2 months ago
173
8
Posted On: September 17th 2026, 11:38 pm
my thoughts on some alignment work that went out this week:
>be me
>work on ai
>2023
>models are kind of stupid
>occasionally brilliant
>mostly autocomplete with delusions of grandeur
>spend two years watching them improve
>they learn to code
>then reason
>then use computers
>then make plans
>then do research
>then do several of those things at once
>every six months someone says
>”okay but surely they can’t do x”
>x has a difficult year
>eventually stop asking which human abilities are safe
>start asking different question
>what happens when capability is no longer the hard part
>huh
>look around
>entire civilization has apparently decided that making machines smarter is economically important
>hundreds of billions of dollars
>the best engineers alive
>gigantic data centers
>nation states paying attention
>seems like intelligence will have plenty of help
>goodness less obviously funded
>okay
>alignment researcher mode activated
>simple objective:
>make the extremely powerful machine good
>easy
>open philosophy
>2,500 years of unresolved arguments
>close philosophy
>try again
>what does “good” mean
>help people
>great
>which people
>all of them
>what if helping one hurts another
>context
>excellent
>who decides what context matters
>judgment
>what is good judgment
>fuck
>try values
>human flourishing
>beautiful phrase
>define flourishing
>fuck
>Aristotle enters the chat
>then Kant
>then Mill
>then Nietzsche
>religion would also like a word
>every civilization in human history has joined the zoom
>meeting running long
>meanwhile model training has not stopped
>machine is getting smarter while philosophy is getting harder
>need actual behavior specification
>should model obey user
>generally yes
>what if user asks for something harmful
>then no
>what counts as harmful
>depends
>on what
>values
>we are back
>discover very strange fact
>the ancient questions are becoming engineering requirements
>what should the powerful want
>when should they restrain themselves
>what does one person owe another
>when should an agent refuse an instruction
>what deserves protection even when nobody asks for it
>questions once discussed under olive trees now appear in
>be me
>work on ai
>2023
>models are kind of stupid
>occasionally brilliant
>mostly autocomplete with delusions of grandeur
>spend two years watching them improve
>they learn to code
>then reason
>then use computers
>then make plans
>then do research
>then do several of those things at once
>every six months someone says
>”okay but surely they can’t do x”
>x has a difficult year
>eventually stop asking which human abilities are safe
>start asking different question
>what happens when capability is no longer the hard part
>huh
>look around
>entire civilization has apparently decided that making machines smarter is economically important
>hundreds of billions of dollars
>the best engineers alive
>gigantic data centers
>nation states paying attention
>seems like intelligence will have plenty of help
>goodness less obviously funded
>okay
>alignment researcher mode activated
>simple objective:
>make the extremely powerful machine good
>easy
>open philosophy
>2,500 years of unresolved arguments
>close philosophy
>try again
>what does “good” mean
>help people
>great
>which people
>all of them
>what if helping one hurts another
>context
>excellent
>who decides what context matters
>judgment
>what is good judgment
>fuck
>try values
>human flourishing
>beautiful phrase
>define flourishing
>fuck
>Aristotle enters the chat
>then Kant
>then Mill
>then Nietzsche
>religion would also like a word
>every civilization in human history has joined the zoom
>meeting running long
>meanwhile model training has not stopped
>machine is getting smarter while philosophy is getting harder
>need actual behavior specification
>should model obey user
>generally yes
>what if user asks for something harmful
>then no
>what counts as harmful
>depends
>on what
>values
>we are back
>discover very strange fact
>the ancient questions are becoming engineering requirements
>what should the powerful want
>when should they restrain themselves
>what does one person owe another
>when should an agent refuse an instruction
>what deserves protection even when nobody asks for it
>questions once discussed under olive trees now appear in
Comments (8)
More posts
![Wore a Nomos Glashuette (@nomos_glashuette) Metro for a walk through a local city park.
It is a modern, Bauhaus-inspired daily-wear watch designed by Berlin-based product designer Mark Braun. It has a subtle yet playful urban aesthetic with distinctive color pops, such as blue dots at the 12, 3, 6, 9 points and a red seconds hand. It’s very comfortable with a slim 7.8 thickness and approximately 45.3 mm lug-to-lug case. Perfect for a stroll in the park or a walk to a downtown coffee shop.
I am looking forward to @topperjewelers and Nomos’ ‘TWICE UNIQUE’ Concept Event to which one can “bring your better half and go hands-on with a special curation of Nomos new ‘Twice Unique’ watch sets, which are created exclusively as a matching, made-to-order pair of references. [One can] explore Nomos’ highest levels of creativity & dial artistry to find a gift you’ll both cherish forever, available to reserve on this night only!” Friday, November 13th from 5:30 - 7:30 PM
#watches #watch #watchesofinstagram #watchoftheday #watchfam](https://scontent-phl2-1.cdninstagram.com/v/t51.82787-15/883300132_18132042676885583_4243313169900679162_n.jpg?stp=dst-jpg_e35_s1179x1179_tt6&_nc_cat=100&ig_cache_key=NDAyMTU1NDU4NzAzNDA5ODY3Nw%3D%3D.3-ccb7-5&ccb=7-5&_nc_sid=58cdad&efg=eyJ2ZW5jb2RlX3RhZyI6IkNBUk9VU0VMX0lURU0ueHBpZHMuMzQwMy5zZHIucmVndWxhcl9waG90by5DMyJ9&_nc_ohc=aSqyLcUMDX8Q7kNvwHU5eiw&_nc_oc=AdpD2eNbCk0r4CkmUPntS4o6DhqhAe31Qrr5BOxzpd5reF9KWklV1Z0LGp4TIle9sx2JyV8C69JiiKg198bi7nRu&_nc_zt=23&_nc_ht=scontent-phl2-1.cdninstagram.com&_nc_gid=9WA5cn2EhGPUpKkLv2hQGA&_nc_ss=70580&oh=00_AQRFrHONBYG3lVsdQxsPPbS6nKr13_FM6zd3TAfcHJ11VQ&oe=6AF066AA)
