Maths student at Bristol. Interested in AI Safety pipelines.
Things Iâve done:
Non Trivial Fellowship. I produced an explainer of the risks posed by improved precision in nuclear warfare.
AI Safety Fundamentals. I produced this explainer of superposition: https://ââchrisclay.substack.com/ââp/ââwhat-is-superposition-in-neural-networks
Absolutely! My estimate in the summer was 4.5% (around 30 fellows), but this was excluding people at frontier labs who were explicitly on the safety teams. If specifics are important Iâd be more than happy to revisit!
Really? I suspect only a small proportion of EAs are pro-Reform UK.
yes exactly thats what Iâve heardâorgs are reluctant to accept inexperienced ops people. Iâd love to see a way round it!
Thanks for this! Youâve changed my mind
This discussion comes at a really good time for me! Iâm actually working on a project in this area.
For the past month(ish) Iâve been researching and building a post (or series of posts) about the AI Safety Pipelineâi.e. how we direct talent most effectively into and between organisations. My main goal is to get more people talking about this in publicâa goal that many people Iâve talked to seem to share!
In particular:
1. How can we improve pipelines and encourage demand for non-technical support roles (e.g. People Management and Operations) in AI Safety?
2. How can we improve resources for the âother 95%â who struggle to get roles within AI orgs but still want to make an impact?
Iâm still deciding which of these areas to prioritize for the final post. Iâve had more conversations about #1, but Iâm personally more excited by the potential impact of #2. Iâd be very grateful to hear which of these you think is more pressing.
Iâm writing to see if anyone else is thinking along these lines. If youâd like to have a conversation about the pipeline, or are even working on the same thing and would like to collaborate, please reach out to me! I have a low bar, and have spoken to people at all levels within the ecosystem.
Iâm looking to publish towards mid-august.
Quick question: are you saying that this is cost-effectively better than going vegan? Or just that if you are going to eat meat, itâs better to switch your consumption to beef?
I see the argument about the US Governmentâs statistical value of a life used a lotâand Iâm not sure if I agree. I donât think it echoes public sentimentârather a governmentâs desire to remove itself of blame. Note how much more is spent per life on say, air transport than disease prevention.
great pointâthanks you changed my view!
Interesting argumentâI donât know much about this argument, but my thoughts are that thereâs not much value in thinking in terms of conditional value. If AI Safety is doomed to fail, thereâs not much value focusing on good outcomes which wonât happen, when there are great global health interventions today. Arguably, these global health interventions could also help at least some parts of humanity have a positive future.
Unless you work in practical AI Safety/âAI policy, I disagree.
The press is very bad at talking about the scale of problems. For example, theyâre always talking about murders, which are not a pressing problem in most parts of the world right now.
The press also tends to focus on the unimportant parts of stories. For instance, when they talk about scientific papers, they often focus on a small aspect of the papers that they think readers will find exciting and blow it out of proportion.
For me also, I often donât have enough time to both read the news and read a book in the day. I think reading books are much more valuable.
Iâve heard this argument before, but I find it un-compelling in its tractability. If we donât go extinct, its likely to be a silent victory; most humans on the planet wonât even realise it happened. Individual humans working on X-risk reduction will probably only impact the morals of people around them.
Wowâof all the replies this makes the most sense to me! Thatâs a great way of looking at things!
Unfortunately as I produced those stats quite quickly in the summer, I didnât have a formal definition and was eyeballing it a bit! Itâs something I might look back into more rigorously in the future.
What I would point out is that MATS had a very similar proportion remaining in AI Safety according to their analysisâhttps://ââwww.lesswrong.com/ââusers/ââryankidd44?from=search_autocomplete. Although I agree it definitely would be surprising if the retention rate of all these fellowships was as high as MATS.