Context:
TL;DR Jack Clark of AI lab Anthropic gave a lecture at Oxford where, among other things, he showed a slide shown in [2] which featured some of his own predictions, which was a subject of The Guardian's article in [1]. The ones above the line are used as basis for questions in this market. The questions are lightly paraphrased to save space; refer to the original slide for full phrasing and see the resolution criteria below for more info.
Notes on resolution criteria:
#1: Both parts of this will need a strong confirmation by a high-profile independent news source, i.e. we must receive a confirmation (1) that a general-purpose AI system (read: not something like AlphaFold) was used to make at least one notable discovery that had eluded human researchers prior to AI assistance, and, likely separately from that, we must receive a confirmation (2) that a general-purpose AI system was used for creating a bioweapon. Successful application of the bioweapon is not necessary, and the two confirmations don't have to refer to the same AI. Resolves YES if we get both by the end of November 2026.
#2: The AI system needs to be involved on the level of a research team collaborator, such as by generating and testing hypotheses, refining methodologies, or gathering experimental data, or any other such equal-grounds activities (i.e. it's not just used as a passive tool for search, writing text/code, or any other task that could be done manually or by other means). This should be confirmed by somebody on the research team in question or their spokesperson of choice. Resolves YES if we get such confirmation by the end of April 2027.
#3: I've taken the liberty to break this in two and only take the first half because I don't want to think about what a "human & AI company" really entails or how that is different from what e.g. Google or Nvidia is today. I expect narrowing the scope to only fully autonomous companies to make resolution easier because those would presumably have to be publicly flagged as such and have separate laws and tax regimes and so on, due to the fact that executive decisions at such companies aren't made by legal subjects (technically a very big deal; we do not have frameworks for such things to legally operate on the open market currently). At least that makes the most sense to me and is highly likely. So this needs a public and legally binding confirmation from any government body with access to this information that such an autonomous company exists and has generated at least $10m in revenue (it does not have to turn net profit since that wasn't a condition on Clark's slide). Resolves YES if we get such confirmation by the end of November 2027.
#4: Clark's description is a pretty vague "useful work in partnership with human tradespeople". Being a coat hanger is arguably useful work, even if unreasonably expensive for a bipedal robot, so for this one, I will wait for any economic analysis picked up or conducted in-house by a high-profile independent news source claiming that using a bipedal robot to assist with any trade they're used for is economically sensible in that context, i.e. provides confirmed net value over not using any (it does not have to be economically competitive with a human laborer in the same position; just net-positive on its own). Resolves YES if we get such confirmation by the end of April 2028.
#5: This refers to what is known as recursive self-improvement, the holy grail of AI research and potentially the Great Filter of our civilization. For the purposes of this market, the RSI loop is considered closed when an AI system working completely autonomously on its successor across every stage of the way (I will make exception for gathering additional training data) can demonstrably achieve statistically relevant improvements over itself without participation of human researchers outside of them acting as real-world object manipulators to e.g. replace faulty GPUs and such. It does not have to be necessarily more time- or resource-efficient or achieve better results than a team of human researchers would, as the original claim only reads "able". The resolution for this, however, is shaky because any AI lab CEO is absolutely incentivized to make that claim as soon as they see fit with nothing forcing them to back it up publicly, and we have no reason to take them at their word. A potential confirmation needs to be provided in one of the following ways:
directly, as a legally binding and publicly broadcasted testimony from the company CEO (e.g. before US Congress) with unambiguous wording;
directly, via a document issued by their respective state government after an official inquiry;
directly, via a document issued by any inter/supra-governmental regulatory body or watchdog agency involved in monitoring superintelligent AI development (if such body, e.g. a UN agency or EU commission, exists by then and is given relevant access);
directly, by documented peer review from multiple independent researchers if weights for the model in question are released publicly;
indirectly, by immediate and visible effects on the labor market such as AI labs ceasing any hiring for AI research/engineering positions, conducting mass lay-offs from such positions, or ceasing to exist in their current organizational forms altogether.
Anything else will be treated as a marketing stunt. Resolves YES if we somehow receive any such confirmation by the end of December 2028.
The market closes with the deadline for the last question on December 31st, 2028, 24:00 UTC but I'll give each individual question up to a month after their due time for confirmations to appear. In case the confirmation arrives with a delay, it must explicitly mention that the event in question took place before the deadline; otherwise, the date of the confirmation being published/released will be used for resolution purposes.
This is my first market, let me know if further clarifications are needed or if you have better ideas for the resolution criteria.
1,000Edit: Er, whoops, didn't mean to repost it here.