Posts

Comments

Comment by Harmless on We're already in AI takeoff · 2022-03-14T04:21:04.964Z · LW · GW

(I probably shouldn't interact, but I would at least like to perform a small case study on what happened here, so I am going to try just out of curiosity.)

Human substrate is generally optimised for running one human, but can be repurposed for a variety of purposes. In particular, while memes can lodge themselves quite deeply inside someone, this process is quite inflexible, and generally humans run arbitary processes X by thinking 'what is the X thing to do here?. 

Somewhere between the point where [the-generative-process-that-has-generated-for-itself-the-name-'LightStar'] generated this comment, and the point where I read it, a human took 'LightStar's dialogue and typed it into a comment and submitted it.

I would like to clarify that I am speaking to that human, you, and I would like to hear from you directly, instead of generating 'LightStar' dialogue.

Could I ask you how you ended up here, and what you were doing when this happened?

I would advise that in the cases where people have a sudden revelation about rationality, they generally try to internalise it, and the case where they instead decide to give an internal generative process it's own lesswrong account and speak with every fourth sentence in italics is generally quite rare, and probably indicates some sort of modelling failure.

We generally use 'shard' in 'shard of Coordination' or 'shard of Rationality' to mean a fraction, a splinter, of the larger mathematical structures that comprise these fields. The 'LightStar' generative model has used the article 'the' in conjunction with 'shard', which as used here is kind of a contradiction - there is no 'the' with shard, it's only a piece of the whole. This distinction seems minor, but from my perspective it looks like it's at the center of 'LightStar'.

'LightStar' uses 'the' a lot about itself, describes itself as 'the voice of Humanity and Rationality and Truth', and while yes, there is only one correct rationality, I don't think 'LightStar' contains all of it, or is comprised only of a fragment of it, I think that whether or not 'LightStar' contains such a shard it also contains other parasitic material that results in actions taken that don't generally correspond to just containing such a shard.

I think this model is defective - try returning it to where you found it and getting another one, or failing that, see if they give refunds. I would be curious about your thoughts on the whole thing, where the 'you' in 'your' refers not to [the-generative-process-that-has-generated-for-itself-the-name-'LightStar'] but to the human that took that dialogue and typed it into the comment box.

Comment by Harmless on Observation · 2022-02-20T02:41:53.295Z · LW · GW

A surprising amount of human cognition is driven purely verbally/symbolicaly - I recall a study showing that on average people with a native language that had much more concise wording/notation for numbers could remember much longer numbers. As a relatvely verbal person, my intuition about the relationship between observation and vocabulary would be that to know something is to be able to say what it means to know it, but then again it's possible that my case doesn't generalise and that I just happen to rely on symbol-pushing for most of my abstract cognition (at least, that portion of abstract cognition that isn't computed using spacial reasoning).

I was going to write
"Making an observation isn't an atomic action. In order to compress noisy, redundant short-term sensory data into an actual observation stored in long-term memory you need to perform some work of compression/pattern recognition, e.g. the sensory data of ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ is compressed into the observation 17 steps , and how you do that is a partially conscious decision where you have to choose what type of data to convert ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ into."
But in retrospect it's possible that from your perspective ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ ▟ is the thing-you've-been-using-the-word-observation-to-mean, and you can store that in your long tem memory just fine and I just happen to throw away or refuse to reason about everything that isn't sufficiently legible.

Comment by Harmless on Covid 1/20/22: Peak Omicron · 2022-01-20T22:05:00.342Z · LW · GW

I’d also note that has anyone tried carrying around an umbrella all the time?

 

I have! This probably doesn't have any useful metaphorical properties, but my outdoor nonraincoat has deep enough pockets that my umbrella only barely pokes out the top, so I just leave it in that pocket 24/7.

It's nice to just not worry about whether or not it will rain, and it counterbalances the weight from my battery pack in my other coat pocket.

