Doesn't #2 (pass long, adversarial Turing Test) mean being unaligned is a "necessary condition" for AGI? I fail to see how an aligned AI would fool its users into believing it's human.

I seems based on the current FAQ, this market is asking "In what specific year will I think we have achieved unaligned AGI?"

@Primer Interesting objection. Suppose we have a human-level AI that's very honest and will always fess up if you ask it point-blank whether it's human. We want that AI to pass our test but it trivially fails a Turing test if it outs itself like that. My first reaction is that I don't think it counts as unaligned if it's willing to role-play and pretend to be human. But if we don't like that, there's another solution: have the humans role-play silicon versions of themselves. Now the humans will also "admit" to being artificial and the judges can no longer use "are you human?" to clock the AI.

This is related to the genius of the Turing test. Usually I think in terms of how all the ways a Turing test can be defeated -- keying in on typing speed or typos or whether the AI is willing to utter racial slurs -- are easy enough to counter by a smart enough AI. It's not hard to mimic human typing speed, insert the right distribution of typos, or override the safety training to fully mimic a human. But you make a good point that we might not want that kind of deception. So instead we can just copyedit the humans' replies, insert delays, and instruct the humans to strictly apply the show-don't-tell rule in asserting their humanity.

In general the idea of using a Turing test to define AGI is that the judges should attempt to identify the AI purely via what human abilities it's lacking. If there are no such abilities, the AI passes the test. Any other way of identifying the AI is cheating.

@dreev Yeah, maybe we could develop a test which mitigates that problem (humans have to try to pass as an AI + fast narrow AI is used to obfuscate things like typing speed or add some muddled words if it's live speech, etc.), but that leads us pretty far from a nice and simple Turing Test, and, more importantly, there's no way such a test will actually be performed.

The only way I see is dropping any requirement of "passing a Turing Test" entirely.

Metaculus has taken a stab at this (thanks to @PeterMcCluskey for the pointer):

https://www.metaculus.com/questions/5121/date-of-artificial-general-intelligence/

In short, they require a unified system that passes all of the following:

1. A long, informed, adversarial Turing test (see FAQ 2)

2. Robotics, putting together a model car with just the human instructions

3. Min 75% and mean 90% on the MMLU benchmark

4. Min 90% on the first try on the APPS benchmark

And by "unified" they mean it can't be something cobbled together that passes each of those. The same system has to be able to do all of it and fully explain itself.

I think we want something even stronger (except the robotics part, which we're explicitly dropping for this market) but this feels like a good start.

Has such a test been performed and “trained” for? I’m pretty sure if someone fine tuned existing SOTA models for this task it could be done today, under the right conditions. Though potentially not meeting the cost or timing requirements.

I’d be curious to know if there’s been any attempts at this. If there aren’t any attempts made in a given year that would qualify for the purposes of this market, that doesn’t necessarily mean we couldn’t have met the criteria. I think it’s still valid criteria to have, but perhaps it’s less informative as a AGI indicator as a result

@MarcusM Sounds like the question is, what if we get AI that could in principle count as AGI but isn't packaged into something that meets the criteria we come up with? In that case I'm inclined to say it doesn't count yet. We wait till it meets our explicit criteria (which we need to continue to pin down!).

Are the years exact or cumulative? i.e. if we get AGI in 2025, how does the 2026 option resolve?

@VitorBosshard Exact. This is a probability density function (pdf), not a cumulative distribution function (cdf). Adding that to the FAQ now; thank you!

PS, I edited the title just now to help avoid people presuming it's cumulative and bidding "2035+" way up.

Given the subjective nature of this market, I don't think the market creator should trade in it

Note this bit from the market description (I edited it just now as well, to be clearer about outsourcing the final decision if there's no way to avoid it coming down to a judgment call):

Note that I'm betting in this market myself and am committing to make the resolution fair. Meaning I'll be transparent about my reasoning and if there are good faith objections, I'll hear them out, we'll discuss, and I'll outsource the final decision if needed. But, again, note the evolving FAQ. The expectation is that bettors ask clarifying questions before betting, to minimize the chances of it coming down to a judgment call.

You can also take a look at my other markets (or perhaps my AGI Friday newsletter) to decide how much you trust me on this. But I'll definitely be trading.

I thought Claude 3.7 identified a couple useful properties:

# AI as "Drop-in Remote Workers" Resolution Criteria

Here's a concise set of capability-based resolution criteria for determining when an AI system qualifies as a "drop-in remote worker":

1. Independent Task Completion: The AI can take a job description with minimal additional context and complete tasks autonomously without human intervention beyond initial instruction and final review.

2. Communication Competence: The AI can participate in meetings, ask clarifying questions, provide progress updates, and collaborate with team members using standard communication channels.

3. Context Adaptation: The AI can understand and adapt to company-specific terminology, processes, and culture based on limited documentation (employee handbook, style guides, etc.).

4. Self-Improvement: The AI independently identifies knowledge gaps and takes initiative to acquire necessary information to complete tasks better over time.

5. Practical Tool Use: The AI can effectively use standard workplace tools (email, chat, productivity software, internal systems) with only standard authorization processes.

6. Reliability Threshold: The AI completes assigned tasks with at least 90% success rate without requiring more supervision than an average human worker in the same role.

Resolution occurs when at least one AI system can consistently satisfy all these criteria across at least three common knowledge-worker domains (e.g., content creation, customer support, data analysis).

I think the "minimal additional context besides a job description" is a bit strong, but maybe that should be interpreted as that the AI can ask for the context it needs, as per #4

@Siebe Nice. Can we avoid getting into the weeds that much with a criterion along the lines of "do bosses hiring for remote positions generally prefer AI to humans?" or "does it generally only make business sense to hire humans because you need their physical skills or for reasons other than actual job performance, such as legal or PR reasons?"?

PS, either way, we may want to pick a threshold for the fraction of the remote workforce that should apply to. Maybe it first becomes true for the cheapest, least skilled human workers. And maybe we want to say we're hitting the AGI threshold when it's true for a median first-world college-educated native speaker?