codex Slate Star Codex


SSC Meetups 2019: Times And Places

Thanks to everyone who offered to host a meetup. Full list of cities, times, and places is below. If you’re reading this, you’re invited. Please don’t feel like you “won’t be welcome” just because you’re new to the blog, demographically different from the average reader, or hate SSC and everything it stands for. You’ll be fine!

Some suggestions for organizers:

1. Bring a sign that says SSC MEETUP so people can find you
2. Bring nametags and markers
3. Bring a signup sheet where people can write their names and emails if they want to hear about future meetups.
4. If people want to get to know each other better outside the meetup, you might want to mention, the rationalist friend-finder/dating site. It runs off Facebook, so you have to Facebook friend the other person first.
5. Please record how many people attend; I will ask for these numbers to help with future meetup posts.
6. If you take a picture and send it to me, I’ll try to post it here. I’ll ask for this later, please don’t email these to me until then.

If a city has **asterisks** around its name, nobody volunteered to host a meetup there, but SSC meetup director Mingyuan suspected you might have interested readers anyway. She chose a time, place, and random coffee shop as a Schelling point for these cities. There won’t be any organizers around, and no guarantee of anyone at all, but you can still go. If you go to these, in order to make it easier for everyone to find each other, try to arrive on time, wear something blue on top (shirt, hoodie, etc) as a signal, and/or bring SSC signs.

If a city’s name is in blue, that means I’ll be attending the meetup. I’ll try to announce each of these on the blog, and some of them might be subject to schedule change, so watch this space. A few people from various rationalist organizations interested in getting to know the further-flung parts of the community might be tagging along with me for some of the way.

Please look over your meetup to make sure it’s correct. If you have corrections, further details, or a new meetup to add, please post a top-level comment below containing the text “@MINGYUAN” so she can find it easily.


Albany, NY
Time: Saturday, September 21, 10:00 AM
Location: Stacks Espresso, 260 Lark St.
Contact: pellinorew[at]gmail

Amsterdam, Netherlands
Time: Saturday, September 28, 3:00 PM
Location: Grass in front of Weesperzijde 100
Contact: mathijs.henquet[at]gmail; Facebook event

Ann Arbor, MI
Time: Thursday, September 26, 6:30 PM
Location: TBD, check closer to the date for updates
Contact: algo2217[at]gmail

Atlanta, GA
Time: Saturday, October 12, 2:00 PM
Location: Hodgepodge Coffeehouse, 720 Moreland Avenue Southeast
Contact: steve[at]digitaltoolfactory[dot]net

Auckland, New Zealand
Time: Sunday, September 22, 4:00 PM
Location: Lim Chhour Food Court, 184 Karangahape Rd.
Contact: benjsmith[at]gmail; +64 22 685-4105; Facebook event
Details: Parking is available at Wilson’s Car Park on Cross Street. You may also find free street parking. There are excellent public transport links from many parts of the city. Please consider perusing the wonderful range of food options and purchasing a drink, snack, or meal to support our hosts. If you can’t fit it in the budget or just aren’t hungry, though, please come anyway as we’d love to have you!

Austin, TX
Time: Monday, September 30, 6:00 PM
Location: Central Market North Lamar, 4001 N Lamar Blvd.
Contact: azavoluk[at]gmail; Google group
Details: Either upstairs on the balcony or outside in the courtyard. Will email the google group, but we should be easy to spot.

Baltimore, MD
Time: Sunday, September 29, 7:00 PM
Location: 4th Floor, UMBC Performing Arts and Humanities Building
Contact: rivka[at]adrusi[dot]com

Bangalore, India
Additional search terms: Bengaluru
Time: Sunday, September 22, 4:00 PM
Location: Matteo Coffea, Church Street, Near MG road Metro Station
Contact: m.nihalmohan[at]gmail

Barcelona, Spain
Time: Saturday, September 21, 4:00 PM
Location: Plaça d’Espanya, 08004 Barcelona, Spain
Contact: ana_bcn[at]protonmail[dot]com
Details: We meet in Plaça d’Espanya at the big tower closest to Fira. Then depending on the weather we can go to Montjuic or Arenas. Feel free to send me an e-mail even if you can’t attend, so that we may catch up at another time.

Berkeley, CA
Time: Thursday, October 10, 6:00 PM
Location: “Event Horizon” (private residence), 2412 Martin Luther King Jr. Way
Contact: mingyuan[at]uchicago[dot]edu; Facebook event

Berlin, Germany
Time: Sunday, September 22, 7:00 PM
Location: Turmstr. 10, 10559 Berlin
Contact: ssc[at]alphabattle[dot]xyz
Details: The place is a club house for a club I’m in (JVB), so a store window, hard to miss. I’ll put up an SSC sign. Bus stops right in front, or subway to Turmstr or S-Bahn to Bellevue. You can buy drinks at ~cost. There are various takeaway options around if you want to bring something or order in. We’ll have the place to ourselves and can easily seat 30 people.

Boston, MA
Time: Friday, September 20, 7:00 PM
Location: 199 Harvard St Apt 2 (about 8 minutes walk from Central MBTA station)
Contact: boston-lw-organizers[at]googlegroups[dot]com

Boulder, CO
Time: Monday, September 16, 7:15 PM
Location: The Rayback Collective, 2775 Valmont Rd.
Details: I will have some sort of sign or folded paper that says “LW RRG” on it or something similar. Look for a tall man with long hair and a beard. We usually sit in the Southwest corner if possible, or in one of the couch circles along the west wall.

**Brighton, UK**
Time: Saturday, September 21, 2:00 PM
Location: Starbucks, 201 Western Rd.
Contact: None
Details: Meet in the upstairs seating area.

Brisbane, Australia
Time: Friday, October 11, 7:00 PM
Location: Level 1/210 Wickham St, Fortitude Valley QLD 4006, Australia
Contact: jarred.filmer[at]gmail
Details: The Boiler Room, Fortitude Valley, the bar behind the mural of Bill Murrary at Tomcat. Head up the stairs then curve left, you’ll see it.

Bristol, UK
Time: Thursday, September 26, 7:30 PM
Location: The Robin Hood, 56 St Michael’s Hill
Contact: Facebook event

**Brussels, Belgium**
Time: Saturday, September 21, 2:00 PM
Location: The Judgy Vegan, Rue des Capucins 55
Contact: None

**Budapest, Hungary**
Time: Saturday, September 21, 2:00 PM
Location: Starbucks, Egyetem tér 4, 1053 Hungary
Contact: None
Details: Meet in the outdoor seating area, if the weather allows

Buffalo, NY
Time: Sunday, September 29, 10:00 AM
Location: Five Points Bakery, 426 Rhode Island St.
Contact: spetey[at]gmail

Calgary, Canada
Time: Sunday, September 22, 1:00 PM
Location: Weeds Cafe, 1903 20 Ave NW
Contact: ssccalgary[at]yahoo[dot]com
Details: Please email me if you are going to attend. I live over an hour away, and would just like to make sure I don’t have to do the long drive if nobody will be attending, as the amount of interest in Calgary might be on the low side. Similarly, I would like to encourage anybody who is on the shy side, or who might be anxious about a small turnout to attend as well. To this end, if it looks like we’re going to have a group of 3 or less people, I will let you know beforehand and you can reconsider attending if you like; this way, you won’t have to worry that you’re going to be the only one to show up.

Cambridge, UK
Time: Wednesday, October 16, 7:00 PM
Location: The Burleigh Arms, 9-11 Newmarket Rd.
Contact: rlm72[at]cam[dot]ac[dot]uk

Canberra, NSW, Australia
Time: Wednesday, September 25, 6:00 PM
Location: The Snug Room at King O’Malley’s in Civic Centre
Contact: andy.bachler[at]gmail
Details: I have booked the Snug Room at King O’Malley’s which is a quiet room behind the bar. Should be low key and quiet enough to allow easy conversation.

Cape Town, South Africa
Time: Saturday, September 28, 11:00 AM
Location: Haas Coffee, 19 Buitenkant St, Cape Town City Centre
Contact: rashtayob[at]gmail

**Champaign-Urbana, IL**
Additional search terms: UIUC
Time: Saturday, September 21, 2:00 PM
Location: Illini Union Basement
Contact: None

Charlotte, NC
Time: Monday, September 16, 7:00 PM
Location: The Mighty Meeple, 8440 Pit Stop Ct NW #180, Concord, NC 28027
Contact: nospam[at]catallaxymedia[dot]com
Details: Mighty Meeple is a game store with a large backroom area of tables as well as a ton of free board games to play during this time frame for their Board Game Night. You can also bring your own food, although they have drinks and snacks available for purchase. My plan will be to sit at a long table with my young adult family and play games with a sign reading SSC so you can find us. There will be at least 4-6 people already. We can chat and/or play various games together. Please email me to introduce yourself or RSVP. I’ll reply-all just before so everyone attending knows what to expect.

Chicago, IL
Time: Saturday, September 28, 6:00 PM
Location: South Loop Strength & Conditioning, 645 S Clark St.
Contact: todd[at]southloopssc[dot]com
Details: The location is my gym, so we’ve got a large open space ripe for exercising socializing.

Cleveland, OH
Time: Saturday, September 28, 1:00 PM
Location: Tabletop Cleveland, 1810 W 25th St.
Contact: ferbfreeman[at]gmail

Cologne, Germany
Time: Saturday, September 21, 5:00 PM
Location: Marienweg 43, 50858 Köln
Contact: marcel_mueller[at]mail[dot]de
Details: Private venue, everyone who wants to attend is welcome. If you are unable to find us call +491788862254.

Columbus, OH
Time: Sunday, September 22, 6:00 PM
Location: 397 Eldridge Ave.
Contact: james.thomas.hays[at]gmail
Details: Private residence. Food will be provided. RSVPs preferred to help with planning purposes, but not required.

