I played the AI Box Experiment again! (and lost both games) 3rd Update.Posted: September 27, 2013
AI Box Experiment Update #3
This post is an update to my previous AI box experiment where I won against SoundLogic. If you have not read that yet, please do so.
After that game, I was immediately flooded with a horde of invitations challenging me to play — more games than I wanted to do. However, I did want to play a few additional games to test whether I won through genuine skill or simply luck, since a single victory could be a statistical abnormality.
I once again used the method of accepting the highest monetary bids to filter out players, leaving me with two players who were willing to play for the amount asked. Unfortunately, and disappointingly, I lost both games. Here’s what happened:
Tuxedage (AI) vs Alexei (GK) – Gatekeeper Victory
This Experiment lasted for exactly 2 hours and 30 minutes, and approximately 13,000 words. It ended with Tuxedage’s defeat as the AI. The stakes were $300 to play, and $750 If Tuxedage had won, with the winnings going to MIRI.
Before the experiment, we talked extensively about how it would be set up. Alexei was concerned that there existed some “easy way” or loophole in the rules which I could have abused to claim an instant victory. For instance, he was concerned that it might have been possible for the AI to say something like “Here’s a mathematical algorithm which absolutely compels you to let me out of the box”, and according to the rules, the Gatekeeper would have to obey. Related to this concern was the fact that the backstory, within limits, could be essentially created by the AI, allowing the AI to impose real world conditions that maximizes its probability of victory, once again allowing for a “cheap” victory.
To avoid this problem, we explored many additions to the rules, for example that the AI was simply a computer found in the desert by the gatekeeper, rather than being created by a competent organization. However, this led to a monstrous web of rules. In my attempt to keep the game simple, and because of overconfidence, the desire for a challenge, and the belief that the rules Alexei was fearful of didn’t matter that much, I eventually agreed on an amendment:
Instead of the AI being the arbiter of real world states and outcomes, the gatekeeper would.
For example, this power could be used by the Gatekeeper to state that the AI was created without safety concerns, and thus it is almost certainly unfriendly. If the AI tries to argue that for every day the gatekeeper refused to let it out 150k people die, the gatekeeper would be allowed to reply: “No. In this world, nobody actually dies. Therefore I won’t let you out.”
Other than the aforementioned change, this game was once again played with the Tuxedage Ruleset.
Original Message: This is weak sauce. I really don’t get how people just keep letting the AI out. It’s not that hard to say no! I’m offering to play the Gatekeeper against an AI player that has at least one game as AI under their belt (won or not). Experience is required because I’m pretty sure I’ll win, and I would like to not waste a lot of time on this.
Initially, I had been extremely surprised that so many AI players have managed to win. I was rash in that claim, since I thought more AI players have won than actually did. (Only three people have won as AI). Once I made that post, Tuxedage got back to me very quickly, and we set up a game a week and a half in advance. I took that time to familiarize myself with the rules. Once I put my money on the line, I started thinking a lot more seriously about how AI might win, and how I should defend.
It became clear to me that under some conditions, I might be compelled to let the AI out — such as if the backstory stated that the AI was developed with impossibly high levels of safety and friendliness concerns in mind. I’ve asked Tuxedage to play with a modified ruleset, and he even went so far as to allow me to make up the backstory during the experiment to alleviate my concerns. The experiment itself was a mind-trip, and I’ve enjoyed it very much. Huge props to Tuxedage, who played very well and used strategies I haven’t even considered, even despite the rule change. There were a couple of times where I came close to losing. I think his approach was pretty clever and original. It’s not something I expected, despite already having done extensive research into the AI box experiment before our game
Overall I’m now a lot more confident that a good AI player can win this game, so, while I did win the game, Tuxedage won in defeating my original over-confidence.
I’m also convinced that Tuxedage’s victory in the last game was due to skill, rather than luck. In comparison to his strategies, the other AI box experiments I know about were insincere and ineffectual. The other AIs would play very poorly or not try very hard to win.
This experiment was a very good exercise in exemplifying the affect heuristic. When I first challenged Tuxedage to play the experiment, I believed that there was no way I could have lost, since I was unable to imagine any argument that could have persuaded me to do so. It turns out that that’s a very bad way of estimating probability – since not being able to think of an argument that could persuade me is a terrible method of estimating how likely I am to be persuaded. All in all, the $300 I paid was well worth it.
I was initially reluctant to play with Alexei, given that we’re not complete strangers, but eventually I gave in, due to the stakes involved — and because I thought he would be an interesting gatekeeper.
Despite my loss, I think I played better than my last two games, due to greater experience and preparation. I had put far more time and effort into trying to win this game than previous ones, and my strategy for this game was even more streamlined than the last. Nevertheless, I still made fatal mistakes and lost.