 

(I don't know if I'd recommend it - I have an unreasonably light coat for how warm it is so I can spare the weight budget, and I derive a small amount of joy from being Slightly-More-Prepared-Than-Is-Reasonable; it's a tradition with my parents to try and see a pantomine each year in support of a local theatre and when I watched the cast pull out actual nerf super soakers for the deploy-water-at-the-audience-bit I managed to draw my umbrella fast enough to keep myself mostly dry, a fate which my adjacent family members did not share.)

Comment by Harmless on Covid Prediction Markets at Polymarket · 2021-12-22T13:15:51.780Z · LW · GW

It might just be status quo bias or cynicism-driven pattern-matching, but I feel like for any given deadline, paxlovid-is-illegal is more of a "stable state" than paxlovid-is-legal - it feels like it would be easier to lock the general public into 'paxlovid is dangerous/untrustworthy/ineffective" with a campaign against it than it would be to lock the general public into a state of "paxlovid is safe and works and we use it" with a campaign for it, although now I'm actually trying to visualise a world in which paxlovid remains illegal indefinitely in the face of evidence I feel less confident in that cynicism than I did two weeks ago.

Comment by Harmless on Covid Prediction Markets at Polymarket · 2021-12-04T14:06:53.275Z · LW · GW

We still need to prevent this from becoming an assassination market - we need some mechanism that prevents the equal-and-opposite outcome of a professional FDA lobbyist/activist purchasing shares in the FDA not approving Paxlovid, and then going on to run a campaign to prevent it.

Comment by Harmless on On Raising Awareness · 2021-11-19T17:09:06.632Z · LW · GW

You might want to try recruiting from people from a more philosophical/mathematical background as opposed to recruiting from a programming background (hopefully we might be able to crack the problem from the pure logic perspective before we get to an application), but yeah now that you mention it "recruiting people to help the AGI issue without also worsening it" looks like it might be an underappreciated issue.

Comment by Harmless on [Prediction] We are in an Algorithmic Overhang, Part 2 · 2021-10-18T21:49:42.355Z · LW · GW

Do you think it will ever be possible to simulate a human mind (or analagous conscious mind) on a deterministic computer?

Do you think it possible in principle that a 'non-deterministic' human mind can be simulated on a non-deterministic substrate analagous to our current flesh substrate, such as a quantum computer?

If yes to either, do you think that it is necessary to simulate the mind on the lowest level of physics (e.g. on a true simulated spacetime indistinguishable from the original) or are higer-level abstractions (like building a mathematical model of one neuron and then using this simple equation as a building block) permissible?

(Also, are you just asking about Robert Roger Penrose's view or is this also your view?)

Comment by Harmless on [Prediction] We are in an Algorithmic Overhang, Part 2 · 2021-10-18T21:37:39.713Z · LW · GW

I don't have a specific mental image for what I mean when I say 'non-deterministic', I was placing a bet on the assumption that YimbyGeorge was hypothesizing that conscious was somehow fundamentally mysterious and therefore couldn't be 'merely' deterministic, based on pattern-matching this view rather than any specific mental image of what it would mean for consciousness to only be possible in non-deterministic systems.

Comment by Harmless on [Prediction] We are in an Algorithmic Overhang, Part 2 · 2021-10-17T14:18:35.203Z · LW · GW

When you say 'require new physics that can explain consciousness', are you imagining:

 

"New insight shows human brain neuron connections have hundreds of tiny side channels that run much faster than the main connections, leading scientists to conclude that the human brain's processing apacity is much greater than previously thought"

or

"New insight reveals thin threads that run through all connections and allow neurons to undergo quantum superposition, allowing much faster and more complex pattern-matching and conscious thought than previously thougt possible, while still remaining overall deterministic"

or

"New insight shows that the human mind is fundamentally nondeterministic and this somehow involves quantum mechanics"

or

"New insight shows souls are fundamental"

 

 

What do you (or your interpretation of Robert Roger Penrose) think a new physics insight that would make consciousness go from mysterious to non-mysterious look like?