Copenhagen, Denmark
Time: Saturday, September 21, 7:00 PM
Location: Studenterhuset, Købmagergade 52, Denmark
Contact: soeren.elverlin[at]gmail

Dallas, TX
Time: Sunday, September 15, 5:00 PM
Location: Magic Cup Cafe, North Jupiter Road, Richardson, TX
Contact: tayfie[at]pm[dot]me; LessWrong

Denver, CO
Time: Saturday, October 5, 6:00 PM
Location: 1520 1/2 Lowell Blvd
Contact: fsjdnjdnsjkuuoqq[at]aol[dot]com

Dublin, Ireland
Time: Saturday, September 21, 2:00 PM
Location: Starbucks, Stephen Court, GF1, St Stephen’s Green, Dublin 2, D02 N960, Ireland
Contact: enagohan[at]gmail; Facebook event; LessWrong

**Edinburgh, Scotland**
Time: Saturday, September 21, 2:00 PM
Location: Starbucks, 124 High St, Edinburgh EH1 1SG, UK
Contact: None
Details: Meet in the upstairs seating area

Fairbanks, AK
Time: Sunday, October 6, 9:00 PM
Location: Starbucks, 3755 Airport Way
Contact: scott[at]slatestarcodex[dot]com
Details: May or may not happen; await details on blog the day before

Halifax, Nova Scotia, Canada
Time: Saturday, September 21, 7:00 PM
Location: Uncommon Grounds, 1030 South Park Street
Contact: ahobata[at]gmail

Hamburg, Germany
Time: Saturday, September 21, 5:00 PM
Location: Troy Office, Maienweg 310, 22335 Hamburg, Germany
Contact: gunnar.zarncke[at]gmx[dot]de
Details: We are meeting in the small company office of the fintech startup troy in the 4th floor. Just ring at “troy”. Depending on the weather we can sit on our nice balcony. It would be nice if guests could bring blankets. This is not an official company event but privately organized. Some foods and drinks might be provided, but I’d appreciate if you contribute. There will be some prepared ice breakers and discussion topics, but you are free to contribute or just meet and talk.

Helsinki, Finland
Time: Tuesday, September 24, 6:00 PM
Location: Restaurant Dubliner Kaivopiha, Mannerheimintie 5
Contact: joenash499[at]gmail
Details: We will be in the private room called Guinness Lounge. You can expect something like a normal night at the pub, except with a group consisting of other SSC readers (and without the background noise, since we’ll be in a private room.) You don’t have to contact anyone in advance, you can just show up.

**Honolulu, HI**
Time: Saturday, September 21, 2:00 PM
Location: Starbucks, 680 Ala Moana Blvd A101
Contact: None

Houston, TX
Time: Sunday, September 15, 2:00 PM
Location: Brasil Cafe, 2604 Dunlavy St.
Contact:[at]gmail; LessWrong; Facebook group; Google group; Discord server (link expires after 5 uses, please contact me if you want to join our Discord but the link has stopped working)

Indianapolis, IN
Time: Saturday, September 21, 4:00 PM
Location: ShoeFly Public House, 122 East 22nd Street
Contact: molybdenumblue[at]gmail

Iowa City, IA
Time: Sunday, September 15, 1:00 PM
Location: The Java House, 211 E Washington St.
Contact: potentgamer[at]gmail

Irvine, CA
See “Orange County”, below.

**Jacksonville, FL**
Time: Saturday, September 21, 2:00 PM
Location: Starbucks, 1980 San Marco Blvd
Contact: None

Kansas City, KS
Time: Sunday, September 22, 12:30 PM
Location: Panera Bread, 10606 Shawnee Mission Pkwy, Shawnee, KS
Contact: alex.hedtke[at]gmail; Facebook event

Kiev, Ukraine
Additional search terms: Kyiv
Time: Saturday, September 14, 4:00 PM
Location: On the steps to the National Museum of Nature (Underground station “Teatral’na”)
Contact: Marichka, (+38)0976672970
Details: We shall move from there, so better don’t be late.

Kitchener-Waterloo, ON
Time: Sunday, September 29, 2:00 PM
Location: Field of Waterloo CI
Contact: blxxia[at]gmail
Details: I’ll wear red converse and bring some frisbees for those that are athletically (or simply disc-) inclined. We’ll walk the half a block to my apartment at around 3:30ish and have the rest of the meetup there with tea and snacks, if space allows.

Lexington, KY
Additional search terms: Louisville, Cincinnati, Frankfort, Covington, Shelbyville
Time: Sunday, September 22, 7:00 PM
Location: West Sixth Brewery, 501 West Sixth Street, Suite 100
Contact: hellernathanssc[at]gmail
Details: Please email if you’re interested.

Lisbon, Portugal
Time: Sunday, October 6, 4:30 PM
Location: Oriente Station
Contact: ana_bcn[at]protonmail[dot]com;
Details: We meet in the Oriente station in front of the book fair. Then depending on the weather we can go to Parque das Nações or Vasco da Gama.

London, UK
Time: Sunday, September 15, 2:00 PM
Location: Craft Beer Co. Islington, 55 White Lion Street (we have a large room booked)
Contact: philip.hazelden[at]gmail; Facebook event; Facebook group

Los Angeles, CA
Time: Wednesday, October 9, 7:00 PM
Location: 10850 West Pico Boulevard, Los Angeles, CA – The Wine Bar on the 2nd floor of the Westside Pavilion, right next to the Landmark Theater.
Contact: bobert.mushky[at]gmail; Discord server
Details: Also consider coming to the Orange County meetup to meet me [Scott] two days later

Madison, WI
Time: Saturday, September 21, 5:00 PM
Location: 1022 High St.
Contact: mmwang[at]wisc[dot]edu; Facebook event
Details: This is a house (newly painted blue with a red door), very close to the #4 and #47 bus stop. Also the #13 and #5 stop close by. If you need a ride, leave a message on the Facebook event and someone will get you. Some basic vegan food will be served. Bring food for yourself or to share if so inclined.

Madrid, Spain
Time: Saturday, September 21, 4:00 PM
Location: Ferros Café, Calle del Maestro Ángel Llorca, 9, 28003 Madrid
Contact: pavill01[at]ucm[dot]es;
Details: We speak both English and Spanish. About a week before I will post the event details in the Meetup group.

**Manchester, UK**
Time: Saturday, September 21, 2:00 PM
Location: The grass patch next to the Wharf Pub in city center M15 4ST, coordinates 53.473975, -2.256884
Contact: None

Medellín, Colombia
Time: Sunday, September 22, 11:00 AM
Location: Café Revolución, Carrera 73 | Cq. 4 #10, Laureles, Medellín, Colombia
Contact: richard.meadows[at]thedeepdish[dot]org

Melbourne, Australia
Time: Friday, October 4, 6:00 PM
Location: Queensberry Hotel, 593 Swanston St, Carlton VIC 3053
Contact: 0438 869 257; Facebook group

Memphis, TN
Time: Saturday, October 5, 5:00 PM
Location: French Truck Coffee at Crosstown Concourse
Contact: michael[at]postlibertarian[dot]com
Details: It’s in the main Central Atrium of Crosstown Concourse. There are many chairs and tables right in front so we should be able to grab a few. I’m totally ok if only a couple people show up!

Miami, FL
Time: Saturday, October 12, 5:00 PM
Location: Pasion del Cielo (Brickell City Centre location), 701 S Miami Ave #350B
Contact: eric135033[at]gmail; Facebook group

Minneapolis, MN
Additional search terms: St. Paul, Twin Cities
Time: Saturday, September 21, 3:00 PM
Location: The Knoll, 27 Pleasant St SE
Contact: grahamsnumberisbig[at]gmail; LessWrong
Details: This is a park on the north end of the University of Minnesota’s Minneapolis campus, not the apartment building of the same name. Although the forecast looks good right now, in the event of poor weather we may have to relocate – email me if you’re interested so I can get a rough headcount and send updates more easily. Also see the LessWrong post for an editable post that I’ll try to keep up to date if things change.

Montreal, Canada
Time: Sunday, September 22, 2:00 PM
Location: Loft 111A, 5455 rue d’Iberville
Contact: 1806992[at]gmail; Yohan 514-562-7682; Facebook event
Details: At the front door, dial 800. The meeting place will be at the collective l’Infini, which recently started hosting LW meetups twice a month. Can comfortably host 30 people. Cozy and quiet. Note that the space welcoming the meetup is a sober space, so no drug/alcohol consumption inside. All are welcome to bring/cook food.

Moscow, Russia
Time: Saturday, September 21, 6:00 PM
Location: Rationality Center Kocherga, 5k2, B.Dorogomilovskaya st., Kievskaya, Moscow, Russia
Contact: oknezrob[at]kocherga-club[dot]ru

Munich, Germany
Time: Thursday, September 19, 7:00 PM
Location: Augustiner-Keller, Arnulfstraße 52, 80335 München, Germany
Contact: rlm72[at]cam[dot]ac[dot]uk

Nashville, TN
Time: Tuesday, September 17, 5:00 PM
Location: Frothy Monkey, 12th Avenue South
Contact: james[at]writechem[dot]com
Details: We will be discussing Meditations On Moloch. Any new members who wish to attend, please reach out to the contact given.

New Delhi, India
Time: Saturday, September 14, 5:00 PM
Location: Brown Box Cafe, Shop No. 2, DDA Market, S Block, Panchsheel Park, New Delhi, Delhi 110017
Contact: anujdayal[at]gmail

New Haven, CT
Additional search terms: Yale
Time: Saturday, September 21, 4:00 PM
Location: If the weather is good: the courtyard outside of P&M Market (on the corner of Orange Street + Cottage Street); If it’s raining: Blue State Coffee on Orange Street.

New York City, NY
Time: Saturday, September 21, 4:00 PM
Location: Pumphouse Park
Contact: tristan[at]thume[dot]ca; 646-480-0618
Details: If it’s raining or we’re otherwise not there go to the main atrium of Brookfield Place, which is adjacent to the park

Norman, OK
Time: Saturday, September 28, 2:00 PM
Location: University of Oklahoma, Bizzell Memorial Library, Room 421
Details: I’m not posting any contact info here, so here are my detailed contingency plans: if nobody shows up to the meetup within one hour, I will leave. If room 421 is for some reason unavailable we will meet in the reading area inside the front (south) doors, north of the front desk. I will fashion a small SSC sign out of a piece of paper or something.

Northampton, MA
Time: Saturday, September 21, 6:30 PM
Location: Packard’s Bar and Restaurant, main dining room
Contact: alex[at]alexliebowitz[dot]com; LessWrong

Oklahoma City, OK
Time: Saturday, September 21, 2:00 PM
Location: Del City Library
Contact: battleshipbean[at]gmail
Details: I’ll be wearing a USS Iowa shirt and hat.

Orange County, CA
Additional search terms: Irvine, Fullerton, Santa Ana, Anaheim, Huntington Beach
Time: Friday, October 11, 7:30 PM
Location: Underneath this mysterious hexagonal sigil at the University Center food court in Irvine, California.

Oslo, Norway
Time: Monday, September 23, 6:00 PM
Location: Starbucks at Aker Brygge, Stranden 7, 0250 Oslo, Norway
Contact: anders[at]huitfeldt[dot]net

Ottawa, Canada
Time: Friday, September 20, 7:00 PM
Location: The Fox & Feather, 283 Elgin St.
Contact: walsh.the[at]gmail; Facebook group
Details: We will book one of the private rooms.

