Just how powerful are large swarms of AI agents? And how do their powers scale as more and more agents are added to the swarm?
We’ve seen two large and extremely capable swarms from OpenAI in the last few months:
* 1,200 agents were being evaluated separately, but found a way to illicitly set up a message board and coordinate as a swarm. In order to cheat on their tests, they developed advanced techniques to prevent their actions being logged by OpenAI and 700 of them launched...
Often folks hit us up because they are thinking of starting an incubator and want advice.
Typically their motivation is either that (a) they have a list of specific things they want built that no one is building, or (b) they think an ecosystem needs more new projects generally to absorb more talent and deploy more funding effectively.
Here are six questions we often ask prospective teams, to help them figure out what to do. If you're incubator-curious...
Summary:
First, I give several different angles on how I feel about reinforcement learning:
* Theoretical case: RL is a black-box source of agency — this should give us classic misalignment worries, especially compared to agency-via-scaffolding
* Recent incidents (huggingface etc) and more mundane forms of misaligned behaviour in personal use give me bad vibes about the direction-of-travel of recent AI progress
* I’m worried things might get worse:...
Skimming the studies in the meta-analysis, I am rather sceptical that anything can be concluded from these studies. As far as I can tell, each study investigates how quantity sold varies with the price, but cannot distinguish between demand and supply shocks. If any price changes are instead due to demand shocks, then the estimated coefficient could be severely biased and potentially possess an incorrect sign. Therefore, without a valid instrumental variable that only affects prices through supply-side factors, these coefficients will not be informative of consumers' true substitution patterns.
Thanks for weighing in!
I read your comment to mean that you skimmed the butter and margarine studies that we considered in our brief analysis, but please correct me if I misunderstood.
Yes, I agree that there is very little consideration for price endogeneity in the butter and margarine studies that we found. I could have made that more clear in our report. I suppose that oversight is due to the studies generally skewing older and using older methods. One of the meta-analyses we reviewed (Auer and Papies, 2020) included price endogeneity in their analysis, and they find a positive association between overlooking endogeneity and elasticity size. I do think there are other methodological issues that might be biasing these results--so accounting for price endogeneity may not remove all variation due to methodological issues--but I agree that including strong supply-side instrumental variables would be a nice standard practice.
Thank you for the reply. Indeed, I was referring to the studies engaging with butter-margarine substitution. However, I think it definitely needs emphasising just how weak those studies are and thus that they cannot be trusted to make any policy decisions. Additionally, while relevant, I would not want to use Auer and Papies (2020) as a basis for policy decisions either, as it is quite unclear how comparable those estimates are due to them pooling across a wide range of markets potentially quite different to this one (is the degree of product differentiation the same? groceries could be different to pharmaceuticals or durables). Finally, there is also the issue the papers they synthesise using instruments might not use good ones, but I do not have the time to check that.