Alex Turner, Oregon State University PhD student working on AI alignment.
TurnTrout
Karma: 5,620
I did this with my family! We had a lot of fun.
(Misread prompt as “you have 10 years of food & water”)1. Kick down the door
2. Unlock the door
3. Use high-pressure water source to explode the room4. Contact someone and get them to open the door
5. Learn how to pick locks and pick the locks
6. Eat through the door
7. Bust through the thin drywall
8. End up in a low-probability universe where you teleport outside the room
9. Wait for someone to come open the door
10. Upload yourself and have yourself reconstructed on the outside
11. Build an AGI to solve the problem
12. Hire someone to come blow up the door
13. Go on an extreme diet and slide under the door
14. Train hard at karate so you can break off the hinges with one blow
15. Use a rocket launcher
16. Climb out through ventilation shaft
17. Synthesize corrosive compound that eats through the wall
18. Tunnel underground with your hands
19. Take screw from glasses, burrow out hole in door so you can reach hand through and unlock it
20. Take your shoe and hit the bolts out
21. Use hemp in food to grow a rope and then slip under the door and wait for someone to trip and then let them know you’re there
22. Scream at resonant frequency of the wall so it explodes
23. Trigger sprinkler system to soften drywall
24. Burn down whole room and hope your remains are enterred outside of the room
25. Wear your way through the wall
26. Eat so much food that you get bigger than the room and then it explodes
27. Download blueprints for the room and then exploit weak points
28. Use avocado pits to cut out cinderblock
29. Find out how to make explosives from food (methane gas?)
30. Use acidity of something to wear away metal in room
31. Train super hard to run fast enough opposite to Earth’s spin to stop its spin and blow the room away via inertia (this wouldn’t work)
32. Build a 3D printer that takes food as input and produces keys to the door
33. Jump hard enough to bust head through ceiling
34. Use light + unlimited energy to wood burn out of the room
35. Wave your hands so fast the air turns to plasma
36. Call the police and tell them you have weed in the room
37. Let food rot; bug eggs hatch; bugs chew through the door
38. Hire a hitman to kidnap you
39. Climb out through the window
40. Rip off two of your toes and rub them together to make a fire
41. Wait for tornado
42. Wait for earthquake
43. Wait until sun expands and then “your” molecules likely end up outside of the former room’s molecules
44. Spin arms at right angle to get lift to turn yourself into a human helicopter and fly out of the skylight
45. Use tine on belt to pick lock
46. Get the building condemned (via your phone) and torn down
47. Sand away log walls with sandblock
48. Drill through the wall by spinning your hand quickly
49. Scratch out with really long fingernails
50. Yell for help
If you’re tempted to write “clearly” in a mathematical proof, the word quite likely glosses over a key detail you’re confused about. Use that temptation as a clue for where to dig in deeper.
At least, that’s how it is for me.
From unpublished work. The answer to this seems obvious in isolation: shaping helps with credit assignment, rescaling doesn’t (and might complicate certain methods in the advantage vs Q-value way). But I feel like maybe there’s an important interaction here that could inform a mathematical theory of how a reward signal guides learners through model space?
I really like this post. Thanks for writing it!
(Why didn’t you mention the Slaughterbot Slip-up of ’24, though?)
Ideally within the next month!
(I have a big google doc analyzing corrigibility & manipulation from the attainable utility landscape frame; I’ll link it here when the post goes up on LW)
Epistemic status: not an expert
Understanding Newton’s third law, .
Consider the vector-valued velocity as a function of time, . Scale this by the object’s mass and you get the momentum function over time. Imagine this momentum function wiggling around over time, the vector from the origin rotating and growing and shrinking.
The third law says that force is the derivative of this rescaled vector function—if an object is more massive, then the same displacement of this rescaled arrow is a proportionally smaller velocity modification, because of the rescaling!
And also, forces have opposite reactions (by conservation of momentum) and equal reactions (by conservation of energy).
I’m going to assume the evil forces knew I had it, but can’t read my mind or know everything I did to hide it. They’ll spend a reasonable amount of resources looking, but not an infinite amount, and they won’t torture me to find out where it is (they know how unreliable torture is as information gain!).
Put the pen in a small steel box, go on a multi-day hike, and randomly bury the box somewhere along the hike. Remember where you put it.
Travel to a new country and keep it in a bank deposit box under a fake name. (Did they have those in 1855?)
Acquire a bunch of pens that look exactly the same, put a different color marking on the interior of each, and give each to a friend. (Figure out some way to make sure the Evil People don’t just steal all of the pens?)
Put the pen inside the wall of a friend’s house.
Destroy the pen and then remake it in 50 years.
Embed it in a tree core. (Would the tree grow to crush it?)
Commission a copy of the pen and give it to them, claiming it’s the real thing.
Have 10 friends over and give them each a pen, only one of which is the Destined Pen. Then, each of them meets 10 people and gives one of them the pen, and so on once again. The final person who receives the actual pen is given the appropriate instructions and payment plan for delivering the pen to Einstein in 50 years (DDOS their resources—they probably can’t investigate 1000 people).
