Thank you for this post.
The only thing about which I want to encourage more reflection is Have more “Move fast and break things” attitude [admittedly there is a bit of context I’m not edit-pasting here, but you do seem to favour this approach to a fair extent].
My gentle nudge here is based on my sense that ‘moving fast and breaking things’ can have pretty bad consequences if you’re (collaboratively) exploring AI safety research tracks that recklessly put into circulation knowledge that can be used to increase capabilities over/without safety ‘points’.
Espedair Street(Mathilde da Rui)
@author—thank you for this fascinating exploration. I’m appreciating your intuition that there might be some interesting structures there and your ability to point to their shape through this research. It seems helpful to have some sense of the ways LLM may exhibit features/behaviours/geographies that we didn’t knowingly ‘put there’ and that are not easily noticeable.
@Shayne—Can you expand on ‘Literary analysis however probably works dramatically better. Thus maybe it might be worth looking at the works of literature critics’?
From what I’m understanding of what your comment is caring for, I don’t actually think that the approach you’re suggesting here gets you richer/more meaningful findings—structuralism provides its own extremely context-specific (both in time and culture) framing of the symbolic meaning landscape it’s concerned with. By comparison, the archetypes you find in the major arcana are pretty steeped in the cultural roots that have underpinned much of the you’re proposing to scrutinise instead, and is made up of figures that would have been more foundationally influential in the substrate of today’s Western culture (especially with its combination of cosmological influences). Therefore, it seems to me from the way you’ve phrased your comment that you would propose a shallower tool, essentially, which doesn’t strike me as more helpful in making sense of how an LLM might derive a blueprint of human cosmology.
Also, if the data that shaped the embedding space was all English language text, I think we’re lying to ourselves somewhat if we’re expecting a diverse, non-Western-centric cosmology to emerge from this kind of analysis. I agree with you that using Lévi-Strauss’ frameworks as another track could still be worthwhile, but heeding the previous sentence rather than thinking that the right lens can cure a lack that’s actually in the object we’re exploring.
Finally, can you clarify what you mean by your very first observation, please?
Quick responses to the questions you asked:
What’s the % likelihood that you will use AISafety.com within the next 1 year? (Please be brutally honest)
For myself, somewhat less likely as I already have first-hand knowledge of a lot of these resources (but I’ll likely still be tempted to come to this website to jump straight through a particular URL if I know it’s easy to find through AISafety.com vs looking for the exact link in the target website).
What list of resources will you use?
Events + training, courses, projects, communities, landscape map (and recommending the reading guide a bunch + likely the donation guide)
What could be changed (features, content, design, whatever) to increase that chance?
Really nit-picking here!
- I’d find it nicer to use if the menu items were listed at the top of the landing page rather than having to scroll down
- I’d like the copy font to be more legible (my laptop isn’t super bright at the best of time so grey font causes strain)
What’s the % likelihood that you will send AISafety.com to someone within the next 1 year?
I estimate 90%—strong expectation in part due to the wide range of target audiences for whom a referral to the site could be useful + confidence in the diversity of perspectives this hub can send people towards.
What could be changed (features, content, design, whatever) to increase that chance?
- This is perhaps going to seem minor but a screenshot of what the Reading Guide landing page view looks like next to the CTA to go read it might give people I recommend the site to a taste of the breezy and interactive feel of the guide. As it is, I can imagine some people might hesitate to click through, maybe expecting a dense 200-page PDF doc that hasn’t been updated in a while and then hesitating as to what to do with it.
- Re intro video: I’d like to see a clickable screenshot of it next to/instead of the CTA. Not the newspaper headline strips, but e.g. the somewhat eerie vibe @ 00′31″
Any other general feedback you’d like to share
I have this nagging sense that it might be good to give newcomers a clear, immediate sense of the intentions of the team bringing these resources to people’s attention, right at the top of the landing page, basically.
As a first-pass suggestion: right under/next to the logo, add ‘A volunteer-curated resources hub about the (critical/essential/existential) safety implications of artificial intelligence. About us [<= link]’ - or something to that effect :)
Otherwise, I think this is a really valuable initiative and I like the site a lot! I also like that the design conveys a sense of urgency, but one that’s grounded in rational concern. Thanks, gang!