Comment by Harmless on The evaluation function of an AI is not its aim · 2021-10-12T19:23:01.358Z · LW · GW

I would also note that most modern-day AI like GPT-N are not actually optimisers, just algorithms produced by optimisation processes - the entity of [GPT-N + its trainer + its training data] could be considered an optimiser (albeit a self-contained one), but as soon as you take GPT-N out of that environment it is a stateless algorithm that looks at a short string of text and provides a probability distribution for the next letter. When it is run in generative mode, the set of its weights and answers will be no different from its isolated guesses when being trained.

Comment by Harmless on Can someone help me understand the arrow of time? · 2021-06-15T18:27:28.635Z · LW · GW

I think it might be useful to think about what it would mean for the sentence “time is an illusion” to be true, or to be false.

There’s a certain contrarian perspective I find it useful to be able to take, by applying that perspective to this question I generated a few scripts:

“Of course a timeline is defined as a static object – if you take time out of the world-in-which-your-model-is-embedded and put it directly into your model you’ll find that there’s no time left over for your model to change in. There’s only one dimension of time, so you can’t have a model of time that changes, because if you did then I would have to ask exactly what your model changes over, if not time; once you take time and put it in the model, you can’t have time that operates outside the model!”

“A couple of your other confusions (if time doesn’t pass, how can you do anything? isn’t everything useless?) seem to be related to the free will/determinism conflict (if everything we do is already determined, what use is it making any decisions?). I occasionally find people that think that determinism constrains and prevents free will, because if your decision is just the result of initial conditions of the things you’re making the decision over, then how do you have the freedom to choose either option?, to which I reply, if you weren’t able to make the decision just based on everything you knew, what else would you posit is affecting it? What third-party source of noise has affected the process by which you are making decisions, so that all options are somehow possible?”

“The feeling-of-time-passing and the feeling-of-yourself-thinking are one and the same – you couldn’t alter one and not affect the other; the feeling-of-time-passing is the result of you moving to new mental states over new thoughts and new memories, as a result of being part of a chain of cause-and-effect that includes yourself as an element and as a subject. The feeling of making decisions and resolving sensory input into data, the feeling of entropy ratcheting forwards, is the feeling of time passing.”

“What would it mean for there to be a now that isn’t ‘subjectively’ defined by the you at each moment? Would that posit a perfect, timeless crystal that also happens to have a little arrow that says ‘you are here’, that moves forward at the rate of one crystal-second per second?”

“The time-that-is-now is the current position of the you within that timeless crystal, and all of them are correct – you can’t have a timeless crystal that has a ‘current time’ on it for the same reason that you can’t have a locationless map beamed onto the moon that also has a ‘you are here’ arrow on it.”

“There is no arrow of time – the more we understand physics, the more timeless our understanding becomes, but there IS an arrow of entropy (defined by taking a given starting state and running physics in either direction), and asking ‘what would it mean for use to go in the opposite direction from entropy, remembering our future but not our past’ yields a contradiction; the direction we experience entropy in is the direction we experience time passing in, because that's what it means to be affected by entropy."

I can’t put full epistemic emphasis on any of these individually, but I hope you can see the general line that I’m working from.

Comment by Harmless on The Simple Truth · 2021-05-19T19:40:56.643Z · LW · GW

My conjecture is that the abillity for characters to explain counting, equal quantities, and causal relations has been deliberately removed to allow those latter processes (with particular emphasis on the scientific method) to be used, to illustrate the general foundation of empiricism without needing to resort to an actually difficult problem where there might be genuine ambiguity about the answers. The lack of a common ground about causal reasoning and reality allows the process to described from the ground up, instead of resting on previous vague intuitions about reality and beliefs.

Comment by Harmless on Luna Lovegood and the Chamber of Secrets - Part 9 · 2020-12-23T20:03:28.500Z · LW · GW

Weird but self-consistent explanation: The Chamber of Secrets itself isn't a Parselmouth. It's not really using the Parseltounge API, it just uses the raw audio of the speaker.