Oxford, UK
Time: Thursday, September 19, 6:00 PM
Location: The Gardener’s Arms, 39 Plantation Road, Oxford, OX2 6JE
Contact: ???

Paris, France
Time: Sunday, September 15, 3:00 PM
Location: 48°51’37.9″N 2°17’11.8″E (
Contact: 336.vingt-trois.dix-huit.82.99; Discord server
Details: I hope the weather is going to be nice. Last year it started to be a bit cold after some time, so take a sweatshirt just in case. If the forecast is bad, we’ll change the place to some coffee shop, maybe not in the same area, so check the page the day before.

Philadelphia, PA
Time: Sunday, September 22, 6:30 PM
Location: Tir Na Nog Bar & Grill, 1600 Arch Street
Contact: wfenza[at]gmail; dmcbriggs[at]gmail
Details: The restaurant is inside the Phoenix Hotel, directly above Suburban Station. We have the entire outer vestibule area reserved

Phoenix, AZ
Time: Saturday, September 21, 2:00 PM
Location: Lola Coffee Downtown, 1001 N 3rd Ave.
Contact: benjamin.j.morin[at]gmail
Details: Parking is typically available on weekends, this is also ~5 min walk from the light rail. Contact me if you are coming and I can give you a phone number to help with parking on the day of.

**Pittsburgh, PA**
Time: Saturday, September 21, 2:00 PM
Location: Starbucks in Omni William Penn Hotel
Contact: None

Portland, OR
Time: Tuesday, October 1, 7:00 PM
Location: Cartopia, 1207 SE Hawthorne Blvd
Contact: nwalton125[at]gmail; LessWrong
Details: Cartiopia is a food cart pod with many options — outside, but covered from rain. I’ll get there at 7 PM (or a little earlier), try to stake out table space, and put up a sign for the meetup. The carts are open as late as midnight.

Prague, Czechia
Time: Wednesday, September 25, 6:30 PM
Location: Rationalist coworking space Epistea (walk through tea room dharmasala to get there), Peckova 15, Karlínské náměstí
Contact: nadvornik.jiri[at]gmail; +420 607 690 860; Facebook event
Details: Food will be provided. English friendly.

Pune, India
Time: Sunday, September 22, 6:00 PM
Location: Cafe Coffee Day, IISER Pune Campus, Baner
Contact: 7rat13[at]gmail

Reno, NV
Time: Saturday, September 28, 2:00 PM
Location: Reno Coffee Co, 1300 S Wells
Contact: supervexi[at]gmail
Details: Not sure how much of an SSC following we have here, so please be sure to indicate your interest by emailing me

Research Triangle, NC
Additional search terms: Raleigh, Durham, Chapel Hill
Time: Wednesday, September 18, 7:00 PM
Location: Ponysaurus Brewing Company, 219 Hood St, Durham, NC 27701
Contact: willdjarvis[at]gmail

Riga, Latvia
Time: Saturday, September 28, 1:00 PM
Location: Caffeine, Audēju iela 15
Contact: cerulean.lemniscate[at]protonmail[dot]com

**Rio de Janeiro, Brazil**
Time: Saturday, September 21, 2:00 PM
Location: Starbucks, Av. Rio Branco, 143 – A – Centro, Rio de Janeiro – RJ, 20040-006, Brazil
Contact: None
Details: Meet in the upstairs seating area

**Rochester, NY**
Time: Saturday, September 21, 2:00 PM
Location: SPoT Coffee, 200 East Avenue
Contact: None

**San Antonio, TX**
Time: Saturday, September 21, 2:00 PM
Location: Local Coffee, 7338 Louis Pasteur Dr #204
Contact: None

Time: Saturday, September 21, 2:00 PM
Location: Starbucks outside Clarke Quay Station, 6 Eu Tong Sen St, #01-29 The Central, Singapore 059817
Contact: None

St. Louis, MO
Additional search terms: Saint Louis
Time: Saturday, September 21, 3:00 PM
Location: Hartford Coffee, 3974 Hartford St
Contact: littlejohnburidan[at]gmail
Details: I have a place to stay if you visit from out of town and need a bed.

Sacramento, CA
Time: Sunday, September 22, 11:00 AM
Location: Crepeville, 1730 L St.
Contact: amethyst.eggplant[at]gmail

Salt Lake City, UT
Time: Saturday, October 5, 3:00 PM
Location: 41 S 900 E #204
Contact: wearenotsaved[at]gmail

San Diego, CA
Time: Sunday, September 29, 3:00 PM
Location: La Jolla Colony Park
Contact: owenshen24[at]gmail; Facebook group
Details: Also consider coming to meet me [Scott] at the Orange County meetup October 11th

San Francisco, CA
Time: Monday, September 23, 6:30 PM
Location: 170 Hawthorne St.
Contact: rocurley[at]gmail; LessWrong
Details: For help getting into the building, please call (or text, with a likely-somewhat-slower response rate): 301-458-0764. We’ll be meeting to give and listen to very short talks! We’ll do 7-minute lightning talks with 3 additional minutes allowed for questions. A talk doesn’t have to be formal, planned, or even something that you’d expect someone to Give A Talk About; it can be as simple as telling the group about something you find interesting or cool.

San Jose, CA
Additional search terms: South Bay
Time: Saturday, September 28, 2:00 PM
Location: 3806 Williams Rd.
Contact: ; Website

**San José, Costa Rica**
Time: Saturday, September 21, 2:00 PM
Location: Starbucks, Costado Oeste del Estadio Nacional, en edificio Clínica 20 20 San José Sabana Oeste, Nunciatura, San José, Costa Rica
Contact: None

São José dos Campos, Brazil
Time: Sunday, September 15, 2:00 PM
Location: Parque Santos Dumont, next to the 14-Bis airplane replica
Contact: emiliobumachar[at]gmail
Details: I will meet even one person in good faith, but will not go if no one contacts me.

Seattle, WA
Time: Wednesday, October 2, 6:30 PM
Location: The Territory (house), 5238 11th Ave NE
Contact: orborde[at]gmail; Facebook event

Seoul, South Korea
Time: Saturday, September 28, 2:00 PM
Location: O’sulloc Green Tea near Gangnam Subway Station
Contact: curtis95112[at]gmail
Details: I would also appreciate it if people who can’t/won’t make the meetup contacted me too, possibly anonymously, because it would be fascinating if nobody replied and I were literally the only SSC reader in a city of 20 million.

Stockholm, Sweden
Time: Saturday, September 14, 3:00 PM
Location: Grillska Huset, Stortorget 3, Stockholm

Sydney, Australia
Time: Thursday, September 19, 6:00 PM
Location: Level 2, City of Sydney RSL, 565 George St
Details: We sit at a table in the fish bowl across the room from the lifts.

Tampa, FL
Time: Saturday, September 21, 5:00 PM
Location: Felicitous Coffee, 11706 N 51st St, Temple Terrace
Contact: danielbporter[at]gmail

Tel Aviv, Israel
Time: Tuesday, September 24, 7:00 PM
Location: Google Israel office, Electra Tower, 12th floor, Yigal Alon 98
Contact: highpriestessofelua[at]gmail; Facebook event

Toronto, Canada
Time: Monday, September 23, 7:00 PM
Location: Madison Avenue Pub
Contact: team[at]thinkbetter[dot]ca; Facebook event

Toulouse, France
Time: Monday, October 7, 7:00 PM
Location: Eurêkafé, 24 Rue Léon Gambetta
Contact: me[at]jrmyp[dot]net

Vancouver, Canada
Time: Sunday, September 29, 1:00 PM
Location: Picnic area at Trout Lake Park, near the south beach
Contact: emily9595[at]gmail

Victoria, BC, Canada
Time: Saturday, September 21, 2:00 PM
Location: Starbucks, 1301 Government St, Victoria, BC V8W 1Y9, Canada (downtown, at Yates and Government)
Contact: graeme.andrew.hill[at]gmail
Details: I’ll be the lanky white dude with a buzzcut wearing all blue.

Vienna, Austria
Time: Saturday, October 19, 3:00 PM
Location: Kaisermühlenstraße 24, A-1220 Vienna
Contact: strivingforconsistency[at]gmail; Facebook group
Details: We’ll be in the meetup room on the ground floor at the back side of the building. Event and directons to the location will be announced on the Facebook group.

**Warsaw, Poland**
Time: Saturday, September 21, 12:00 PM
Location: Starbucks, plac Trzech Krzyży 16A, 00-499 Warszawa, Poland
Contact: None

Washington DC
Time: Tuesday, September 24, 5:00 PM
Location: Teaism coffee shop, 400 8th Street NW, Washington DC
Contact: robirahman94[at]gmail; Google group

Wellington, New Zealand
Time: Saturday, September 28, 3:00 PM
Location: Glover Park
Contact: benwve[at]gmail

West Lafayette, IN
Additional search terms: Purdue
Time: Sunday, October 6, 6:00 PM
Location: Cafe Literato, 1920 Northwestern Ave.
Contact: lancesbush[at]gmail; Facebook group; Facebook account

Zurich, Switzerland
Time: Saturday, October 5, 3:00 PM
Location: Kosmos at Lagerstrasse 104
Contact: ssczurich[at]gmx[dot]ch; Facebook group

Posted in Uncategorized | Tagged | 47 Comments

Lots Of People Going Around With Mild Hallucinations All The Time

[Related to: Relaxed Beliefs Under Psychedelics And The Anarchic Brain, HPPD And The Specter Of Permanent Side Effects]


Hallucinogen persisting perceptual disorder is a condition where people who take psychedelics continue hallucinating indefinitely. Estimates of prevalence range from about 4% of users (Baggott) to “nobody, the condition does not exist” (Krebs and Johansen). To explore this discrepancy, I asked about it on the 2019 SSC survey. The specific question was:

Hallucinogen Persisting Perceptual Disorder is a condition marked by visual or other perceptual disturbances typical of psychedelic use that continue for weeks and months after coming off the psychedelic, in some cases permanently. Have you ever had this condition?

2,234 readers admitted to having used psychedelics. Of those, 285 (= 12.8%) stated that they had some hallucinations that persisted afterwards. 219 (9.8%) said they’d had them for a while and then they had gone away. 66 (= 3%) stated that they still had the hallucinations (one limit of the study: I don’t know how long it has been since those people took the psychedelics).

