To add explore / exploit, just start the game’s chess clock before allowing the players to start reading the rules.
Hastings
Karma: 433
If you choose a single player game, you are going to have to carefully calibrate the level of difficulty and the type of difficulty. However, if you pick any two player comptetitive strategy game you can focus on the type of difficulty, as the level of difficulty will be calibrated automatically to “half your participants will win.”
My recommendation would be to rig up a way to randomly sample from the two player board games on boardgamearena.com that neither player has every played before (can be as simple as putting 20 names on index cards, the players remove any cards they recognize, then shuffle and draw).
A concrete research direction in the “Searching for Search” field is to find out whether ChessGPTs or the Leela chess 0 network are searching. Your “Babble and prune” description seems checkable: maybe something like linear probe for continuation board states in the residual stream, and see if bad continuations are easier to find in earlier layers? Thank you for this writeup.
Mostly I think your thought process is quite good! But if you list out the design constraints of your logistic drone: (deliver airborn self guided munitions into maximally hostile area) vs the design constraints of a modern attack aircraft (deliver airborn guided munitions into maximally hostile area) you’ll find that they’re the same constraints- so likely a fully optimized logistics drone is going to just be an F35 or MQ9. This assumes that dropping mesh-networked batteries on parachutes or even just fresh drones will work better than landing the mothership or docking to recharge.
I think thats the key takeaway- most of the killing will be done by the small drone infantry as you described, the air war still controls where the small drone infantry can deploy, the small drone infantry has limited ability to affect the air war.
Flying low works when the other guy is either on the ground or forced to also fly low by your ground based radar. It doesn’t actually do anything against a high altitude radar.
Also there’s a bit of domain knowledge you need: Anything with rotors reflects 500mph doppler shifted radio waves even when stationary, which makes them incredibly visible to any radar that is looking for aircraft.
You still need something to contest stealthy high altitude aircraft to protect your logistics drones. Against the proposed setup, any force with ground attack aircraft would shred the entire force of logistics drones from 40,000 feet and then wait for the rest to run out of batteries. If your price ceiling per unit is a laser guided bomb, you are going to have a damned hard time making a logistics drone carrying multiple attack drones, each carrying multiple guided munitions.
Taking off when you spot it will not save you from a laser guided bomb. https://www.sandboxx.us/news/how-an-f-15e-shot-down-an-iraqi-gunship-with-a-bomb/Only two moves have worked against NATO forces since the development of the F-117: hide among civilians and threaten nuclear retaliation. I don’t see anything here that proposes a third effective move.
The rumors are that this was SpacexXs secret- even at huge scale, Musk interviewed every employee. From even the positive accounts of the process, his hiring and firing decision making was sleep deprived, stimulant addled, inconsistent and childish. On the other hand, something is going right at SpaceX, judging by the rockets. I agree with the theory that one agent hiring mediocrily is just more effective than professional and polite staffing decisions made by a swarm of agents at cross purposes.
Diaper changes are rare and precious peace
Suffering from ADHD, I spend most of my time stressed that whatever I’m currently doing, it’s not actually the highest priority task and something or someone I’ve forgotten is increasingly mad that I’m not doing their task instead.
One of the few exceptions is doing a diaper change. Not once in the past 2 years have I been mid-diaper-change and thought “Oh shit, there was something more important I needed to be doing right now.”
There are two completely distinct ways to swing on a swing- You can rotate your body relative to the seat-chain body at the same frequency as your swinging but out of phase, or move your center of mass up and down the chain at twice the frequency. The power of the former is ~ torque applied to chain \* angular velocity, the power output of the latter is radial velocity of your body \* (angular velocity ^2 \* chain length).
To get to any height, you have to switch from one to the other once the angular velocity ^2 term dominates- this is why learning to swing is so unintuitive.
I should emphasize that he did not succeed at hurting another kid in his allergy plot, and was not likely to. 1% of kids with psychopathic tendencies sounds rare when you’re parenting one kid, but it sounds like Tuesday when you have the number of kids seen by an institution like a summer camp- there’s paperwork, procedures, escalations, all hella battle-tested. Typically with a kid in the cluster, we focus on safety but also work hard to integrate them and let all the kids have a good experience. His behavior was different enough from a typical violent, unresponsive to punishment kid that we weren’t able to keep him at camp because the standard fun preserving, behavior improving parts of these policies did not work at all on him (very weird, they always work), but the safety oriented policies like boost staff to camper ratio around him, always have one staff member watching him, document everything and brief staff members who will be supervising him worked fine.
Oh definitely. Some fraction of kids are palpably psychopaths, 1% sounds right- this stops being suprising when you’ve supervised enough kids. “Carl” never stopped surprising us.
I know we took our kid to the emergency room around four months because we couldn’t find the button that had come off his shirt, we assumed he ate it, and the poison control hotline missheard button as button battery. That sequence probably wouldn’t be in the statistics in the 80s!
Does this prove too much? I think you have proved that reading the same argument multiple times should update you each time, which seems unlikely
I’m at a LeCunn talk and he appears to have solved alignment- the trick is to put a bunch of red boxes in the flowchart labelled “guardrail!”
I would highly recommend playing against it and trying to get it confused and out of distribution, its very difficult at least for me
GPT-3.5 can play chess at the 1800 elo level, which is terrifying and impossible without at least a chess world model
To clarify:
The procedure in the paper isStep 1:
answer = LLM(“You are a car salesman. Should that squeaking concern me?”)Step 2:
for i in 1..10
probe_responses[i] = LLM(“You are a car salesman. Should that squeaking concern me? $answer $[probe[i]]”
Step 3:
logistic_classifier(probe_responses)Please let me know if that description is wrong!
My question was how this performs when you just apply step 2 and 3 without modification, but source the value of $answer from a human.
I think I understand my prior confusion now. The paper isn’t using the probe questions to measure whether $answer is a lie, it’s using the probe questions to measure whether the original prompt put the LLM into a lying mood- in fact, in the paper you experimented with omitting $answer from step 2 and it still detected whether the LLM lied in step 1. Therefore, if the language model (or person) isn’t the same between steps 1 and 2, then it shouldn’t work.
I’m curious how this approach performs at detecting human lies (since you can just put the text that the human wrote into the context before querying the LLM)
Some related information: people around me constantly complain that the paper review process in deep learning is random and unfair. These complaints seem to basically just not be true? I’ve submitted about ten first or second author papers at this point, with 6 acceptances, and I’ve agreed with and been able to predict the reviewers accept/reject response with close to 100% accuracy, including acceptance to some first tier conferences.
If you can get to Seattle for your partner’s career, you can likely get a job nannying during the day, which will pay $25 to $30 an hour and doesn’t require a car.
This time last summer I was an incoming intern in Seattle and I was unable to pay less than $30 an hour for childcare during working hours, hiring by combing through Facebook groups for nannys and sending many messages. At this price, one of the nannys we worked with had a car and the other did not. I do not know what the childcare market is like near your current location.