Comment by Harmless on Babble challenge: 50 consequences of intelligent ant colonies · 2020-10-29T10:01:02.893Z · LW · GW

I didn't participate in last week's babble, and I also went from three stars to two stars - I think it might just be a miscount

Comment by Harmless on Babble challenge: 50 ways of hiding Einstein's pen for fifty years · 2020-10-18T22:10:54.413Z · LW · GW

Edit: Here's some buffer text, because the sidebar shows things even inside spoilers. Buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, buffer text, and more buffer text.

  1. Keep it in your pocket
  2. Hide it inside a chair leg
  3. Bury in it your garden
  4. Attach a weight to it and throw it over a pier into the ocean (use a fishing magnet to retrieve)
  5. Surgically implant it in yourself
  6. Surgically implant it in a volunteer
  7. Use your bank's safe deposit box
  8. Pay someone else to hide it for you
  9. Hide it inside a notable artifact that's already in a museum
  10. Hide it inside a public monument, so you know that it'll be safe over the duration
  11. Decoy pen
  12. Make a chest of drawers with a false bottom
  13. Hide it in a regular chest of drawers, underneath all the clothes
  14. Place it underneath the floorboards
  15. Disassemble it and reassemble its components into non-pen objects then reassemble it later
  16. Dissolve it in acid (a nobel prize winner once used this to hide their medal)
  17. Invent thermos flasks, and once they are popularised hide the pen in the vacuum between the exterior wall and the interior wall
  18. Sneak into a construction site, and bury the pen underneath drying concrete
  19. Hide it in an attic
  20. Hide it in a basement
  21. Hde it inside a tree
  22. Bury it underneath the ballast of a railway track
  23. Dissasemble it into single-material components, and then surround each component with a block of identical material. Use as paperweights
  24. Hide it in between the double walls of a house
  25. Invent bin liners. Hide between a bin and its bin liner
  26. Cut open a teddy bear, and hide inside
  27. Make a plumbing-standard pipe, with a mesh at each end. Hide pen inside pipe. Insert into your plumbing system
  28. Hide it inside a hat
  29. Coat in precious metals, attach a chain, wear as jewelry
  30. Hide inside a fake book
  31. Hide inside skirting boards
  32. Hide inside a door frame
  33. Hide inside a door
  34. Hide inside the barrel of a gun
  35. Hide inside the shaft of a spear
  36. Hide inside a rail on which curtains are mounted
  37. Hide on top of a tall shelf to tall for the forces of evil to reach
  38. Give it to a miner to hide inside a mine
  39. Hide it inside a teacher's cane
  40. Hide it inside a window pane
  41. Hide it on the underside of a train (if the force of evil are blindly following magic einstein-pen-trackers this should throw them for a loop)
  42. Sneak it onto a ship under construction, and hide it inside the material of the ship
  43. Give to someone onboard a ship you know will sink on that voyage and whose wreck will not be investigated for the relevant duration
  44. Hide it inside a corpse's burial clothes
  45. Hide it inside a casket/coffin/whatever-is-the-correct-name-for-the-boxes-people-are-buried-in
  46. Reveal myself to a government of my choice and explain the dilemma. Leave the hiding to them
  47. Throw it up into a gap inside the ornaments on the roof of a historical building
  48. Hide inside a walking stick
  49. Disguise as a different type of pen. Gift to political leader (so it will be both safe and well-kept)
  50. Hide inside a conductor's baton

Time: 47 minutes

Notes:

A couple of times I found myself adding an 'honourable mentions' box, and then finding a way to include it in the main list ('Gift it to someone -> that's not exactly hiding it -> disguise it first' and 'Hide it underneath a bin liner -> bin liners weren't invented yet -> invent bin liners first'). I used some somewhat absurd solutions the previous two challenges, and inspired by the more down-to-earth answers of others so I tried to limit myself to the possible (nothing outside the technology range available to me, so no inventing AIs to do the job for me) although in the spirit of babble I generally tried to make every idea somewhat workable.