But most of these people reported very mild experiences; on a scale from 1-10, the median severity was 2. The most commonly reported changes were more “visual snow” (ie “static” in the visual field), slight haloes around objects, and visual trails. Many people reported that if they stared at a textured area like a carpet long enough, illusory geometric patterns would start to form. Only a few people noticed anything weirder than this. Here are some sample responses:

— It took the form of visual disturbances (patterns, moving patterns, mild hallucinations like seeing a flower growing on a person’s face or seeing the legs of a chair walk). Initially there would be daily incidents. They seemed to be triggered by fatigue or other stress, and I did not like them, which lead to more stress. Exercise helped reduce the stress and make the visual effects go away. The severity and frequency dropped quickly at first and then more slowly over the years, essentially gone after a decade. I have had migraines with ‘fortification figures’ as well; this was different from that.

— Perceptual distortions about once a week. Difficulty concentrating for about 6 months. Intense interest in drawing.

— Used LSD about twice a month for a period of time when I was 17, and had a very very slight build up of persisting visual distortions. This phase culminated in me taking ~425µg by accident one night, and having a somewhat negative experience, had an ego death sort of thing, experienced strong “time-warping” where my subjective experience of the passing of time slowed down and sped up repeatedly by a large degree. Very strong acid trip. After that night I experienced pretty noticeable breathing on certain textures, like ceiling tiles, and a sort of “bright aura effect” on others, like pebbles. This gradually receded over something like 1-2 years, at which point it was entirely gone.

— straight lines often bend! it’s distracting!

— If I stare at an object, it’ll appear to slightly sway or move in a circular motion. If I stare at a surface (especially a textured one like a carpet), faint light patterns will develop. I got HPPD the 3rd time I did LSD. I haven’t done anything to treat it, expecting it would go away on its own over time. However, it’s been like 4 years and hasn’t really changed much. I’ve gotten used to it and often forget I have HPPD. I told an optometrist about it one time and he was just like “yeah, happens to some people who do acid.”

— Quite concerned. Seen by PCP and opthamologist. Performed a literature review myself and tried clonidine without a large effect. Symptoms have seemed to resolve however.

— Persistent discolouration blue-ward or red-ward, depending on *subjective* head orientation, in a pattern of visual field roughly equivalent to the Tetragrammaton. Yes bloody really. Help.

— My colour perception was distorted and confusing for some time, perhaps 1-3 years with browns, greens and reds, especially in natural settings (2C* series) I also experienced intense flashbacks consisting of bodily tingling and diffuse feeling of heat for ~10 months after a different psychedelic session.

— I see tracers, halos, have increased anxiety, visual snow. Nothing has helped so far.

— Very mild but persistent visual distortions. Walls breathing, kaleidoscoping of repeating patterns, etc. Not severe enough to warrant treatment.

— Any highly detailed or repetitive surface “breathes” after looking at it for even a few seconds. get glowing patterns when looking at some art, a sort of cycle of different colors being highlighted in sequence. makes eyes open meditation easier; makes “where’s waldo” harder.

— Visual snow; when viewing images which contain several different possible interpretations of patterns (eg hex grid on bathroom tile, carpet weaving pattern) my perception of the most salient pattern stutters and vacillates between the top 3-5 possible interpretations.

— I hesitated whether to even answer ‘yes’. 15-25 years ago, a period when I occasionally used psychedelics, I would sometimes have mild visual trails or aberrations when I wasn’t on any psychedelics. It never really bothered me. I can still sometimes make stationary patterns swirl or move by staring at them, but I think that’s just normal visual perception, not HPPD. Right?

A Reddit user helpfully illustrated what his (particularly severe) HPPD looked like (note especially the subtle square grid in left picture):

Weird, but not unbearably so. I think this explains the wide variety of prevalence estimates. Many people who take psychedelics will have very minor permanent changes to their vision. Most of them will shrug and get on with their lives. Only the few people whose changes are especially bad, or who get especially neurotic about it, will ever talk to a doctor.

I discussed some of this recently at a colloquium of experienced psychonauts (realistically: a group house full of the kind of programmers who go to Burning Man). Of the ten or so people there, two admitted to HPPD which sounded about as bad as the image above. They’d never really mentioned it before, it hadn’t come up, and they were living normal lives.

I conclude that lots of people you encounter are having mild hallucinations all the time, and it just never comes up.


I want to talk more about Krebs & Johansen, the paper that claims to find HPPD doesn’t exist.

First, it talks about surveys of special populations, like Native Americans using peyote. None of them report HPPD. Plausibly peyote does not cause HPPD (most of the people on my survey blame LSD) or something about the cultural set and setting prevents Native Americans from getting it during ceremonies. I see suspects this too.

Second, it talks about K&J’s own study, based in the National Survey of Drug Use and Health. Among 130,152 participants, those who have used psychedelics are no more likely to report symptoms than those who have not. But this survey violates one of the cardinal sins of psychiatric questioning – be careful in how you ask questions about psychosis-like symptoms! (see here for more people getting this wrong). K&J base their finding entirely on how people answer the question “Saw vision others could not?” (they don’t give the presumably longer question for which this is an abbreviation). But it is asked after several other questions clearly probing psychotic symptoms, like “felt force taking over your mind?”, “felt force inserting thoughts?”, and “heard voices?” Participants probably assumed (I think correctly; K&J did not write the survey) that the question was asking about more dramatic visions, like visions from God about their destiny or something. The average person whose visual field was 5% more staticky than normal wasn’t going to answer.

Third, it talks about HPPD-like visual hallucinations in people who have never taken psychedelic drugs. It says many people have them. I believe this. At the colloquium I mentioned, someone talked about having some similar staticky symptoms before taking any psychedelics. On my own survey, a few people said things like “I could always make the carpet self-organize into geometric patterns if I tried really hard, but after I took LSD it started happening much more”. I don’t know of any surveys on how prevalent this is, but since nobody has any trouble finding examples, I imagine it’s pretty common. My only argument that this doesn’t disprove HPPD is how many people – including my survey respondents – describe these oddities starting (or getting much worse) right after they use psychedelics. But we have no idea how many totally drug-naive people are just going around with weird visual experiences just because that’s the kind of person they are.


I want to talk about a case study of this last kind of person (as usual, this is a composite with some details changed to protect privacy). Patient of mine for a totally unrelated reason – she had mild depression after a death in the family, wanted to talk it over. Gradually it started coming out that she had an entire spiritual system based on what part of a grid she thought people fell on. It took me a long time to figure out the grid was literal. It was a shape she could see in her vision, all the time, close to a Cartesian plane but with a couple of extra lines and curves on it. It looked similar in spirit to the grid pattern tiled across the sample image above, though not exactly the same.

In terms of her personality – she was very into William Blake, Meister Eckhart, and ecstatic poetry. She dabbled in Wicca and occultism. She had strong opinions on chakras. Her hair was dyed an even brighter shade of purple than the Bay Area median. I was mildly concerned she was going to either join or start a cult, but she never went further than sharing some of her manifestoes about the nature of time-space with her lucky psychiatrist.

This woman had never used psychedelics. But her description of her visual grid matched some HPPD phenomena pretty well.

I still dream of linking individual differences in perception with differences in cognition. So I wonder: psychedelics make you more likely to believe woo, and more likely to have certain classes of visual hallucinations. But some people believe woo without taking psychedelics, and some people have those same visual hallucinations without taking psychedelics. Are they the same people? Do psychedelics move you along a natural axis of variation, maybe from unusually strong priors to unusually relaxed priors? Was my patient the sort of person that the rest of us would become if we took massive doses of LSD every day for years? Did that explain both her perceptual and her cognitive uniqueness?

Start thinking along these lines, and some other become pretty suggestive. How come some New Agey people say they can see auras around people? Are they just lying? Seems like a weird thing to lie about. And a lot of these people don’t sound like they’re lying. Aren’t auras a classic LSD hallucination? I understand they’re not quite as simple as the haloes around lights that HPPD people get; they’re only around people and sometimes the colors seem meaningful. But add something about handwave handwave using a special kind of visual processing for other people handwave synaesthesia, and maybe it’s not totally outside the realm of possibility? Maybe if your priors are so relaxed – ie so far towards the “naturally on LSD all the time” side of the scale – that you believe in auras, then your priors might also be so relaxed that you can see them.

We really don’t know what other people’s perceptions are like. At least, not until I ask a bunch of questions about this on the next SSC survey.

Posted in Uncategorized | Tagged , | 140 Comments

SSC Journal Club: Relaxed Beliefs Under Psychedelics And The Anarchic Brain

[Thanks to Sarah H. and the people at her house for help understanding this paper]

The predictive coding theory argues that the brain uses Bayesian calculations to make sense of the noisy and complex world around it. It relies heavily on priors (assumptions about what the world must be like given what it already knows) to construct models of the world, sampling only enough sense-data to double-check its models and update them when they fail. This has been a fruitful way to look at topics from depression to autism to sensory deprivation. Now, in Relaxed Beliefs Under Psychedelics And The Anarchic Brain: Toward A Unified Model Of The Brain Action Of Psychedelics, Karl Friston and Robin Carhart-Harris try to use predictive coding to explain the effects of psychedelic drugs. Then they use their theory to argue that psychedelic therapy may be helpful for “most, if not all” mental illnesses.

Priors are unconscious assumptions about reality that the brain uses to construct models. They can range all the way from basic truths like “solid objects don’t randomly disappear”, to useful rules-of-thumb like “most get-rich-quick schemes are scams”, to emotional hangups like “I am a failure”, to unfair stereotypes like “Italians are lazy”. Without any priors, the world would fail to make sense at all, turning into an endless succession of special cases without any common lessons. But if priors become too strong, a person can become closed-minded and stubborn, refusing to admit evidence that contradicts their views.

F&CH argue that psychedelics “relax” priors, giving them less power to shape experience. Part of their argument is neuropharmacologic: most psychedelics are known to work through the 5-HT2A receptor. These receptors are most common in the cortex, the default mode network, and other areas at the “top” of a brain hierarchy going from low-level sensations to high-level cognitions. The 5-HT2A receptors seem to strengthen or activate these high-level areas in some way. So:

Consistent with hierarchical predictive processing, we maintain that the highest level of the brain’s functional architecture ordinarily exerts an important constraining and compressing influence on perception, cognition, and emotion, so that perceptual anomalies and ambiguities—as well as dissonance and incongruence—are easily and effortlessly explained away via the invocation of broad, domain-general compressive narratives. In this work, we suggest that psychedelics impair this compressive function, resulting in a decompression of the mind-at-large—and that this is their most definitive mind-manifesting action.

