
I’d like to recommend that part of the process review for providing travel grant funding includes consideration of the application process timing for CEA-run or supported events. In my experience, key dates in the process (open, consideration/decision, notification of acceptance, notification of travel grant funding) happen much closer to the date of the event than other academic or trade conferences.
For example, in 2022, several Australian EAs I know applied ~90 days in advance of EAG London or EAG SF, but were accepted only around 30-40 days before the event.
A slow application process creates several issues for international attendees:Notice is needed for employment leave. Prospective attendees who are employed usually need to submit an application for leave with 1+ months notice, especially for a trip of ~1 week or longer needed for international travel. Shorter notice can create conflict or ill-feeling between the employee and employer.
Flight prices increase as the travel date approaches. An Australian report recommended booking international flights 6 months ahead of the date of travel. A Google report recommended booking international travel (US<>Europe) at least 50-180 days [~2-6 months] ahead of the date of travel. By 30 days out—when my colleagues received notice of acceptance—flights were much more expensive, and some of the most convenient travel dates were unavailable.
Fit with other commitments and needs. For parents, people with caring responsibilities, or people with disabilities, a slow process can lead to stress and conflict about whether to accept the invitation and how to make arrangements to support attendance at short notice.
Visa issues. Visa applications can take weeks or months to approve by the destination country.
Providing travel grant funding can help to “smooth over” some of these issues, e.g., by subsidising the increase in flight costs, offsetting the (literal or emotional) costs of navigating / negotiating commitments and needs. It is not a panacea—the application process itself also needs to be reviewed to reduce these issues. If the travel grant funding is significantly reduced but no change is made to the application process, there may be an unintended consequence of fewer international attendees who would otherwise be a good fit for events.
I support a review of travel grant funding processes. I ask that you also consider the application process (especially timing) and its relationship with the travel grant funding process, to improve the experience for international attendees so that the flagship events of EA Global can continue to live up to their name.
Thanks Peter! I appreciate the work you’ve put in to synthesising a large and growing set of activities.
Nicholas Moes and Caroline Jeanmaire wrote a piece, A Map to Navigate AI Governance, which set out Strategy as ‘upstream’ of typical governance activities. Michael Aird in a shortform post about x-risk policy ‘pipelines’ also set (macro)strategy upstream of other policy research, development, and advocacy activities.
One thing that could be interesting to explore is the current and ideal relationships between the work groups you describe here.
For example, in your government analogy, you describe Strategy as the executive branch, and each of the other work groups as agencies, departments, or specific functions (e.g., HR), which would be subordinate.
Does this reflect your thinking as well? Should AI strategy worker / organisations be deferred to by AI governance workers / organisations?
Thanks for the plausible explanation!
Re: adding images to your post, I literally just copy and paste. But you could also read a longer post on how to enable advanced editing features such as tables and images.
Thanks for pointing this out, Peter. As I understand it, you found this by searching for “effective altruism” and then sorting by date, not relevance.
I did not see any results for “less wrong”
But I did see similar results to your observation for “alignment forum”
Thanks for this detailed write-up, Ninell. I’ll be applying several of the principles for organisation and roles to a version of AGISF I’m facilitating in Australia in late 2022.
This was great fun, and I enjoyed contributing to it!
Thanks for your work on this! I’ve signed the letter and have shared with colleagues. I’m proud that the Australian community is able to engage lawfully and transparently with the democratic process like this, without fear of retribution. Not everyone is so lucky.