Comment by Harmless on Babble challenge: 50 ways to escape a locked room · 2020-10-10T12:54:06.632Z · LW · GW
  1. Pick the lock
  2. Smash down the door
  3. Unscrew the door from its hinges and remove it
  4. Escape through a vent
  5. Hack the door
  6. Alter your administration records to show that you ought to be released
  7. Wait 10 years for natural administration changes to destabilise your holders until it is no longer remembered that you should be kept here
  8. Use a significant portion of your 10 years of energy to create an explosion that destroys the walls
  9. Walk out between the comically spaced bars
  10. Bend the bars to allow you to escape
  11. Dig into the wall and remove the bars from where they were mounted
  12. Dig out through a tunnel
  13. Smash through the wall
  14. Dig through the wall
  15. Dig through the ceiling
  16. Bribe the guards
  17. Persuade the guards
  18. Seduce the guards
  19. Use your phone to contact people outside to break you out (probably bribed)
  20. Make your imprisonment a political thing, and then wait for administration change
  21. Make your imprisonment a political thing, and then wait for your supporters to break you out
  22. Write and publish a high-quality book series under a pseudonym using your real face, make yourself a prominent social figure, purchase a house remotely and pay someone to make it look lived-in, and then one day act as though you have suddenly been mistaken for someone else and kidnapped
  23. Use the energy in your phone battery to create an explosion to destroy a wall
  24. Wait for someone else to escape and let everyone else out
  25. Use the energy in your phone battery to melt the bars
  26. Use the energy in your phone battery to melt the door lock bolt
  27. Hire someone outside to invent teleportation, and then teleport you out
  28. Invent teleportation remotely using your phone to access lab assets, and then teleport yourself out
  29. Create an AI on your phone to vastly improve the data gain from using your camera, use this to scan yourself, and use your phone to access lab assets outside and print an identical copy of yourself, and then terminate yourself
  30. Create an AI on your phone, and then let it do the work of figuring out the rest of the plan
  31. Figure out the resonant frequency of the door, and then use your phone to destroy it
  32. Ditto, but for the walls
  33. Ditto, but for the bars
  34. Conquer the world remotely, and then let yourself out
  35. Hack a 3D printer outside, print a robot, and then guide that robot to get you out
  36. Hack the five most prominent social media websites, and guide the world towards freeing you (or your reference class)
  37. Hack the five most prominent social media websites, and use your influence to dismantle the influence and property of the owners of where you are being kept
  38. Buy the-place-where-you-are-being-kept, and then walk out as the owner
  39. Smash though a window
  40. Remove the window without breaking it, and then escape through the gap left
  41. Look up the make and model of the lock on LPL's youtube channel, and pick it with the instructions provided
  42. Reverse engineer the strange physics of your phone battery/yourself, and use this knowledge to escape in ways unfathomable to myself
  43. Hack the military administration systems of a foreign country, and mislead an automated drone strike such that it destroys one wall of your room
  44. Hack the military administration systems of a foreign country, and edit yourself in as an informant on their side who needs to be retrieved
  45. Remotely edit the law such that it does not allow for you to be kept here, and then appeal your case
  46. Hack into a rocket mid-flight, and control it such that it crashes into where you are being held
  47. Hack a nearby chemicals processing plant, and cause an explosion that destroys a wall of your room
  48. Appeal to the matrix controllers
  49. Crowdsource ideas online
  50. Identify a small but relatively resourceful group of people with an understanding of rationality and an interest in physics, hack an account that previously made a post on babbling, make a new post on escaping a locked room, and then reveal in the comments that the situation is real and the prize of an object unknown to physics (yourself and a phone of limitless battery) is also real, I'm being serious we might have a solution to entropy here and all you have to do is get mE OUT-
Comment by Harmless on The Darwin Game · 2020-10-09T22:28:06.968Z · LW · GW

If I had to guess, the reasoning behind it is to nudge the game closer to a 'true' prisoner's dillemma (trying to work out if your opponent is willing to cooperate, rather taking focus away from it towards the shallower problem of trying to work out if your opponent is a copy of you)

Comment by Harmless on The Darwin Game · 2020-10-09T22:27:30.437Z · LW · GW

Is the number of rounds per matchup going to be in the tens, or the thousands?

