Iâm a software engineer on the CEA Online team, mostly working on the EA Forum. We are currently interested in working with impactful projects as contractors/âconsultants, please fill in this form if you think you might be a good fit for this.
You can contact me at will.howard@centreforeffectivealtruism.org
I think this table from the paper gives a good idea of the exact methodology:
Like others Iâm not convinced this is a meaningful âred line crossingâ, because non-AI computer viruses have been able to replicate themselves for a long time, and the AI had pre-written scripts it could run to replicate itself.
The reason (made up by me) non-AI computer viruses arenât a major threat to humanity is that:
They are fragile, they canât get around serious attempts to patch the system they are exploiting
They lack the ability to escalate their capabilities once they replicate themselves (a ransomware virus canât also take control of your car)
I donât think this paper shows these AI models making a significant advance on these two things. I.e. if you found this model self-replicating you could still shut it down easily, and this experiment doesnât in itself show the ability of the models to self-improve.
I just sent out the Forum digest and I thought there was a higher number of underrated (and slightly unusual) posts this week, so Iâm re-sharing some of them here:
Detection of Asymptomatically Spreading Pathogens by @Jeff Kaufman (I think the slide-show format works really well)
The anti-clickbait titled Next-gen LLINs with chlorfenapyr by @Scott Smith, which analyses the effect of using chlorfenapyr rather than pyrethroid as an insecticide in bednets. This reduces the effect of insecticide-resistance
I donât work on the EAG team, but I believe applications havenât opened yet because the exact date and location havenât been decided (cc @RobertHarling)
Iâm curating this post. I really enjoyed the argument-rebuttal format used here, and it does a great job of tiling out the common flavours of PAV arguments.
I think itâs a shame the Nucleic Acid Observatory are getting so few votes.
They are relatively cheap (~$2M/âyear) and are working on a unique intervention that on the face of it seems like it would be very important if successful. At least as far as Iâm aware there is no other (EA) org that explicitly has the goal of creating a global early warning system for pandemics.
By the logic of it being valuable to put the first few dollars into something unique/âneglected I think it looks very good (although I would want to do more research if it got close to winning).
Ah, I hadnât thought of that, and I can see how this makes the results indeterminate (because reallocating the votes from one joint-last candidate could bump the other joint-last candidate up from the bottom).
Iâll have a think about how to handle this and get back to you, my initial thought is still to break ties randomly (with a stable-but-random ranking of the precedence of each candidate in a tie).
(Discussed separately) I think it would be best to split the pot 4 ways if this happens, because there is some chance of introducing a bias by deciding when to end based on a property of the votes. Or if there is some reason we canât do this that Iâm not aware of (e.g. legal constraints), then breaking the tie with a coin flip.
(@Lorenzo Buonannođž You can consider this the official answer unless I hear otherwise).
Iâm curating this post. This was my favourite post from Funding Strategy Week. It makes a straightforward but important point that is useful to keep in mind.
Yup new accounts can downvote immediately, unlike on LessWrong where you need a small amount of karma to do so. Canât confirm whether this happened on this post
Iâm curating this post (and by extension, part 2). These posts are very clearly written, and include a lot of detail like:
Figure 2[1] on the lifecycle of viruses within the body
Which viruses we might be able to detect in blood
In part 2, an overview of where collected blood and plasma comes from in the US
As someone who is not familiar with this topic, this was a great concise introduction.
I stumbled across this via Trendlines in AIxBio evals, also by @ljusten. This authorâs contributions seem highly underrated to me.
See the doc linked in the quick take for our thinking on this. These are the main reasons from there (ones below number 3 are not that important imo), upon reflection I would now swap 2 and 3 in terms of importance:
You can browse past editions with a much nicer + more familiar UI. I would guess this would convert a lot more people per impression than a signup box with no context
We want to start crossing over with substack in various other ways (getting authors from there to crosspost on the forum), so it would be useful for us (Toby) to become more familiar with how the platform works, particularly in terms of social dynamics rather than features per se. E.g. I donât really understand whether most people discover newsletters on substack itself vs being linked from elsewhere
Substack has a good recommendations algorithm, which will hopefully recommend people other EA relevant content (this feels complementary with the thing above, where itâs facilitating some cross-flow of users between our owned channels and substack)
Iâm curating this post. The facets listed are values that I believe in, but that are easy to forget due to short term concerns about optics and so on. I think itâs good to be reminded of the importance of these things sometimes. I particularly liked the examples in the section on Open sharing of information, as these are things that other people can try and emulate.
Thanks for the comments everyone. This wonât be done in time for this weekâs edition, but hopefully the next one
do they now
They do⊠as an exercise try and find them
Maybe Iâd advocate for putting them all somewhere CEA owns as well in case substack stops being the right place.
Seems reasonable đ
Weâre thinking of moving the Forum digest, and probably eventually the EA Newsletter to Substack. Weâre at least planning to try this out, hopefully starting with the next digest issue on the 23rd. Hereâs an internal doc with our reasoning behind this (not tailored for public consumption, but you should be able to follow the thread).
Iâm interested in any takes people have on this. Iâm not super familiar with Substack from an author perspective so if you have any crucial considerations about how the platform works that would be very helpful. General takes and agree/âdisagree (with the decision to move the digest to Substack) votes are also appreciated.
Separate from your comment, I have seen comments like this elsewhere (albeit also mainly from Bob and other RP people), so I still think itâs interesting additional evidence that this is a thing.
It seems like some people find it borderline inconceivable that neuron counts correspond to higher experience intensity/âsize/âmoral weight, and some people find it inconceivable the other way. This is pretty interesting!
Iâm one of the people who agreed with @titotalâs comment, and it was because of something like this.
Itâs not that Iâm worried per se that the survey designers will write a takeaway that puts a spin on this question (last time they just reported it neutrally). Itâs more that I expect this question[1] to be taken by other orgs/âpeople as a proxy metric for the EA communityâs support for hits-based interventions. And because of the practicalities of how information is acted on the subtlety of the wording of the question might be lost in the process (e.g. in an organisation someone might raise the issue at some point, but it would eventually end up as a number in a spreadsheet or BOTEC, and there is no principled way to adjust for the issue that titotal describes).
And one other about supporting low-probability/âhigh-impact interventions