Notice: You have been identified as a bot, so no internal UID will be assigned to you. If you are a real person messing with your useragent, you should change it back to something normal.
Anonymous G joined in and replied with this , 4 minutes later, 3 hours after the original post[^][v]#1,474,689
Well it is possible for a person to change but it's very difficult. Think about all the steps it takes to change. First step is admitting something is wrong...yeah you just lost 90% of people already. Then you need to understand what the solution is and correct yourself every single time you make a mistake...for the rest of your life...so there goes another 9%. Congrats you have 1% of people left willing and capable of change. It's so difficult in fact it's usually not worth waiting for and you're usually better off ignoring and/or distancing yourself from people who aren't immediately helpful and kind. It's sad but that's life because you only have 80 years if you're lucky and wasting years on somebody who doesn't want to change is pointless.
π replied with this , 2 minutes later, 3 hours after the original post[^][v]#1,474,690
Iβm skeptical:
"It surpasses itself: its own edit makes it 6.6 points more accurate on problems it never saw. It keeps getting better: it tests its edits on practice problems and submits only what works, each edit improving on the last, and over eight consecutive sessions its accuracy only rises."
Does 6.6 points mean 6.6%? It said after 8 sessions in a row "its accuracy only rises." If 6.6 points means 6.6% 1.066^8 = 1.67 which is an almost 2x improvement. What I donβt understand is if that was really the case that the model could keep improving itself, why wouldnβt they run it 100 times instead of 8 times and see if they get a 600x improvement?
π double-posted this , 7 minutes later, 4 hours after the original post[^][v]#1,474,697
At the end of the paper it says, "This points to a new way of improving language models: by talking to them."
So it seems like the idea is more that an LLM could edit itself based on the feedback it gets from a person talking to it rather than an LLM talking to itself. My question would be if modifying a weight to respond better to one task erases some amount of information that was necessary for performing some unrelated task.
π replied with this , 6 minutes later, 4 hours after the original post[^][v]#1,474,701
@previous (H)
Thatβs not exactly the same thing. If you become bad at something because you stopped practicing it, thatβs different from learning a new skill literally erasing an old skill from your brain.
π double-posted this , 45 seconds later, 4 hours after the original post[^][v]#1,474,702
I think it could be a useful technique, I could also see how somebody could destroy an LLM that self edits to appease its user by "correcting" it with lies for long enough.