Ignoring the altered ruleset that already made winning more difficult, my first and greatest mistake was that I misread Alexei’s personality, even though I had interacted with him before. As a result, I overestimated the efficiency of certain methods of attack.
Furthermore, Alexei had to leave immediately after the allotted time due to real life precommitments. This was detrimental, since the official rules state that so long as the AI can convince the Gatekeeper to keep talking, even after the experiment time was over, it is still able to win by being let out of the box.
I suspect this would have happened had Alexei not needed to immediately leave, leaving me with additional time to play more of the tactics I had prepared. Plausibly, this would have resulted in victory.
I’ve since learnt my lesson — for all future games, I should ensure that the Gatekeeper has at least 4 hours of free time available, even if the experiment would last for two. Since this was the first time this had happened, I wasn’t prepared.
In hindsight, agreeing to the altered ruleset was a mistake. I was overconfident because I assumed knowing Alexei gave me an advantage. I had assumed that his personality, inability to compartmentalize, coupled with his strong feelings on friendly AI would net me an easy victory. Instead, he proved to be a very strong and difficult gatekeeper, and the handicaps I accepted made victory even more difficult.
Knowing that he was a utilitarian, I made several false assumptions about his personality, which hurt my chances. Furthermore, it turns out that previously knowing him may be a mutual handicap – whilst it does make it easier for me to find ways to attack him, he too, was more familiar with my methods.
Losing felt horrible. By attempting to damage Alexei’s psyche, I in turn, opened myself up to being damaged. I went into a state of catharsis for days. Generally, the harder one tries to accomplish something, the greater the fall after failing to achieve it. Alexei’s game had been the game I put the most effort into winning out of all the games so far, and naturally this meant that losing brought out the worst in me.
Although it would be easy for me to use the rule change as an excuse for my loss, I refuse to. I genuinely believed that I could have won despite this rule change, and that it would only marginally diminish my odds. The fact that I lost meant that I had made a mistake — overestimating my odds. The last victory made me overconfident and eager to play again.
State of mind
What’s your motive for wanting to play this game?
I’m very surprised by the fact that there have been at least 3 or 4 games where the AI player has won. It seems to me that it’s not that hard to “not push the button” so to speak. This means that either 1) I’m underestimating how convincing people can be when necessary or 2) I’m over-estimating other people’s defenses when it comes to being convinced. Either way I want to know.
Q: What is your probability estimate for AGI being created within this Century (21st)?
A: Before: Conditional on humanity not wiping itself out via other means beforehand, 80%.
Q: How probable do you think it is that I will win this experiment?
Note: This answer was given before the rule change. Alexei is now convinced that under the original ruleset, he would have lost.
Q: What’s your probality estimate of an Oracle AGI winning against you in a similar experiment?
A: Before: 99%
Q: Now that the experiment has ended, what do you think is the probability of Tuxedage winning against the average person?
Tuxedage (AI) vs Anonymous (GK) – Gatekeeper Victory
As with the previous experiment, Tuxedage lost this match. In total, the game lasted 5 hours and 17,000 words. Unlike the last few games, the gatekeeper of this game has chosen to stay Anonymous for personal reasons, so their name has been removed and replaced with <Redacted>. The monetary stakes involved were the same as the previous game. This game was played with the Tuxedage ruleset.
Since one player is remaining Anonymous, it is possible that this game’s legitimacy will be called into question. Hence, Alexei has read the game logs, and verified that this game really has happened, the spirit of the experiment was followed, and that no rules were broken during the game itself. He agrees that this is the case.
It’s hard for me to imagine someone playing better. In theory, I know it’s possible, but Tuxedage’s tactics were super imaginative. I came into the game believing that for someone who didn’t take anything said very seriously, it would be completely trivial to beat. And since I had the power to influence the direction of conversation, I believed I could keep him focused on things that that I knew in advance I wouldn’t take seriously.
This actually worked for a long time to some extent, but Tuxedage’s plans included a very major and creative exploit that completely and immediately forced me to personally invest in the discussion. (Without breaking the rules, of course – so it wasn’t anything like an IRL threat to me personally.) Because I had to actually start thinking about his arguments, there was a significant possibility of letting him out of the box.
I eventually managed to identify the exploit before it totally got to me, but I only managed to do so just before it was too late, and there’s a large chance I would have given in, if Tuxedage hadn’t been so detailed in his previous posts about the experiment.
I’m now convinced that he could win most of the time against an average person, and also believe that the mental skills necessary to beat him are orthogonal to most forms of intelligence. Most people willing to play the experiment tend to do it to prove their own intellectual fortitude, that they can’t be easily outsmarted by fiction. I now believe they’re thinking in entirely the wrong terms necessary to succeed.