Find a box and put the pen in it.
Disassemble the pen and entrust the pieces to a pensmith (is that a thing?)
Preserve the pen in some kind of resin
Form a concrete block around the pen, or otherwise embed it in some kind of wall, and then take it out when needed
Go spelunking and leave the pen wedged in one of the caves in a memorable but rarely frequented location (I notice I’m focusing on the “hide from dedicated group of people” constraint, which is slowing me down. I’m going to focus on just “places you could put a pen” without worrying at all about hiding for now.)
On the ground.
In the attic.
Bird’s nest
Buried in dirt
Library
Carriage
Warehouse (Ok, back to solutions that have at least a small chance of working)
Pay someone to hide the pen, split up the location info into a cryptographically secure 3-out-of-5 secret-sharing scheme, and then dole out the info to me and four close friends who live far away and move frequently. In 1855, it would be logistically hard to coordinate to track 3 of us down at the same time. They pass on the secret to their children and spouses, in case any of us die before 50 years pass.
Unethical: kill the evil forces?
Unethical: take over the local government and have them protect the pen.
Talk the bad guys out of wanting the pen.
Set up an elaborate dungeon below my house (Legend of Zelda-style), at the end of which is a fake copy of the pen. Keep the real pen in my desk drawer.
Similarly, go to a lot of trouble to look like I’m building up a vault for the pen, all the while keeping “the rest” of my pens on a pedestal in the middle of my house. Of course, the real pen is there.
Persuade the bad guys that pens are really annoying and you should use a mechanical pencil (that totally existed in 1855, right?), if you aren’t going to use a digital writing device (which they aren’t going to use).
Give a fake pen to a major world power, and then tell the bad guys it’s already out of my hands, and eat popcorn while the two groups fight.
Find Einstein’s ancestors and convince them to pass down the pen.
Using extreme computational resources and technological prowess, build a capsule that’s shot into LEO which is timed to reenter Earth’s atmosphere in front of Einstein right when he’s looking for a new pen.
Deduce exactly what kind of simulation we’re on, then execute a bizarre policy that mind-hacks whoever’s observing us into spawning 1x Pen of Destiny for Einstein at exactly the right time.
Compute a Butterfly Effect policy for a molecularly identical pen spontaneously assembling itself in front of Einstein at the appropriate time (this is totally a thing, right?)
Wedge the sanitized pen between my radius and ulna. That totally wouldn’t go wrong, and it’d be too gross for them to want to retrieve.
Buy a farm and hide it in a hay bale.
Go to a stream and instruct a friend to wait some unknown # of miles downstream. Put the pen in a bottle, float it downstream, and the friend retrieves it and waits downstream. Maybe a festival could be going on near the stream at the same time. In any case, even if I’m being trailed at that moment, they wouldn’t know how far downstream it went, which would give my friend an important head start for hiding it in a different town.
Pay someone to take it to a tribe which will be convinced that it should be worshipped as a sacred object, and then steal it back in 50 years.
Memorize a binary sequence using a memory palace, which I use as an XOR cipher on a series of coin flips which indicate: “heads: go north 100 feet; tails: go east 100 feet”. Flip 100 coins and write down the result, and then bury the coin in the place indicated by the flips XOR the sequence. (This is basically a one-time pad for north-eastern lattice paths)
Do the above, but just through the pseudo-randomly generated memorized sequence. Also, have a habit of taking these north-eastern lattice path walks randomly for a few days before and after actually burying the pen, so they don’t know the walk on which I buried the pen.
Enter into a 2-out-of-2 secret sharing scheme with someone difficult to intimidate, like a major world leader.
Do 38, but using a fall tree branch as my (not technically) pseudorandom source: given a branch, choose the one that has the reddest leaf on it, using an appropriate embedding from branch choices into directional bearings for the next step.
Put the pen inside a football that won’t be used. Who would do that?
Preserve the pen in permafrost somewhere stable (is that a thing? would it be crushed by changing pressures?)
Encase the pen in diamond or some even harder material, which Einstein will need to invent molecular nanotech to undo. He wants his miracle papers, let him do the impossible!
If I’m making myself as capable as in #43, why not just build an AGI to ensure Einstein gets it? Might as well.
Toss the pen into a furnace in front of the bad guys; unbeknownst to them, the furnace has a secret compartment in which the pen will be safe from the heat.
Keep the pen inside a wooden compartment, which I nail to the underside of the nearest bridge (so that if the pen falls, it falls over land).
Fake my own death and inter the pen with the fake body.
47, but hide the pen in the gravestone.
48, but a friend surreptitiously removes the pen from the gravestone compartment during the funeral.
Just give the bad guys the pen and take it back before Einstein needs it.
