Astral Codex Ten

Dan Schwarz

14 min read

Dan Schwarz

Article

Dan Schwarz is a recurring person in the Astral Codex Ten archive, appearing 2 times across 2 issues between February 20, 2024 and May 13, 2024. The archive places it in contexts such as “including former Metaculus CTO (and Google internal prediction market veteran) Dan Schwarz”; “Dan Schwarz and the FutureSearch team say that forecasting has a “rationale-shaped hole”“. It most often appears alongside Biden, Democrats, FutureSearch.

Metadata

  • Category: People
  • Mention count: 2
  • Issue count: 2
  • First seen: February 20, 2024
  • Last seen: May 13, 2024

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

24
February 20, 2024 · Original source
How many residents will live in Prospera, a new special economic zone in Honduras, on Jan 1, 2026? Answer: 600 (80% confidence interval 100-2,000) This seems like a good guess (except that my confidence interval would have included zero because there’s a 20%+ chance that it gets shut down). So overall its forecasts seem pretty impressive. But I was concerned by its reasoning even in some of the questions it got “right”. For example, the Nikki Haley question tried to get a base rate by asking what percent of elections Haley had won before, and found she had won 71% of them - these were mostly elections for South Carolina governor. You can see what the AI is trying to do - but it’s not going to work. Then it got confused and read a lot of news stories about how she’s currently losing the 2024 presidential election, and seemed to think they were about 2028. So either the AI only got a reasonable probability by coincidence, or it was testing many different strategies, throwing out the useless ones, and updating only on the useful ones, in a way that was kind of opaque to the casual reader. Still, if the company says it beats most human forecasters, this doesn’t seem totally impossible based on what I’ve seen. And that would be exciting! An AI that can generate probabilistic forecasts for any question seems like in some way a culmination of the rationalist project. And if you can make something like this work, it doesn’t sound too outlandish that you could apply the same AI to conditional forecasts, or to questions about the past and present (eg whether COVID was a lab leak). I would be most excited if at some point this graduated from its geopolitical focus and was able to answer questions on any topic (eg “what is the chance that Astral Codex Ten gains paid subscribers this year?”), maybe if the questioner gives it links or feeds it some of the appropriate information. FutureSearch is run by a team formerly from Metaculus, including former Metaculus CTO (and Google internal prediction market veteran) Dan Schwarz. They’re looking for potential clients and/or investors; if you’re interested, email hello@futuresearch.ai. Vitalik On AI Prediction Markets Vitalik Buterin, Ethereum-founder-turned-cryptocurrency-public-intellectual, has a blog post on The Promise And Challenge Of Crypto + AI Applications. One of them is a prediction market. He writes: Prediction markets have been a holy grail of epistemics technology for a long time; I was excited about using prediction markets as an input for governance ("futarchy") back in 2014, and played around with them extensively in the last election as well as more recently. But so far prediction markets have not taken off too much in practice, and there is a series of commonly given reasons why: the largest participants are often irrational, people with the right knowledge are not willing to take the time and bet unless a lot of money is involved, markets are often thin, etc. One response to this is to point to ongoing UX improvements in Polymarket or other new prediction markets, and hope that they will succeed where previous iterations have failed. After all, the story goes, people are willing to bet tens of billions on sports, so why wouldn't people throw in enough money betting on US elections or LK99 that it starts to make sense for the serious players to start coming in? But this argument must contend with the fact that, well, previous iterations have failed to get to this level of scale (at least compared to their proponents' dreams), and so it seems like you need something new to make prediction markets succeed. And so a different response is to point to one specific feature of prediction market ecosystems that we can expect to see in the 2020s that we did not see in the 2010s: the possibility of ubiquitous participation by AIs. AIs are willing to work for less than $1 per hour, and have the knowledge of an encyclopedia - and if that's not enough, they can even be integrated with real-time web search capability. If you make a market, and put up a liquidity subsidy of $50, humans will not care enough to bid, but thousands of AIs will easily swarm all over the question and make the best guess they can. The incentive to do a good job on any one question may be tiny, but the incentive to make an AI that makes good predictions in general may be in the millions. Note that potentially, you don't even need the humans to adjudicate most questions: you can use a multi-round dispute system similar to Augur or Kleros, where AIs would also be the ones participating in earlier rounds. Humans would only need to respond in those few cases where a series of escalations have taken place and large amounts of money have been committed by both sides. This is a powerful primitive, because once a "prediction market" can be made to work on such a microscopic scale, you can reuse the "prediction market" primitive for many other kinds of questions: Is this social media post acceptable under [terms of use]?
