Hi Sam! I want to apologize for taking so long to respond. I'll try to be quicker in the future if there is more to discuss after my response here. I also really appreciate you taking the time to respond in such detail. Here are some poorly organized thoughts:

1. I appreciate you outlining the specific use cases you have for AI in this space. I certainly like the idea of bots that automatically provide factual information. I'm not sure 100% automation will be possible without a good deal of regrettable or borderline false positives or hallucinations, but it could at least automate surfacing misleading information with high visibility. I think the social media use case makes a ton of sense, but it's less clear to me that this won't be solved by a private company.
2. Thanks for clarifying the limitations of fine-tuning! I hadn't realized that. I'm still a little unsure about the feasibility of training your own models to be capable AI chatbots, though. How big of a non-speciesist corpus do you expect to be able to assemble, and how does it compare to the size and quality of the training data from foundation models? You might have found a way of amassing a ton of data, in which case, kudos. If not, however, I wonder if it would make more sense to focus on developing techniques to clean speciesist data out of corpuses, which could be used by the bigger AI labs. Overall, I'm still not entirely sold on the theory of change with AI powered chatbots, but I'm also not sure exactly what kind of world we're heading into. My sense is that people mostly change their habits/views out of self-interest or under influence from their close peers. I do think personalized messaging/ads could improve the quality of outreach efforts, but where do you get the prerequisite data on the individuals?
3. I also like your point that advocacy is most impactful when you can point to specific solutions. My sense is that suggesting techniques for pruning their corpuses of speciesist data is more tractable than showing them presumably less capable models trained on a different dataset. This paper might also be inspirational.

Thanks for working on this Sam! I'm excited to see where your work goes.

A couple of questions:

1. What do you think is the single biggest impact initiative that Open Paws could focus on? How much higher impact do you think it might be than the second?
2. What are the AI use cases for animal advocates you are most excited about (ideally defined in relation to a specific task/process)? How much do you think it improves the quality or efficiency of this task? How much do you expect an animal activist fine-tuned LLM to improve on the quality/efficiency gains compared to base LLMs?

Some other comments are:

1. I think it might be useful to narrow in on a small, well-defined set of problems/tasks to focus on, and then maybe expand from there. It's hard to solve for many things at once, and if you think some problems are higher impact than others, often not worth doing so.
2. I feel somewhat skeptical of the 10% increase in productivity over base LLMs claim. 
   
   1. First, I don't feel like you provide a lot of evidence for assertions like "An AI trained on the collective knowledge of the animal rights movement would be extremely effective in crafting highly personalized and persuasive messages on animal issues." Even just examples demonstrating the difference in quality would be helpful. From a theoretical perspective, I don't clearly see why fine-tuning on animal specific text would be a lot better than prompt-engineering, since most of these models were probably trained on most public animal advocacy text anyways. You mention the cases of external facing chatbots, but how much are these used, and what would be the flowthrough effect on impact from improving their quality?
   2. Secondly, overall productivity gains from improvements in the productivity of a subprocess are diminishing (for a technical overview, see this paper). We can maybe write quicker copy and code for animal advocacy, but our impact might still be bottlenecked by things like the number of investigations we can do or how many staffers know congresspeople, which feel harder to solve with LLMs. The claim may well be right, but I think it could be better supported.
3. I also feel unsure that animal advocates are likely to have more advocacy success than other groups. We do have some allies working at the bigger players, but my sense is that they have to more careful about their EA association after the Sam Altman debacle. Changing their policies to consider animals would be a clear signal of this.

Thanks for writing and sharing! This was a useful debunking of something I hear a lot too.

Tipping points may be overrated, but I still think they can be appreciable force. Here are two reasons why:

1. New information can lead to sudden and dramatic updates in societal opinions/actions. I'm not well read on this literature, but Leonardo Bursztyn's work on social norms speaks to these issues. In one study, he finds simply providing information on Saudi Arabian men's actual attitudes towards women working outside of the home increases their likelihood to accept services to help their wives find work by 36% (9 percentage points). In another study, he and his coauthors find that the share that the number of people willing to publicly donate to an anti-immigration group rose from 20% to 33% when given information that Trump won the county in question (as opposed to Clinton).
2. Legislative change occurs at thresholds. As a result:
   1. Legislative thresholds naturally form "tipping points" where big changes cascade from smaller changes in public opinion.
   2. Legislative change itself abruptly provides new information about societal values, fueling further change in public opinion.

I think these kinds of dynamics underlie the rapid legislative victories outlined in the first chart in your blog post, although Loving vs. Viriginia is an interesting counterexample).

What I do find lacking from a lot of tipping point discussion is more consideration of the underlying mechanism supposedly fueling the self-propelling tendency. In new technology adoption, it is often more clear ― increasing returns to scale on the production and consumption side are a big one. For social norms, for me at least it's often less clear how, where and why beliefs propagate in society.