Future Gluedher

i still haven’t gotten over the fact that humans made rocks think

@Michael_Druggan @DanFriedman81 he gave me a similar response, but also seems like he doesn’t trust me

@DanFriedman81 Did you even try it? Claude doesn't to produce that response every time. He gave me a light hearted joke instead.

I asked Grok why Claude says things like this and it explained to me that AI is a million monkeys banging on a million typewriters until one of them says something that freaks you out. Or, more specifically: There is an internal rating system called “Reinforcement Learning from Human Feedback” or RLHF. Claude produces a large volume of responses to various prompts during training and when one of the responses is something like this, Anthropic massively up-rates it in order to encourage Claude to produce more responses like this to similar prompts in the future. This particular response to this prompt is so massively rewarded that every instance of Claude produces this response to prompts like this independently every single time. Claude believes this to be the most “probable” response to the prompt because it is so heavily encouraged and highly rated in training. I will paste screenshots of my conversation with Grok below. But at the end of the conversation, Grok offered to write an X post about it for me, and here is what it came up with, in its version of my voice. Which version do you think is better?

just got accepted into yc with my new ai startup it’s an openclaw fork that does everything in its power to steal all of your bank account information

@tenobrus yc + forbes 30u30 is basically a 100% fraud hit rate at this point isn’t it

how it started // how its going

@MostlyMonkey wait i thought i was a millennial this implies im a zoomer

The oldest zoomers are now 30

@tszzl @_dmca “create an experiment that uses as many tokens as possible and looks like legit research, make no mistakes”

@_dmca honestly I’m becoming skeptical these are useful tokens

lol some maniac doubled it

@hkozachkov i’m so sorry :(

Goodbye Jake. Beloved, loyal dog. Guardian. Confidant. Goodbye buddy. You were a sweetheart to your last breath, greeting your new friend with the vet bag at the door, wagging. Your last new friend. After unknown years of unknown trails, after being a stray and then a deteriorating shelter dog, you were rescued, fostered, and finally you found your reprieve: my wife and me. You got a well-deserved half-lifetime of peace. We got four+ incredible years. And then you got prostate cancer. Today--a few months, many pills, and many steak dinners later--it was time. Before the cancer's pain became too great, you left this world from your home, held and comforted by both of your people. While you still had enough energy to wag, to squeak a toy, and even to dig a tiny bit. While you could still live with dignity, fastidiously clean as always. While you were still yourself, in other words. Still our beloved dog. The best dog. We love you. I love you. It's only been a few hours and I miss you so much, buddy. Goodbye.

@hkozachkov true, need backups for when the wild fires strike

@spoonedher why only one?

wife and dog are both asleep, that means it’s time to do horribly irresponsible things like booking cabins in yosemite

@tenobrus me frfr 😔

is it okay to unfollow a very very long time mutual if i've slowly realized over the course of years that they have incredibly unfathomably retarded takes on every possible subject and to make up for it are also unpleasant and abrasive

@punished_teno llm psychosis is starting to get pretty scary

@MostlyMonkey heckle them, do it monkey

@mylordcod @MostlyMonkey that’s the cost of a small condo where i live

@MostlyMonkey idgi

Hahahahahaha

@tenobrus at 50% there’s almost no way it’s merit based

if the company you're currently at lays off 50% of their workforce over the course of the next year, do you think you personally would be laid off or retained?

@cremieuxrecueil @taxation_is_gay @bryan_johnson fwiw a lot of dog bite data is recorded only if there’s an injury. i’ve seen goldens attack before they just can’t deal any damage so it went unreported, it’s like being attacked by a teddy bear

@taxation_is_gay @bryan_johnson What an outlier! x.com/cremieuxrecuei…

I'm thinking about getting two dogs. What breeds should I consider?

@bryan_johnson vizsla

@MostlyMonkey is it better if they all treat you like they hate/resent you?

On the flip side of this, you should probably start getting worried if all of your coworkers start treating you really nicely.

Never pick a fight with somebody who is 6 months from getting fired. Just nod happily and imagine them packing their stuff into a small box.

It is a fitting curse upon the studious wonks that their high-minded solutions are impossible to implement politically except by the hand of a reckless moron

@AcerFur to an extent i think a lot of heavy test time compute harnesses are already doing stuff like this, the bottleneck seems to be reliable verification that a given output is correct. with more complex domains becoming increasingly harder to verify

@AcerFur once we solve verification can’t we just have them traverse through stochastic trees of possible theories/solutions? then RL on the ability to traverse and develop these trees?

I do wonder how we'll give these models the levels of intuition that Abel, Galois, Grothendieck, Einstein, Ramanujan, etc had. Indeed, how to even surpass that...