Inline links: FutureSearch, hello@futuresearch.ai, The Promise And Challenge Of Crypto + AI Applications, back in 2014, played around with them extensively, more recently, Polymarket, tens of billions on sports, LK99, Augur
24
May 13, 2024 · Original source
People changed their minds a little over time, but not in a very consistent way that mattered much in the end. What was the “client feedback”? The report says: Client feedback was provided to the Superforecasters on December 21. The client posed questions to the Superforecasters about their assessments up to that date and asked for their reactions to several studies and articles. In the days following the client engagement, the Superforecasters lowered their confidence in the natural zoonosis hypothesis from 73% to 67%, although zoonosis remained the most likely potential cause in their assessment. But following an active engagement with recent genomic studies and historical base rates of zoonotic spillovers, those numbers began to return to earlier levels. January also saw increased attention to the geopolitical context and transparency issues, particularly related to research activities in Wuhan Is this bad? I’m imagining a pro-lab-leak client saying “But what about [this list of pro-lab-leak arguments]?” and then the superforecasters read them and adjust. In one sense, it’s good that they got to see more arguments; on the other, it seems like a potential route by which clients could bias the results - probabilities never quite got back to where they were before the feedback, though they got pretty close. The last-minute spike for zoonosis might be the Rootclaim debate results, which were released on 2/18. So maybe the client feedback and the Rootclaim results both slightly affected the numbers, but mostly the superforecasters started out pro-zoonosis and stuck to their guns. Dan Schwarz and the FutureSearch team say that forecasting has a “rationale-shaped hole”. Despite the report making this sound like a pretty intense process, we don’t get much information about details: In their extensive discussions , Good Judgment’s Superforecasters assessed base rates and historical patterns, existing evidence and scientific analysis, geopolitical context and transparency concerns, trust in intelligence communities, and methodological constraints. 1. Base Rates and Historical Patterns: The Superforecasters frequently referenced base rates, i.e., the history of pandemics emerging from natural zoonosis versus the history of laboratory leaks, to anchor their probabilities. For the former, they discussed how the base rates are changing as the climate warms and as expanding human populations push farther into natural environments that previously saw little human presence. For the latter, they acknowledged that it has only been 12 years since the advent of CRISPR gene- editing tools, and the base rate of lab leaks in the short synthetic biology era is not yet well established. 2. New Evidence and Scientific Analysis: Throughout the period, the Superforecasters adapted their forecasts in light of new scientific evidence, including genomic analyses of SARS-CoV-2 and its relation to bat viruses, and the debate over potential laboratory manipulation. 3. Geopolitical Context and Transparency Concerns: The geopolitical implications of the virus’s origins, particularly in relation to China’s transparency and the involvement of international research institutions, played a significant role in the analysis. Concerns over data veracity, and over the political ramifications of determining that the pandemic’s origins were other than zoonosis, were extensively debated. 4. Trust in Intelligence: Commentary on trust in intelligence communities and discussions about the impact of geopolitical biases on the interpretation of evidence illustrated the complex interplay between science, politics, and human behavior in assessing the pandemic’s origins. 5. Methodological Critiques and the Evaluation of Evidence: The Superforecasters engaged in methodological critiques of the evidence base, including the scrutiny of laboratory practices and biocontainment levels [...] In the end, most Superforecasters were in rough agreement on issues like the base rates of zoonotic spillover. Where they most often disagreed was on the interpretation of actions by Chinese officials and whether their actions reflected how an authoritarian government would react in any crisis over which it did not have full control, or whether those actions were indicative of attempts to cover up a biomedical research-related accident that allowed the SARS-CoV-2 virus to enter circulation in China and, ultimately, the entire globe. Probably it would be too much to ask for to get a transcript of all their discussions - then they’d be nervous saying things that might make them look bad to an audience. What would be a good balance between getting more information and not imposing on their time? Forecasting is an unusually legible and easy-to-judge domain. One of the theories of change for forecasting was to use it to identify smart people with good reasoning, then turn them loose on less well-behaved problems. This is one of the first big attempts to do this at scale. How did it work? We can’t tell, because it’s inherently an illegible and hard-to-judge domain. Darn. I don’t know what I expected. Notes From A Local