OK, but there’s a difference between “here’s a definition of manipulation that’s so waterproof you couldn’t break it if you optimized against it with arbitrarily large optimization power” and “here’s my current best way of thinking about manipulation.” I was presenting the latter, because it helps me be less confused than if I just stuck to my previous gut-level, intuitive understanding of manipulation.
Edit: Put otherwise, I was replying more to your point (1) than your point (2) in the original comment. Sorry for the ambiguity!
In your original comment, you wrote:
one possible implication being he will also not accept the results
The first article in that search says:
Before leaving the state, Biden told reporters his comments were ‘taken a little out of context’ and added that ‘I’m going to accept the outcome of this election, period.’
Not Stuart, but I agree there’s overlap here. Personally, I think about manipulation as when an agent’s policy robustly steers the human into taking a certain kind of action, in a way that’s robust to the human’s counterfactual preferences. Like if I’m choosing which pair of shoes to buy, and I ask the AI for help, and no matter what preferences I had for shoes to begin with, I end up buying blue shoes, then I’m probably being manipulated. A non-manipulative AI would act in a way that increases my knowledge and lets me condition my actions on my preferences.
Sources?
I’m not very satisfied by answers like “X won’t support him, because that’s illegal” and “it’s unconstitutional for the federal government to do X, so they won’t.” I think these usually are correct, but over the last four years we have seen rapid deterioration of our ability to: agree on an objective reality, have an executive branch which abides by the law absent immediate and tangible enforcement mechanisms (remember when congressional subpoenas were at least often answered? now they’re ~always ignored AFAICT), have common knowledge that the law is the Law and if you break it you will be punished (obviously, rich&powerful would get more leeway in this calculation), etc.
I think many of these things have degraded and am no longer sure that anything would really stop red states from ignoring the popular results and sending their own set of electors. It’s already being discussed, and red officials have admitted they are discussing it without immediately walking it back / distancing themselves from the prospect. Maybe the Supreme Court would be enough to stop that, if they so chose. (Would they? Aren’t states technically allowed to choose electors however they please?)
I went to the doctor’s yesterday. This was embarrassing for them on several fronts.
First, I had to come in to do an appointment which could be done over telemedicine, but apparently there are regulations against this.
Second, while they did temp checks and required masks (yay!), none of the nurses or doctors actually wore anything stronger than a surgical mask. I’m coming in here with a KN95 + goggles + face shield because why not take cheap precautions to reduce the risk, and my own doctor is just wearing a surgical? I bought 20 KN95s for, like, 15 bucks on Amazon.
Third, and worst of all, my own doctor spouted absolute nonsense. The mildest insinuation was that surgical facemasks only prevent transmission, but I seem to recall that many kinds of surgical masks halve your chances of infection as well.
Then, as I understood it, he first claimed that coronavirus and the flu have comparable case fatality rates. I wasn’t sure if I’d heard him correctly—this was an expert talking about his area of expertise, so I felt like I had surely misunderstood him. I was taken aback. But, looking back, that’s what he meant.
He went on to suggest that we can’t expect COVID immunity to last (wrong) but also that we need to hit 70% herd immunity (wrong). How could you even believe both of these things at the same time? Under those beliefs, are we all just going to get sick forever? Maybe he didn’t notice the contradiction because he made the claims a few minutes apart.
Next, he implied that it’s not a huge deal that people have died because a lot of them had comorbidities. Except that’s not how comorbidities and counterfactual impact works. “No one’s making it out of here alive”, he says. An amusing rationalization.
He also claimed that nursing homes have an average stay length of 5 months. Wrong. AARP says it’s 1.5 years for men, 2.5 years for women, but I’ve seen other estimate elsewhere, all much higher than 5 months. Not sure what the point of this was—old people are 10 minutes from dying anyways? What?
Now, perhaps I misunderstood or misheard one or two points. But I’m pretty sure I didn’t mishear all of them. Isn’t it great that I can correct my doctor’s epidemiological claims after reading Zvi’s posts and half of an epidemiology textbook? I’m glad I can trust my doctor and his epistemology.
I read “feasible” as something like “rationalizable.” I think it would have been much clearer if you had said “if no strategy profiles are Pareto over any others.”
Wikipedia says if , it’s a “non-expansive map”. But yes, contraction maps have some Lipschitz constant that enforces the behavior you describe. However, notice we still have the math “intuitive contraction” here, so it has the reverse-direction correspondence. Intriguingly, we’re missing part of “intuitive contraction = bring things closer together” for the case, as you point out, so we don’t have the forward direction fulfilled.
Math That Clicks: Look for Two-Way Correspondences
You’re thinking of a Kaldor-Hicks optimality frontier for {outcomes with maximal total payoff}, while the Pareto frontier is {maximal elements in the unanimous-agreement preference ordering over outcomes}.
Right, I understand how this correctly labels certain cases, but that doesn’t seem to address my question?
Forgive me if I missed it in the post, but might it be cheaper to use some kind of visor if you’re only illuminating for personal benefit? Much easier and cheaper to fill your FOV with light that way.