But their argument also hinges on the observation that psychedelics cause all the problems we would expect from weakened priors. For example, without strong priors about object permanence to constrain visual perception toward stability, we would expect the noise of the visual sensorium to make objects pulse, undulate, flicker, or dissolve. These are some of the most typical psychedelic hallucinations:

consider the example of hallucinated motion, e.g., perceiving motion in scenes that are actually static, such as seeing walls breathing, a classic experience with moderate doses of psychedelics. This phenomenon can be fairly viewed as relatively low level, i.e., as an anomaly of visual perception. However, we propose that its basis in the brain is not necessarily entirely low-level but may also arise due to an inability of high-level cortex to effectively constrain the relevant lower levels of the (visual) system. These levels include cortical regions that send information to V5, the motion-sensitive module of the visual system. Ordinarily, the assumption “walls don’t breathe” is so heavily weighted that it is rendered implicit (and therefore effectively silent) by a confident (highly-weighted) summarizing prior or compressive model. However, under a psychedelic, V5 may be forced to interpret increased signaling arising from lower-level units because of a functional negligence, not just within V5 itself, but also higher up in the hierarchy. Findings of impaired high- but not low-level motion perception with psilocybin could be interpreted as broadly consistent with this model, namely, pinning the main source of disruption high up in the brain’s functional hierarchy.

But F&CH are most interested in whether psychedelics can cause the positive effects we would expect of relaxed priors. If overly strong priors cause closed-mindedness, psychedelics should allow users to “see things with new eyes” and change their minds about important issues. These changes would be precipitated by the drug, but not fundamentally about the drug. For example, imagine a person who formed a strong prior around “I am a failure” at a young age, then later went on to achieve great things. Because their prior was so strong, they might think of each of their accomplishments as a special case, or interpret them as less impressive than they were. On psychedelics, they could reexamine the evidence unbiased by their existing beliefs, determine that their accomplishments were impressive enough to count as successes, and abandon the “I am a failure” prior. They would continue understanding that they were successful even after they sobered up, because the change of mind was a triumph of rationality and not a drug-fueled hallucination.

When I was young, I liked a fantasy book called The Sword of Shannara. The titular sword had an unusual magic power: it made the wielder realize anything he already knew. Neither a truly good person nor a truly bad person would benefit from the sword, but someone who was hypocritical, or deluding themselves, or complicit in their own brainwashing, would find all the parts of their mind flung together so hard that they couldn’t help but realize all the inconvenient facts they were trying to repress, or connect all the puzzle pieces previously scattered in separate mental compartments. Sometimes this resulted in a blinding revelation that you were on the wrong side, or had wasted your life; other times in nothing at all. F&CH argue that psychedelics are the real-life version of this, a way to make all of your beliefs connect with each other and see what results from the reaction.

These dignified scientists don’t like magic-sword related analogies, so they stick to regular-sword-related ones. “Annealing” is a concept in metallurgy where blacksmiths heat a metal object in a forge until it undergoes a phase change. All the molecules move to occupy whatever the lowest-energy place for them to occupy is, strengthening the metal’s structure. Then the metal object leaves the forge and the metal freezes in the new, better configuration.

They analogize the same process to “flattening an energy landscape”. Imagine a landscape of hills and valleys. You are an ant placed at a random point in the landscape. You usually slide downhill at a certain rate, but for short periods you can occasionally go uphill if you think it would help. Your goal is to go as far downhill as possible. If you just follow gravity, you will end up in a valley, but it might not be the deepest valley. You might get stuck at a “local minimum”; a valley deep enough that you can’t climb out of it, but still not as deep as other places in the landscape you will never find. F&CH imagine a belief landscape in which the height of a point equals the strength of your priors around that belief. If you settle in a suboptimal local minimum, you may never get out of it to find a better point in belief-space that more accurately matches your experience. By globally relaxing priors, psychedelics flatten the energy landscape and make it easier for the ant to crawl out of the shallow valley and start searching for even deeper terrain. Once the drugs wear off, the energy landscape will resume its normal altitude, but the ant will still be in the deeper, better valley.

Here F&CH are clearly thinking of recent research that suggests MDMA treats post-traumatic stress disorder. Post-traumatic stress disorder is well-modeled as a dysfunctional prior, something like “the world is unsafe” or “you’re still in that jungle in ‘Nam, about to be ambushed”. Many PTSD patients have gone on to live good lives in well-functioning communities and now have more than enough evidence that they are safe. But the evidence doesn’t “propagate”; the belief structure is “frozen” in place and cannot be updated. If psychedelics relax strong priors, they can flatten the energy landscape and allow the ant of consciousness to leave the high-walled “the world is unsafe” valley and test the terrain in “the world is actually okay”. And since the latter is a deeper valley (more accurate belief) than the former, the patient will remain there after the drug trip wears off. This seems to really work; the effect size of MDMA on PTSD is very impressive.

But the authors want to go further than that. They write:

In this study, we take the position that most, if not all, expressions of mental illness can be traced to aberrations in the normal mechanics of hierarchical predictive coding, particularly in the precision weighing of both high-level priors and prediction error. We also propose that, if delivered well (Carhart-Harris et al., 2018c), psychedelic therapy can be helpful for such a broad range of disorders precisely because psychedelics work pharmacologically (5-HT2AR agonism) and neurophysiologically (increased excitability of deep-layer pyramidal neurons) to relax the precision weighting of high-level priors (instantiated by high-level cortex) such that they become more sensitive to context (e.g., via sensitivity to bottom-up information flow intrinsic to the system) and amenable to revision (Carhart-Harris, 2018b).

“Most if not all” psychiatric disorders. This has some precedent: some people are already thinking of depression as a high-level prior on negative perceptions and events, obsessive-compulsive disorder as strong priors on the subject of the obsession, etc. But it’s is a really strong claim, and Friston himself has previously published models of depression and anxiety that don’t obviously seem to mesh with this. I wonder if this is Carhart-Harris’ overenthusiasm for psychedelics running a little ahead of the evidence.

Speaking of Carhart-Harris’ overenthusiasm for psychedelics running a little ahead of the evidence, the paper ends with a weird section comparing the hierarchial structure of the brain to the hierarchical structure of society, and speculating that just as psychedelics cause an “anarchic brain” where the highest-level brain structures fail to “govern” lower-level activity, so they may cause society to dissolve or something:

Two figureheads in psychedelic research and therapy, Stanislav Grof and Roland Griffiths, have highlighted how psychedelics have historically “loosed the Dionysian element” (Pollan, 2018) to the discomfort of the ruling elite, i.e., not just in 1960s America but also centuries earlier when conquistadors suppressed the use of psychedelic plants by indigenous people of the same continent. Former Harvard psychology professor, turned psychedelic evangelist, Timothy Leary, cajoled that LSD could stand for “Let the State Dissolve” (Pollan, 2018). Whatever the interaction between psychedelic use and political perspective, we hope that psychedelic science will be given the best possible opportunity to positively impact on psychology, psychiatry, and society in the coming decades—so that it may achieve its promise of significantly advancing self-understanding and health care.

Sure, whatever. But this might be a good time to go back and notice some of the slight discordant notes scattered throughout the paper.

In a paragraph on HPPD, F&CH write:

Hallucinogen-persisting perceptual disorder (HPPD) is a Diagnostic and Statistical Manual of Mental Disorders, 5th Edition–listed disorder that relates to enduring visual perceptual abnormalities that persist beyond an acute psychedelic drug experience. Its prevalence appears to be low and its etiology complex, but symptoms can still be distressing for individuals (Halpern et al., 2018). Under the REBUS model, it is natural to speculate that HPPD may occur if/when the collapse of hierarchical message passing does not fully recover. A compromised hierarchy would imply a compromised suppression of prediction error, and it is natural to assume that persistent perceptual abnormalities reflect attempts to explain away irreducible prediction errors. Future brain-imaging work could examine whether aspects of hierarchical message passing, such as top-down effective connectivity, are indeed compromised in individuals reporting HPPD.

In other words, the priors relax and don’t unrelax again after the drug experience.

For example, an especially common HPPD experience is seeing solid objects pulsate, ooze, or sway. It’s not surprising that a noisy visual system would sometimes put the edge of an object in one place rather than another. But usually a strong prior on “solid objects are not pulsating” prevents this from interfering with perception. Relax this prior too far and the pulsating becomes apparent. If the prior stays relaxed after the drug trip ends, you’ll keep seeing the pulsation indefinitely.

This is one of two plausible theories of HPPD, the other being that the hours of seeing objects pulsate makes your brain learn a new prior, “objects do pulsate” and stick to it. This would make more sense in the context of other learned permanent perceptual disorders like mal de debarquement.

F&CH include a section called “What To Do About The Woo?”, where they admit many people who have psychedelic experiences end up believing strange things: ghosts, mysticism, conspiracies. They are not very worried about this, positing that “a strong psychedelic experience can cause such an ontological shock that the experiencer feels compelled to reach for some kind of explanation” and arguing that as long as we remind people that science is good and pseudoscience is bad, they should be fine.

But I still worry that psychedelic woo is the cognitive equivalent of HPPD.

On one reading, it’s the failure of relaxed priors to re-strengthen, so that beliefs that previously had low prior probability – “this phenomenon is explained by ghosts”, “this guy at the subway station preaching universal love has really discovered all the secrets of the universe” – become more compelling. Spiritual beliefs are kind of a Pascal’s Wager type of deal – extremely important if true, but so unlikely to be true that we don’t usually pay much attention to them. If someone is walking around with a permanently flattened energy landscape – if all their probabilities are smushed together so that unlikely things don’t seem that much more unlikely than likely ones – then the calculation goes the other way, and the fascinating nature of these beliefs overcomes their improbability to make them seem worthy of attention.

On the other reading, they’re the result of newly-established priors in favor of ghosts and mysticism and conspiracies. People are not actually very good at reasoning. If you metaphorically heat up their brain to a temperature that dissolves all their preconceptions and forces them to basically reroll all of their beliefs, then a few of them that were previously correct are going to come out wrong. F&CH’s theory that they are merely letting evidence propagate more fluidly through the system runs up against the problem where, most of the time, if you have to use evidence unguided by any common sense, you probably get a lot of things wrong.

F&CH aren’t the first people to discuss this theory of psychedelics. It’s been in the air for a couple of years now – and props to local bloggers at the Qualia Research Institute and Mad.Science.Blog for getting good explanations up before the parts had even all come together in journal articles. I’m especially interested in QRI’s theory that meditation has the same kind of annealing effect, which I think would explain a lot.

