🧵 View Thread
🧵 Thread (27 tweets)
A useful exploration on the topic of Goodhart's law"The dynamics highlighted are hopefully useful to explain many situations of interest in policy design, in machine learning, and in specific questions about AI alignment."https://t.co/Ka60eX7ra6 https://t.co/cymqBtsDxm
Goodhart's law in a nutshell."Asimov's three laws"https://t.co/7AxWjbmMIL
Good overview of the space of ideas here: https://t.co/Cxs6H9wfKshttps://t.co/DnabOdZCur
The not-so-obvious extension from economic theory to Schelling points, extended to biology. <3"It can be easier and more reliable to just learn and copy the proxy than to evolve the hidden quality and hope other people catch on" https://t.co/TQ1BSkHG9xhttps://t.co/l6hudT4wiE
"multi-agent failure modes are more complex, more problematic, and less well understood than the single-agent case, and are also already occurring, largely unnoticed. [...] why these failure modes are in some senses unavoidable." https://t.co/PsOXtSDW21 https://t.co/N3tyQqcrDc
Why I think it's funny to just not read books.https://t.co/VR8bJ4316u
Why should I read books when your library's book reading metric system spy's on what you are reading and falls for Goodhart's law in new and interesting ways!https://t.co/qgLqLyMqci
Your research repository platforms and library book checkout tools are watching what you read.It's even encouraged as best practice https://t.co/QdEueduXuHPay no attention to massive data tracking conglomerate buying up services behind the curtain.https://t.co/oq7DjAukmA
"citizens dutifully complied and began breeding venomous snakes to kill and bring to the British. By the time the experiment was over, the snake problem was worse than when it began." https://t.co/iXb10YXD9Uhttps://t.co/eU6jUS0fSf
"it’s better to take a dead snake as an indicator of what point you are at in the cycle—i.e. you’re at the end of something in your life." https://t.co/GV9YmW8x9Z
"This is the superset of Goodhart-Campbell. The solution is the superset of costly signaling. The greater the divergence in the incentive structure between proxy and proxied, the greater the incentives to optimize for appearance." https://t.co/tsFuOEdVe8 https://t.co/ZsXeZnJSkQ
Sure is Noisy in here.https://t.co/uNQ2RL97Rq
I hear carts clashing and cars honking. I feel the sun gazing down at me in the dead of winter as I lose the ability to feel my hand.https://t.co/a103p3xqwN
"Don't act so innocent, this was no accidentYou planned this in the end and now it's overSay what you wanna say, you lied and I got playedYou threw it all away and now it's over" 🎶https://t.co/XeYh7vMGTEhttps://t.co/ITmUnvhTXE
Thinking a lot about the principle-agent problem in the context of actor-network theory right now.How does one gain more agency over their own life?https://t.co/gZ5OTovUdc
I did not.https://t.co/TJ64BFY2U6
Hey @lukestein please have your bot not harass my tweets. kthnx. https://t.co/2SfAExNFUw
@lukestein Your bot haunting my misspelling literally hits a part of me being suicidal. So having random tweets show up hidden is kinda shitty. have you considered the principle agent problem of writing a spelling correction bot and it's impact on behavior? https://t.co/TYb0UDXE9P
@lukestein This is disgustinghttps://t.co/5eqCTBAATS
Neat."We suggest that the analysis of incomplete contracting developed by law and economics researchers can provide a useful framework for understanding the AI alignment problem and help to generate a systematic approach to finding solutions"https://t.co/jHzuJESoCx
Oooh, they got some good publications https://t.co/RqETNzqF4X
"from social sciences, e.g., Actor-Network Theory (ANT) can support the development of tracing networks of association amongst many actors, which can help understand how, [...] humans may offload value-laden behavior onto the technology around us." https://t.co/g0RFxbuvMn
Tired: AI alignment problemWired: AI going mad.https://t.co/EBh9efgz6x
I've got a lovely bunch of paperclips...https://t.co/GWKoTbA4qj
huh."They’re converging on the idea that it’s time to stop using gross domestic product (GDP) as the world’s main measure of prosperity, and to complement it with a dashboard of indicators, possibly ones linked to the [Sustainable Development Goals]."https://t.co/Jfg24xcXBx
Runaway utility maximization, in my economic proxy for maximizing paper clip extraction services? https://t.co/JG8fGLmPq4
While it's true that the cliche of "the cobra effect" is anecdotal. The wikipedia has examples of this effect happening that are documented. https://t.co/Vd0FU1N1R9Good to be nuanced here.https://t.co/GpZlq4kVhM https://t.co/OuXW8lRqkc
