Highly recommend the video walkthrough by @jungofthewon : youtube.com/watch?v=mJXZuisO…

Q: How will Elicit have impact if AI timelines are short? A: The high-level picture I have in mind is that the world will become less chill and more fast-paced, with increasing chaos in many areas. Everyone realizes that AI is going to be a transformative force, and that they need to embrace it. I worry that things will go wrong—not for any particularly interesting reasons, but simply because people make bad decisions under time pressure in rapidly changing circumstances. When I consider how Elicit can improve the situation, it’s primarily by helping people make better decisions amid the chaos. People will use AI regardless, and eventually they'll say that they can’t make their decisions and stay competitive without AI augmentation. They will turn to tools like ChatGPT, Claude, or others. However, it’s not clear that these tools will be optimized for robust, trustworthy, high-stakes decision-making. The guarantees needed for this use case are different from those required for everyday use. The role I see for Elicit is to serve as a nonpartisan, trusted platform that helps ensure decisions are made correctly. Elicit wouldn’t rely on a single language model. Instead, it would aggregate or verify information using multiple sources, remain independent of major LLM providers, and be available for larger organizations or government agencies when it really matters. Q: You're saying that the scenario in which you think Elicit can help the most is this chaotic “people make poor decisions” scenario. What does the version of Elicit look like that helps with that? How is it different from current Elicit which mostly helps with existing information? A: Starting from where Elicit is now, users are already leveraging it to inform decisions. Deloitte is assisting the government in choosing which suicide prevention program to fund. They use Elicit as one input and then incorporate other factors—such as cost effectiveness and additional considerations—into their decision-making process. From a bottom-up perspective, a natural next step could be to absorb more of that work by understanding what users do with these outputs and ultimately converting them into decisions and support. The question then becomes: what does that require? You need some fairly straightforward data analysis capabilities. You probably also need better modeling of the world—”if we do this, then what are the consequences”. The top-down perspective is more about how to make good decisions. There is often a very structured process that we approximate: listing all possible options, assessing the likely consequences of each, and weighing the costs and benefits. This is a process that machines could execute in far greater detail than humans, considering thousands of options and exhaustively enumerating all their consequences. The bull case is to define this idealized decision-making process and then systematically scale it up using language models. Even with today’s models, you could already perform a version of this, though it will work better once the models have a somewhat better understanding of the world. Q: How would Elicit be different from future ChatGPT / Claude / Deep Research? A: The main difference is in the guarantees you can provide. I expect the guarantees from model answers to remain relatively weak. You run the system—similar to using Deep Research—and receive a report that appears to be good, assuming it didn't overlook anything. The process, however, was rather haphazard. This approach works well for many use cases, especially when you're searching for a needle in a haystack. But it is less ideal if you require a worst-case guarantee—if you need assurance that all downsides have been considered and nothing has been missed that could catastrophically derail your plan. It ultimately comes down to whether you optimize for the average case or the worst case. Part of the story will have to be people are willing to spend a lot of money on some decisions. Maybe the reason that this product hasn't been built yet is, it's really expensive to do with current models because you want to use the largest and best current models because they're barely at the point where they get this stuff right. Who is willing to pay many thousands dollars for fleshing out a decision process? Even in the future we'll want to specifically focus on that niche. Q: I wouldn’t just type my decision problem into Deep Research. I would do the breakdown I think Elicit would naturally nudge me towards. Why does Elicit win against me typing each of those steps myself into a reasoning model? A: How reliable will Deep Research be at the component steps? If it’s reliable, that’s great. If not, you’ll have to break it down further. But then at some point, you'll be doing too much work relative to how much work automation should do for you. The flipside of this is that, even if Deep Research is reliable at some problem size, as the problems you have to solve get more complex and the time available to make decisions goes down, human work for supervision & decomposition won’t scale. So either humans are doing very little work and the component pieces are already robust and everything is good, or you're doing a lot of work and at some point can’t keep up with the size of the problems you face. We most care about the world where models aren’t automatically trustworthy decision advisers for the size of problem the world is about to encounter, and so everything isn’t good by default. Q: Isn’t there a big gap between current Elicit use and the GTM required to impact big decisions? A: It's large but not insurmountable. Not many people are trying to do that thing. If you look at who's applying for FedRAMP to get things into government - there's not that much competition. The main alternative will be just using ChatGPT directly. But I think we have a good shot, especially if the need for better decision-making becomes clearer. I think the go-to-market can be accelerated a lot in that case. Maybe a one-year timeline is rough, but on a two-year timeline, it seems quite feasible to me. Q: What’s the most likely reason we fail? A: It's mostly an execution challenge. I don't think it'll be like "oh, we just had completely the wrong idea." I think it's still a very hard product to build, and we are trying to ramp this up as quickly as the quickest startups in history. That will just require hiring a bunch of people, we don't want to lower our bar. Which has always been our struggle, and so I think that will remain the struggle. In five years, we could either be like "oh no, we hired a bunch of people and totally derailed the whole thing because they didn't get it" or "we didn't hire enough people and we had good people but not enough of them." So that seems like a possible failure mode.