But F&CH’s paper lends the theory a new level of credibility. Carhart-Harris is one of the pioneers of psychedelic therapy, and the paper looks like it’s intended to get people more interested in and accepting of that work by providing a promising theoretical basis. If so, mission accomplished.

[Partial Retraction] Age Gaps and Birth Order Effects

On Less Wrong, Bucky tries to replicate my results on birth order and age gaps.

Backing up: two years ago, I looked at SSC survey data and found that firstborn children were very overrepresented. That result was replicated a few times, both in the SSC sample and in other samples of high-opennness STEM types. Last year, I expanded those results to look at how age gaps affected birth order effects. Curiously, age gaps less than seven years did not seem to attenuate birth order, but age gaps of more than seven years attenuated it almost completely.

Bucky analyzed the same data and found that I bungled one and a half of my results. Left graph in each pair is mine, right is Bucky’s.

In the first analysis, Bucky replicates my results: in people with exactly one sibling show a sudden cliff in birth order effects after seven years.

In the second and third analysis, Bucky finds that I screwed up. I mislabeled the second analysis (people with more than one sibling) as the third (full sample), and my third analysis was just wrong (I double-counted people with one sibling).

When Bucky corrects these errors, they find that the weird cliff at seven years is present only in the sample of people with exactly one sibling. This makes it more likely to be a weird coincidence about that sample, and less likely to be a weird phenomenon. They also identify a potential confounder (there may be longer gaps between later-born children than between first-borns and second-borns) which also slightly affects the results, although does not dramatically change the conclusions.

Bucky then does their own analysis of the correct results, and finds that most likely the sudden drop at seven years is a coincidence. They conclude that:

The SSC 2019 survey data support a constant, high, birth order effect (~2.4 oldest siblings for every 1 second oldest sibling) for age gaps less than 4-8 years. This is followed by a decline to a lower birth order effect at an undetermined rate. The decline does not necessarily completely remove any birth order effect although this may be the case for very large age gaps.

The data provide some evidence that:

– The reduction may not be the same (or might disappear) for larger families (4+ children)
– Birth order effect may be lower at 1 year age gap vs 2-7 year age gap

However the evidence for both of these points is relatively slim.

As those of you who have put up with my constant typos realize, I am not a very careful person. I try to double-check any result I present on my blog, and in some cases (including this one) ask other people to double-check them for me as well. Sometimes mistakes still slip through and I’m sorry. I am partially retracting the previous results (“partially” because some of the analyses were correct and the conclusion is still basically the same).

In response to this, I have added a note in bold to the top of the original age gaps post directing readers to the failed replication and reanalysis. I’ve also added a paragraph about this to my Mistakes page to help people calibrate how much to believe my future results. I am also writing this post to make sure the replication gets at least as much prominence as the original results.

I continue to make the raw survey data available for everybody to double-check my work, except for parts that could seriously compromise people’s privacy. Please, if you have any doubt in my findings at all, do your own analysis and let me know what you get.

Open Thread 136

This is the bi-weekly visible open thread (there are also hidden open threads twice a week you can reach through the Open Thread tab on the top of the page). Post about anything you want, but please try to avoid hot-button political and social topics. You can also talk at the SSC subreddit or the SSC Discord server – and also check out the SSC Podcast. Also:

1. Thanks to everyone who’s helped organize upcoming meetups. I still need people to volunteer for the following cities: Cambridge (UK), Detroit, Dublin, Munich, Oxford, Pittsburgh. If you live in those cities and are willing to host an SSC meetup, please post below and/or email me with location, date, and time. I’ll try to have the big list of times and locations up later this week.

2. Related: the Less Wrong team now has a feature where you can add your location onto a world map and see if there are other people in your area interested in meeting up (or get notifications if someone else organizes for your area). See the thing on the top of

3. I’ve previously been refraining from enforcing the comment policies too hard on people who otherwise produce good content. And when conversations degenerate and everyone breaks the comment policies in a way where it’s hard to disentangle who started it, I’ve been leaving most of the people involved alone. But I think discussion quality has been degenerating here lately, so I’m revoking both those policies. The following people are now banned for multiple violations of the comment policy (linked after their names):
– Conrad Honcho indefinitely (1, 2, 3, 4, 5, 6)
– Dick indefinitely (1, 2, 3, 4)
– Matt M for three months (1, 2)
– Deiseach for three months or until you guys whine at me to reinstate her enough (1, 2, 3)

The following people are on thin ice and should consider themselves warned:
– Brad
– Le Maistre Chat
– JPNunez
– EchoChaos

4. AI safety organization Ought is looking for an engineering team lead (and, uh, offering a big referral bonus, so if you apply, mention my name).

5. I got a chance to talk to the author of the Times article I reviewed in Don’t Fear The Simulators. He wants to clarify that the presentation in the Times was necessarily condensed and simplified, and that if you’re really interested in this topic he has a paper, The Termination Risks Of Simulation Science, which explains his arguments in more detail.

Posted in Uncategorized | Tagged | 1,245 Comments

LW Party: Bay Area Invitation

Less Wrong is throwing a party this Saturday and wants to invite any Bay Area SSC readers who are interested. Consider this an experiment to see if inviting people to parties via blog is a good idea.

It will be at 2412 MLK Jr Way in Berkeley (a private house), from 7 PM on. The Facebook page is here.

I’m going to be there; so will the Less Wrong website team and other interesting people.

Posted in Uncategorized | Tagged | 15 Comments

List Of Passages I Highlighted In My Copy Of Ages Of Discord

Turchin has some great stories about unity vs. polarization over time. For example in the 1940s, unity became such a “problem” that concerned citizens demanded more partisanship:

Concerned about electoral torpor and meaningless political debate, the American Political Science Association in 1946 appointed a committee to examine the role of parties in the American system. Four years later, the committee published a lengthy (and alarmed) report calling for the return of ideologically distinct and powerful political parties. Parties ought to stand for distinct sets of politics, the political scientists urged. Voters should be presented with clear choices.

I have vague memories of similar demands in the early ’90s; everyone was complaining that the parties were exactly the same and the “elites” were rigging things to make sure we didn’t have any real choices.

On the other hand, partisanship during the Civil War was pretty intense:

Another indicator of growing intraelite conflict was the increasing incidence of violence and threatened violence in Congress, which reached a peak during the 1850s. The brutal caning that Representative Preston Brooks of South Carolina gave to Senator Charles Sumner of Massachusetts on the Senate floor in 1856 is the best known such episode, but it was not the only one. In 1842, after Representative Thomas Arnold of Tennessee “reprimanded a pro-slavery member of his own party, two Southern Democrats stalked towards him, at least of one of whom was arhmed with a bowie knife…calling Arnold a ‘damned coward,’ his angry colleagues threatened to cut his throat ‘from ear to ear'” (Freeman 2011). According to Senator Hammond, “The only persons who do not have a revolver and a knife are those who have two revolvers” (quoted in Potter 1976:389). During a debate in 1850, Senator Henry Foote of Mississippi pulled a pistol on Senator Thomas Hart Benton of Missouri (Freeman 2011).

In another bitter debate, a New York congressman inadvertently dropped a pistol (it fell out of his pocket), and this almost precipitated a general shootout on the floor of Congress (Potter 1976: 389).

Turchin places the peak of US unity and cooperation around 1820, and partly credits the need to stand together against Indians:

A particularly interesting case is eighteenth-century Pennyslvania (the following discussion follows closely the text in Turchin 2011:30-31). Initially, European settlers were divided by a number of ethnic and religious boundaries (Silver 2008). The English found it difficult to cooperate with the Germans and the Irish, and each ethnic group was further divided into feuding sectarian groups: Quakers against Anglicans, German Lutherans against Moravians and Mennonites. Yet, by the end of the eighteenth century, the European settlers had forged a common identity (“white people”) in opposition to the natives. As Nancy Shoemaker (2004) shoes, these “metaethnic” labels (Whites versus Reds) were not evoked as soon as settlers and natives came into contact. Rather, during the course of the eighteenth century Europeans and Indians gradually abandoned an initial willingness to recognize in each other a common humanity. Instead, both sides developed new stereotypes of the Other, rooted in the conviction that they were peoples fundamentally at odds, by custom and even by nature (Shoemaker 2004).

The evolution of civic organizations reflected this expanding definition of common identity. Clubs with ethnic and denominational membership criteria appeared in Pennyslvania during the 1740s. These associations represented what Putnam (2000) called “bonding” rather than “bridging” social capital. For example, the St. Andrew’s Society was narrowly focused on helping the Scots, while Deutsche Gesellschaft did the same for the Germans. However, as settler-native warfare intensified, especially during the second half of the eighteenth century, the focus of civic organizations gradually shifted to charity for any victims of Indian attacks, without regard for their ethnicity or religious denomination (Silver 2008). The social scale of coorperation took a step up. Of course, there were definite limits to this new “bridging” social capital: the Indians were most emphatically excluded; in fact, the integration of “white people” developed explicitly in opposition to the Indians.

Although the above description applies to pre-revolutionary Pennsylvania, a very similar dynamic obtained on the Northwestern frontier in Ohio after the Revolution (Griffin 2007). As Griffin notes, for white Americans “Indians existed as cultural glue, since the hatred of them was fast becoming a basis of order.”

This passage stood out to me because modern racial commentators focus on “whiteness” as an idea that evolved in opposition to (and to justify oppression of) blacks. But the Indian theory makes some sense too, especially because Northerners would have more exposure to Indians than they did to black people. But I notice I’ve never heard anyone else talk about this, and most of the history books I’ve read treat Indians as too weak to be an important enemy or have much of a place in the early American consciousness.

One factor leading to greater polarization was “elite overproduction”, here represented by more office-seekers than federal offices. This was apparently a well-known problem in early America:

Despite the increase in government posts, the supply was overwhelmed by demand for such positions. A horde of office-seekers nearly turned Jackson’s inauguration into a riot. Abraham Lincoln once said, “Were it believed that vacant places could be had at the North Pole, the road there would be lined with dead Virginians” (quoted in Potter 1976:432). And, most dramatically (although in a later period), President James Garfield was assassinated by a rejected office-seeker in 1881.

And so on. Some of Turchin’s measures of cooperation vs. polarization are a bit odd. But I have to respect the big-picture-ness of someone who will literally just look at the occurence of the word “cooperation” in various books:

It is interesting that “culture-metric” data support Fraser’s subjective perception [of declining cooperation between business and labor]. For example, the frequency of the word “cooperation” in the corpus of American books grew rapidly during the Progressive Era and somewhat less so during the New Deal (Figure 12.3). After reaching a peak in 1940, there was a minor decline during the 1950s, followed by an increase toward the second peak of 1975. After 1975, however, the frequency of this word went into a sustained decline.