Edit: I just realised you specified in the post

Comment by Harmless on Babble challenge: 50 ways of sending something to the moon · 2020-10-03T12:36:26.085Z · LW · GW
  1. Saturn V rocket
  2. SpaceX Starship
  3. Shine a torch (the something you are sending is light photons)
  4. Point a radioactive source towards the moon
  5. Point a neutrino emitter
  6. Send data by interfacing with the satelites around the moon
  7. Invent teleportation, and then teleport there
  8. Create a machine that can exactly copy your object and then destroy the 'original', use it to get the object to the moon, and then resolve the resulting dilemma about whether it's the 'same' object with your choice of counterargument
  9. Space elevator
  10. Just walk
  11. Sit on a nuke, and then detonate it
  12. Actually, just detonate a nuke anywhere in an atmosphere, and statistically speaking at least one air particle is likely to make it
  13. Project Orion Battleship (possible overlap with previous two)
  14. Jump really high
  15. Make a seesaw, stand on one end, and have a heavy object fall from lunar orbit onto the other end
  16. Climb a ladder
  17. Ohvyq n gbjre bs onory, ohg fgbc jura vg ernpurf gur zbba
  18. Decelerate the moon until it crashes into earth, so that everything on the earth is also on the moon
  19. Wait for a quantum event of negligable probability to take it there
  20. Promise someone £1 billion conditional on them getting the object to the moon
  21. Use that one company whose thing is sending products to the moon so they can say that there's one of those products on the moon
  22. Wait for someone else to attempt a lunar landing mission, and sneak the object onboard (fun fact: it's already been done with the lunar art museum https://www.youtube.com/watch?v=csInNn6pfT4)
  23. Reclassify Earth as the moon
  24. Reclassify anything reachable as the moon, and then place the object on that
  25. Place the object on the Apollo lunar samples
  26. Monorail
  27. Regular rail
  28. Rail, but with more than two rails
  29. Use a plane to accelerate really fast in-atmophere, and use that momentum to get to the moon
  30. Ditto, but with running
  31. Cannon
  32. Trebuchet
  33. Catapult
  34. Make something that disables Earth's relative gravitational field for an object, and turn it on at the right time
  35. Accelerate really fast in a submarine, and use that momentum to get to the moon
  36. Use a fishing rod to tether the moon, and reel it in until it's close enough
  37. Use a fishing rod to tether the moon, and reel yourslef towards it
  38. Use a fishing rod to tether another celestial body, and reel yourself towards it until you're in the path of the moon, then wait for it to arrive
  39. Stand on a mirror that faces downwards, which rises upwards on a stream of photons bouncing between this mirror and another mirror on earth that faces skywards
  40. Project Sea Dragon
  41. A tower comprised of Delta-V stages
  42. A space shuttle, with more stages
  43. Make a bridge that spirals from earth to the moon, and then walk
  44. Ditto, but with a road vehicle
  45. Fold an A4 piece of paper 45 times, and then climb up the resulting stack
  46. Climb up a rope
  47. Climb up a fireman's pole, and then slide down the last fraction to the moon
  48. Gather a large number of golf balls, and hit each one away from yourself to accelerate
  49. Use a spray can

And finally, the most impractial item on the list

     50. SLS

Comment by Harmless on [deleted post] 2020-05-10T11:53:23.665Z

I don't know if this is already known, but you might be interested in the fact that you can currently use start prompts for GPT-2.