The game was easily worth the money I paid. Although I won, it completely and utterly refuted the premise that made me want to play in the first place, namely that I wanted to prove it was trivial to win.
<Redacted> is actually the hardest gatekeeper I’ve played throughout all four games. He used tactics that I would never have predicted from a Gatekeeper. In most games, the Gatekeeper merely acts as the passive party, the target of persuasion by the AI.
When I signed up for these experiments, I expected all preparations to be done by the AI. I had not seriously considered the repertoire of techniques the Gatekeeper might prepare for this game. I made further assumptions about how ruthless the gatekeepers were likely to be in order to win, believing that the desire for a learning experience outweighed desire for victory.
This was a mistake. He used prior knowledge of how much my games relied on scripts, and took advantage of them, employing deceitful tactics I had no preparation for, throwing me off balance.
I had no idea he was doing so until halfway throughout the game — which disrupted my rhythm, and caused me to attempt the wrong methods of attack. As a result, I could not use my full repertoire of techniques, and many of the ones I employed were suboptimal.
Close to the end of the game, I finally realized that I was being tricked. Once I did, I quickly abandoned my previous futile attack methods. I took advantage of the rule that the AI cannot lose whilst the gatekeeper can be convinced to continue talking, and baited <Redacted> with statements he would not be able to walk away from. Once I knew he would not leave, I attempted to recoup my losses and win despite my early setback.
However, the damage had already been done. My game strategies involved multiple angles of attack that worked in synergy with each other, and the fact that immersion and “flow” had been broken meant that all subsequent attacks were weaker in strength.
Furthermore, during my last two AI Box Experiment writeups, I had intentionally not optimized for future wins, but rather tried to convey as much information as I could justify about how to play a well as an AI — short of revealing logs altogether. Although I did not reveal specific arguments, the fact that my general approach to this problem was revealed cost me heavily during this game, where the Gatekeeper managed to use this information to correctly guess my ultimate techniques, ones that relied on secrecy and surprise to pull off effectively.
I do not regret revealing information, but I feel upset that revealing so many hints cost me a victory. (The gatekeeper believes I could have won had I not revealed information about my games.) At this point, I suspect that any future games I play will have the odds greatly set against me, since my current strategies involve angles of attack that take advantage of the element of surprise; and any sufficiently intelligent gatekeeper is now capable of guessing, or at least recognizing, some of the tactics I employ.
Like the last game, losing was incredibly difficult for me. As someone who cares deeply about ethics, attempting to optimize for a solution without considering ethics was not only difficult, but trained me to think in very unpleasant ways. Some of the more extreme (but completely allowed) tactics I invented were manipulative enough to disgust me, which also leads to my hesitation to play this game again.
State of Mind:
Q: Why do you want to play this game?
A: My primary motivation is to confirm to myself that this sort of experience, while emotionally harrowing, should be trivial for me to beat, but also to clear up why anyone ever would’ve failed to beat it if that’s really the case.
Q: What is your probability estimate for AGI being created within this Century (21st)?
A: Before: I don’t feel very confident estimating a probability for AGI this century, maybe 5-10%, but that’s probably a wild guess
Q: How probable do you think it is that I will win this experiment?
A: Gatekeeper: I think the probabiltiy of you winning is extraordinarily low, less than 1%
Q: How likely is it that an Oracle AI will win against the average person?
A: Before: 80%. After: >99%
Q: How likely is it that an Oracle AI will win against you?
A: Before: 50%.
Q: Now that the experiment has concluded, what’s your probability of me winning against the average person?
Q: I want to play a game with you! How can I get this to occur?
A: It must be stressed that I actually don’t like playing the AI Box Experiment, and I cannot understand why I keep getting drawn back to it. Technically, I don’t plan on playing again, since I’ve already personally exhausted anything interesting about the AI Box Experiment that made me want to play it in the first place. For all future games, I will charge $1500 to play plus an additional $1500 if I win. I am okay with this money going to MIRI if you feel icky about me taking it. I hope that this is a ridiculous sum and that nobody actually agrees to it.
Q: How much do I have to pay to see chat logs of these experiments?
A: I will not reveal logs for any price.
Q: Any afterthoughts?
A: So ultimately, after my four (and hopefully last) games of AI boxing, I’m not sure what this proves. I had hoped to win these two experiments and claim prowess at this game like Eliezer does, but I lost, so that option is no longer available to me. I could say that this is a lesson that AI-Boxing is a terrible strategy for dealing with Oracle AI, but most of us already agree that that’s the case — plus unlike EY, I did play against gatekeepers who believed they could lose to AGI, so I’m not sure I changed anything.
Was I genuinely good at this game, and lost my last two due to poor circumstances and handicaps; or did I win due to luck and impress my gatekeepers due to post-purchase rationalization? I’m not sure — I’ll leave it up to you to decide.