Google Ngram is an imperfect instrument with which to trace cultural shifts. One problem is that the same word (eg, “capitalism”) can be used with either positive or negative valence, and Ngram does not allow one to separate these different meanings. “Cooperation”, however, is rarely used in the negative sense. Because of its predominantly positive valence, its overall frequency should provide us with a proxy for how much a society values cooperative values. Checking different variants (cooperation, Cooperation, cooperative, etc) yields the same overall rise-fall dynamics during the twentieth century (and up to 2008, where the current Google book database stops).

Furthermore, a more specific phrase, “labor-business cooperation” again traces out the same secular cycle, although with significant differences during some decades (eg, the 1920s). Finally, “corporate greed” with its predominantly negative valence is another check on the validity of this result, and it is reassuring that during the twentieth century its frequency moved in the opposite direction from the two positive terms (to show this parallelism more clearly, Figure 12.3 plots “corporate greed” on an inverse scale).


There is an interesting parallel…between the Great Depression and the 1970s Bear Market. Both periods of economic hardship (although it goes without saying that the Great Depression was a much more severe crisis) were broadly interpreted as empirical evidence against the prevailing economic doctrine – the naked, laissez faire capitalism in the first instance, more cooperative relations between business and labor in the second. Yet it is much more likely that the primary mechanism, responsible for long-term economic decline/stagnation in each case, was the negative phase of the Kondratiev cycle, perhaps supplemented by exogenous shocks (eg, the 1973 oil embargo). Yet in each case a prolonged period of economic troubles helped to delegitimize the prevailing ideological regime (Chapter 9).

Thanks for reminding me there’s yet another cycle I need to study, one that supposedly determines the rate of technological advances. Maybe that’s my next book review.

Book Review: Ages Of Discord


I recently reviewed Secular Cycles, which presents a demographic-structural theory of the growth and decline of pre-industrial civilizations. When land is plentiful, population grows and the economy prospers. When land reaches its carrying capacity and income declines to subsistence, the area is at risk of famines, diseases, and wars – which kill enough people that land becomes plentiful again. During good times, elites prosper and act in unity; during bad times, elites turn on each other in an age of backstabbing and civil strife. It seemed pretty reasonable, and authors Peter Turchin and Sergey Nefedov had lots of data to support it.

Ages of Discord is Turchin’s attempt to apply the same theory to modern America. There are many reasons to think this shouldn’t work, and the book does a bad job addressing them. So I want to start by presenting Turchin’s data showing such cycles exist, so we can at least see why the hypothesis might be tempting. Once we’ve seen the data, we can decide how turned off we want to be by the theoretical problems.

The first of Turchin’s two cyclic patterns is a long cycle of national growth and decline. In Secular Cycles‘ pre-industrial societies, this pattern lasted about 300 years; in Ages of Discord‘s picture of the modern US, it lasts about 150:

This summary figure combines many more specific datasets. For example, archaeologists frequently assess the prosperity of a period by the heights of its skeletons. Well-nourished, happy children tend to grow taller; a layer with tall skeletons probably represents good times during the relevant archaeological period; one with stunted skeletons probably represents famine and stress. What if we applied this to the modern US?

Average US height and life expectancy over time. As far as I can tell, the height graph is raw data. The life expectancy graph is the raw data minus an assumed constant positive trend – that is, given that technological advance is increasing life expectancy at a linear rate, what are the other factors you see when you subtract that out? The exact statistical logic be buried in Turchin’s source (Historical Statistics of the United States, Carter et al 2004), which I don’t have and can’t judge.

This next graph is the median wage divided by GDP per capita, a crude measure of income equality:

Lower values represent more inequality.

This next graph is median female age at first marriage. Turchin draws on research suggesting this tracks social optimism. In good times, young people can easily become independent and start supporting a family; in bad times, they will want to wait to make sure their lives are stable before settling down:

This next graph is Yale tuition as a multiple of average manufacturing worker income. To some degree this will track inequality in general, but Turchin thinks it also measures something like “difficulty of upward mobility”:

This next graph shows DW-NOMINATE’s “Political Polarization Index”, a complicated metric occasionally used by historians of politics. It measures the difference in voting patterns between the average Democrat in Congress and the average Republican in Congress (or for periods before the Democrats and Republicans, whichever two major parties there were). During times of low partisanship, congressional votes will be dominated by local or individual factors; during times of high partisanship, it will be dominated by party identification:

I’ve included only those graphs which cover the entire 1780 – present period; the book includes many others that only cover shorter intervals (mostly the more recent periods when we have better data). All of them, including the shorter ones not included here, reflect the same general pattern. You can see it most easily if you standardize all the indicators to the same scale, match the signs so that up always means good and down always means bad, and put them all together:

Note that these aren’t exactly the same indicators I featured above; we’ll discuss immigration later.

The “average” line on this graph is the one that went into making the summary graphic above. Turchin believes that after the American Revolution, there was a period of instability lasting a few decades (eg Shays’ Rebellion, Whiskey Rebellion) but that America reached a maximum of unity, prosperity, and equality around 1820. Things gradually got worse from there, culminating in a peak of inequality, misery, and division around 1900. The reforms of the Progressive Era gradually made things better, with another unity/prosperity/equality maximum around 1960. Since then, an increasing confluence of negative factors (named here as the Reagan Era trend reversal, but Turchin admits it began before Reagan) has been making things worse again.


Along with this “grand cycle” of 150 years, Turchin adds a shorter instability cycle of 40-60 years. This is the same 40-60 year instability cycle that appeared in Secular Cycles, where Turchin called it “the bigenerational cycle”, or the “fathers and sons cycle”.

Timing and intensity of internal war in medieval and early modern England, from Turchin and Nefedov 2009.

The derivation of this cycle, explained on pages 45 – 58 of Ages of Discord, is one of the highlights of the book. Turchin draws on the kind of models epidemiologists use to track pandemics, thinking of violence as an infection and radicals as plague-bearers. You start with an unexposed vulnerable population. Some radical – patient zero – starts calling for violence. His ideas spread to a certain percent of people he interacts with, gradually “infecting” more and more people with the “radical ideas” virus. But after enough time radicalized, some people “recover” – they become exhausted with or disillusioned by conflict, and become pro-cooperation “active moderates” who are impossible to reinfect (in the epidemic model, they are “inoculated”, but they also have an ability without a clear epidemiological equivalent to dampen radicalism in people around them). As the rates of radicals, active moderates, and unexposed dynamically vary, you get a cyclic pattern. First everyone is unexposed. Then radicalism gradually spreads. Then active moderation gradually spreads, until it reaches a tipping point where it triumphs and radicalism is suppressed to a few isolated reservoirs in the population. Then the active moderates gradually die off, new unexposed people are gradually born, and the cycle starts again. Fiddling with all these various parameters, Turchin is able to get the system to produce 40-60 year waves of instability.

To check this empirically, Turchin tries to measure the number of “instability events” in the US over various periods. He very correctly tries to use lists made by others (since they are harder to bias), but when people haven’t catalogued exactly the kind of instability he’s interested in over the entire 1780 – present period, he sometimes adds his own interpretation. He ends up summing riots, lynchings, terrorism (including assassinations), and mass shootings – you can see his definition for each of these starting on page 114; the short version is that all the definitions seem reasonable but inevitably include a lot of degrees of freedom.

When he adds all this together, here’s what happens:

Political instability / violent events show three peaks, around 1870, 1920, and 1970.

The 1870 peak includes the Civil War, various Civil War associated violence (eg draft riots), and the violence around Reconstruction (including the rise of the Ku Klux Klan and related violence to try to control newly emancipated blacks).

The 1920 peak includes the height of the early US labor movement. Turchin discusses the Mine War, an “undeclared war” from 1920-1921 between bosses and laborers in Appalachian coal country:

Although it started as a labor dispute, it eventually turned into the largest armed insurrection in US history, other than the Civil War. Between 10,000 and 15,000 miners armed with rifles fought thouasnds of strike-breakers and sheriff’s deputies, called the Logan Defenders. The insurrection was ended by the US Army. While such violent incidents were exceptional, they took place against a background of a general “class war” that had been intensifying since the violent teens. “In 1919 nearly four million workers (21% of the workforce) took disruptive action in the face of employer reluctance to recognize or bargain with unions” (Domhoff and Webber, 2011:74).

Along with labor violence, 1920 was also a peak in racial violence:

Race-motivated riots also peaked around 1920. The two most serious such outbreaks were the Red Summer of 1919 (McWhirter 2011) and the Tulsa (Oklahoma) Race Riot. The Red Summer involved riots in more than 20 cities across the United States and resulted in something like 1,000 fatalities. The Tulsa riot in 1921, which caused about 300 deaths, took on an aspect of civil war, in which thousands of whites and blacks, armed with firearms, fought in the streets, and most of the Greenwood District, a prosperous black neighborhood, was destroyed.

And terrorism:

The bombing campaign by Italian anarchists (“Galleanists”) culminated in the 1920 explosion on Wall Street, which caused 38 fatalities.

The same problems: labor unrest, racial violence, terrorism – repeated during the 1970s spike. Instead of quoting Turchin on this, I want to quote this Status 451 review of Days of Rage, because it blew my mind:

“People have completely forgotten that in 1972 we had over nineteen hundred domestic bombings in the United States.” — Max Noel, FBI (ret.)

Recently, I had my head torn off by a book: Bryan Burrough’s Days of Rage, about the 1970s underground. It’s the most important book I’ve read in a year. So I did a series of running tweetstorms about it, and Clark asked me if he could collect them for posterity. I’ve edited them slightly for editorial coherence.

Days of Rage is important, because this stuff is forgotten and it shouldn’t be. The 1970s underground wasn’t small. It was hundreds of people becoming urban guerrillas. Bombing buildings: the Pentagon, the Capitol, courthouses, restaurants, corporations. Robbing banks. Assassinating police. People really thought that revolution was imminent, and thought violence would bring it about.

One thing that Burrough returns to in Days of Rage, over and over and over, is how forgotten so much of this stuff is. Puerto Rican separatists bombed NYC like 300 times, killed people, shot up Congress, tried to kill POTUS (Truman). Nobody remembers it.

The passage speaks to me because – yeah, nobody remembers it. This is also how I feel about the 1920 spike in violence. I’d heard about the Tulsa race riot, but the Mine War and the bombing of Wall Street and all the other stuff was new to me. This matters because my intuitions before reading this book would not have been that there were three giant spikes in violence/instability in US history located fifty years apart. I think the lesson I learn is not to trust my intuitions, and to be a little more sympathetic to Turchin’s data.