Optimum Austin’s concern - that forecasting has reached a local optimum - is widely shared. We have some good sites: Manifold, Metaculus, Polymarket, GJO, etc - all doing good work. We have good-ish probabilities for a few important questions. Every so often a news source cites them. Sometimes a decision-maker looks at them behind the scenes, maybe. Is this all there is? The FutureSearch team says the next step is to focus on “rationale”. We need to use forecasting not just to get a raw probability, but to explain what’s going on and why we think something. Then instead of just convincing policy-makers to trust forecasts, we can tell them why something is true, or inform their discussions even if they’re not willing to blindly trust a number. Is this a betrayal of the forecasting ethos? The original dream was that instead of a bunch of people giving arguments, we could just test who was right. Now we’re going back to the arguments? People have argued forever; what does forecasting add to that? Well, they add the knowledge that the arguments are from people who have been right a lot before and are incentivized to be right again. Still, it’s not a natural fit. Probably it’s relevant here that FutureSearch’s forecasting AI does a really good job of this by default, in a way humans can’t match. Nuno’s yearly forecasting roundup doesn’t have a single thesis, but the first part is a well-supported complaint that most forecasting sites aren’t good business. They either burn VC money, burn EA donations, or converge towards casinos to support themselves. He gives an honorable exception to Cultivate Labs, which sells prediction market software rather than the results themselves. Open Philanthropy (billionaire Dustin Moskovitz’s EA-aligned charitable foundation) has at least given forecasting a vote of confidence, recently choosing to promote it to one of their main donation areas. Still, they got a lot of pushback on the decision, for example SuperDuperForecasting here: This will be a total waste of time and money unless OpenPhil actually pushes the people it funds towards achieving real-world impact. The typical pattern in the past has been to launch yet another forecasting tournament to try to find better forecasts and forecasters. No one cares, we already know how to do this since at least 2012! The unsolved problem is translating the research into real-world impact. Does the Forecasting Research Institute have any actual commercial paying clients? What is Metaculus's revenue from actual clients rather than grants? Who are they working with and where is the evidence that they are helping high-stakes decision makers improve their thought processes? Incidentally, I note that forecasting is not actually successful even within EA at changing anything: superforecasters are generally far more relaxed about Xrisk than the median EA, but has this made any kind of difference to how EA spends its money? It seems very unlikely. And Marcus Abramovich here: I'm in the process of writing up my thoughts on forecasting in general and particularly EA's reverence for forecasting but I feel, similar to @Grayden that forecasting is a game that is nearly perfectly designed to distract EAs from useful things. It's a combination of winning, being right when others are wrong and seemingly useful, all wrapped into a fun game. I'd like to see tangible benefits to more broad funding of forecasting that seems to be done in t he millions and tens of millions of dollars. I would also be the type of person you would think would be a greater fan of forecasting. I'm the number one forecaster on Manifold and I've made tens of thousands of dollars on Polymarket. But I think we should start to think of forecasting as more of a game that EAs like to play, something like Magic the Gathering that is fun and has some relations to useful things but isn't really useful by itself. Eli Lifland has a long and hard-to-summarize comment here, response from Ozzie Gooen here, podcast between them on “Is Forecasting A Promising EA Cause Area?” here. I’m split on this. My previous hope was that the field would gradually grow, without any qualitative changes or discontinuities, until it became big enough that journalists and policy-makers were aware of it and took it seriously (compare eg the growth of the Internet as a scholarly resource). I think the strongest argument against this is Manifold’s relatively flat user numbers. Is there a new hope? I think if nothing else, forecasting might be useful as a testing ground: First, to create forecasting AIs (like FutureSearch) which can then get consulted on a variety of questions, eg by policy-makers. The biggest holdup has always been the need to gather 20 or 50 or however many hard-to-find superforecasters for whatever question you’re asking, and then trust their advice even though they’re fallible fleshbag humans. If you can use the 20 to 50 superforecasters to inspire an AI, and then test the AI and prove it’s good, people might be more interested. This is especially true if the AI can branch out beyond traditional forecasting questions. Once we have a few of these, we can start comparing the next generation of AIs to the previous generation, and skip the superforecasters.
Inline links: forecasting has a “rationale-shaped hole”, Nuno’s yearly forecasting roundup, Cultivate Labs, choosing to promote it to one of their main donation areas, here, superforecasters are generally far more relaxed about Xrisk, here, @Grayden, here, here, “Is Forecasting A Promising EA Cause Area?” here