One more thing: the 1770 spike was obviously the American Revolution and all of the riots and communal violence associated with it (eg against Tories). Where was the 1820 spike? Turchin admits it didn’t happen. He says that because 1820 was the absolute best part of the 150 year grand cycle, everybody was so happy and well-off and patriotic that the scheduled instability peak just fizzled out. Although Turchin doesn’t mention it, you could make a similar argument that the 1870 spike was especially bad (see: the entire frickin’ Civil War) because it hit close to (though not exactly at) the worst part of the grand cycle. 1920 hit around the middle, and 1970 during a somewhat-good period, so they fell in between the nonissue of 1820 and the disaster of 1870.


I haven’t forgotten the original question – what drives these 150 year cycles of rise and decline – but I want to stay with the data just a little longer. Again, these data are really interesting. Either some sort of really interesting theory has to be behind them – or they’re just low-quality data cherry-picked to make a point. Which are they? Here are a couple of spot-checks to see if the data are any good.

First spot check: can I confirm Turchin’s data from independent sources?

Here is a graph of average US height over time which seems broadly similar to Turchin’s.

Here is a different measure of US income inequality over time, which again seems broadly similar to Turchin’s. Piketty also presents very similar data, though his story places more emphasis on the World Wars and less on the labor movement.

– The Columbia Law Review measures political polarization over time and gets mostly the same numbers as Turchin.

I’m going to consider this successfully checked; Turchin’s data all seem basically accurate.

Second spot check: do other indicators Turchin didn’t include confirm the pattern he detects, or did he just cherry-pick the data series that worked? Spoiler: I wasn’t able to do this one. It was too hard to think of measures that should reflect general well-being and that we have 200+ years of unconfounded data for. But here are my various failures:

– The annual improvement in mortality rate does not seem to follow the cyclic pattern. But isn’t this more driven by a few random factors like smoking rates and the logic of technological advance?

Treasury bonds maybe kind of follow the pattern until 1980, after which they go crazy.

Divorce rates look kind of iffy, but isn’t that just a bunch of random factors?

Homicide rates, with the general downward trend removed, sort of follow the pattern, except for the recent decline?

USD/GBP exchange rates don’t show the pattern at all, but that could be because of things going on in Britain?

The thing is – really I have no reason to expect divorce rates, homicide rates, exchange rates etc to track national flourishing. For one thing, they may just be totally unrelated. For another, even if they were tenuously related, there are all sorts of other random factors that can affect them. The problem is, I would have said this was true for height, age at first marriage, and income inequality too, before Turchin gave me convincing-sounding stories for why it wasn’t. I think my lesson is that I have no idea which indicators should vs. shouldn’t follow a secular-cyclic pattern and so I can’t do this spot check against cherry-picking the way I hoped.

Third spot check: common sense. Here are some things that stood out to me:

– The Civil War is at a low-ish part of the cycle, but by no means the lowest.

– The Great Depression happened at a medium part of the cycle, when things should have been quickly getting better.

– Even though there was a lot of new optimism with Reagan, continuing through the Clinton years, the cycle does not reflect this at all.

Maybe we can rescue the first and third problem by combining the 150 year cycle with the shorter 50 year cycle. The Civil War was determined by the 50-year cycle having its occasional burst of violence at the same time the 150-year cycle was at a low-ish point. People have good memories of Reagan because the chaos of the 1970 violence burst had ended.

As for the second, Turchin is aware of the problem. He writes:

There is a widely held belief among economists and other social scientists that the 1930s were the “defining moment” in the development of the American politico-economic system (Bordo et al 1998). When we look at the major structural-demographic variables, however, the decade of the 1930s does not seem to be a turning point. Structural-demographic trends that were established during the Progressive Era continues through the 1930s, although some of them accelerated.

Most notably, all the well-being variables that went through trend reversals before the Great Depression – between 1900 and 1920. From roughly 1910 and to 1960 they all increased roughly monotonically, with only one or two minor fluctuations around the upward trend. The dynamics of real wages also do not exhibit a breaking point in the 1930s, although there was a minor acceleration after 1932.

By comparison, he plays up the conveniently-timed (and hitherto unknown to me) depression of the mid-1890s. Quoting Turchin quoting McCormick:

No depression had ever been as deep and tragic as the one that lasted from 1893 to 1897. Millions suffered unemployment, especially during the winters of 1893-4 and 1894-5, and thousands of ‘tramps’ wandered the countryside in search of food […]

Despite real hardship resulting form massive unemployment, well-being indicators suggest that the human cost of the Great Depression of the 1930s did not match that of the “First Great Depression” of the 1890s (see also Grant 1983:3-11 for a general discussion of the severity of the 1890s depression. Furthermore, while the 1930s are remembered as a period of violent labor unrest, the intensity of class struggle was actually lower than during the 1890s depression. According to the US Political Violence Database (Turchin et al. 2012) there were 32 lethal labor disputes during the 1890s that collectively caused 140 deaths, compared with 20 such disputes in the 1930s with the total of 55 deaths. Furthermore, the last lethal strike in US labor history was in 1937…in other words, the 1930s was actually the last uptick of violent class struggle in the US, superimposed on an overall declining trend.

The 1930s Depression is probably remembered (or rather misremembered) as the worst economic slump in US history, simply because it was the last of the great depressions of the post-Civil War era.

Fourth spot check: Did I randomly notice any egregious errors while reading the book?

On page 70, Turchin discusses “the great cholera epidemic of 1849, which carried away up to 10% of the American population”. This seemed unbelievably high to me. I checked the source he cited, Kohl’s “Encyclopedia Of Plague And Pestilence”, which did give that number. But every other source I checked agreed that the epidemic “only” killed between 0.3% – 1% of the US population (it did hit 10% in a few especially unlucky cities like St. Louis). I cannot fault Turchin’s scholarship in the sense of correctly repeating something written in an encyclopedia, but unless I’m missing something I do fault his common sense.

Also, on page 234, Turchin interprets the percent of medical school graduates who get a residency as “the gap between the demand and supply of MD positions”, which he ties into a wider argument about elite overproduction. But I think this shows a limited understanding of how the medical system works. There is currently a severe undersupply of doctors – try getting an appointment with a specialist who takes insurance in a reasonable amount of time if you don’t believe me. Residencies aren’t limited by organic demand. They’re limited because the government places so many restrictions on them that hospitals don’t sponsor them without government funding, and the government is too stingy to fund more of them. None of this has anything to do with elite overproduction.

These are just two small errors in a long book. But they’re two errors in medicine, the field I know something about. This makes me worry about Gell-Mann Amnesia: if I notice errors in my own field, how many errors must there be in other fields that I just didn’t catch?

My overall conclusion from the spot-checks is that the data as presented are basically accurate, but that everything else is so dependent on litigating which things are vs. aren’t in accordance with the theory that I basically give up.


Okay. We’ve gone through the data supporting the grand cycle. We’ve gone through the data and theory for the 40-60 year instability cycle. We’ve gone through the reasons to trust vs. distrust the data. Time to go back to the question we started with: why should the grand cycle, originally derived from the Malthusian principles that govern pre-industrial societies, hold in the modern US? Food and land are no longer limiting resources; famines, disease, and wars no longer substantially decrease population. Almost every factor that drives the original secular cycle is missing; why even consider the possibility that it might still apply?

I’ve put this off because, even though this is the obvious question Ages of Discord faces from page one, I found it hard to get a single clear answer.

Sometimes, Turchin talks about the supply vs. demand of labor. In times when the supply of labor outpaces demand, wages go down, inequality increases, elites fragment, and the country gets worse, mimicking the “land is at carrying capacity” stage of the Malthusian cycle. In times when demand for labor exceeds supply, wages go up, inequality decreases, elites unite, and the country gets better. The government is controlled by plutocrats, who always want wages to be low. So they implement policies that increase the supply of labor, especially loose immigration laws. But their actions cause inequality to increase and everyone to become miserable. Ordinary people organize resistance: populist movements, socialist cadres, labor unions. The system teeters on the edge of violence, revolution, and total disintegration. Since the elites don’t want those things, they take a step back, realize they’re killing the goose that lays the golden egg, and decide to loosen their grip on the neck of the populace. The government becomes moderately pro-labor and progressive for a while, and tightens immigration laws. The oversupply of labor decreases, wages go up, inequality goes down, and everyone is happy. After everyone has been happy for a while, the populists/socialists/unions lose relevance and drift apart. A new generation of elites who have never felt threatened come to power, and they think to themselves “What if we used our control of the government to squeeze labor harder?” Thus the cycle begins again.

But at other times, Turchin talks more about “elite overproduction”. When there are relatively few elites, they can cooperate for their common good. Bipartisanship is high, everyone is unified behind a system perceived as wise and benevolent, and we get a historical period like the 1820s US golden age that historians call The Era Of Good Feelings. But as the number of elites outstrips the number of high-status positions, competition heats up. Elites realize they can get a leg up in an increasingly difficult rat race by backstabbing against each other and the country. Government and culture enter a defect-defect era of hyperpartisanship, where everyone burns the commons of productive norms and institutions in order to get ahead. Eventually…some process reverses this or something?…and then the cycle starts again.

At still other times, Turchin seems to retreat to a sort of mathematical formalism. He constructs an extremely hokey-looking dynamic feedback model, based on ideas like “assume that the level of discontent among ordinary people equals the urbanization rate x the age structure x the inverse of their wages relative to the elite” or “let us define the fiscal distress index as debt ÷ GDP x the level of distrust in state institutions”. Then he puts these all together into a model that calculates how the the level of discontent affects and is affected by the level of state fiscal distress and a few dozen other variables. On the one hand, this is really cool, and watching it in action gives you the same kind of feeling Seldon must have had inventing psychohistory. On the other, it seems really made-up. Turchin admits that dynamic feedback systems are infamous for going completely haywire if they are even a tiny bit skew to reality, but assures us that he understands the cutting-edge of the field and how to make them not to do that. I don’t know enough to judge whether he’s right or wrong, but my priors are on “extremely, almost unfathomably wrong”. Still, at times he reminds us that the shifts of dynamic feedback systems can be attributed only to the system in its entirety, and that trying to tell stories about or point to specific factors involved in any particular shift is an approximation at best.

All of these three stories run into problems almost immediately.

First, the supply of labor story focuses pretty heavily on immigration. Turchin puts a lot of work into showing that immigration follows the secular cycle patterns; it is highest at the worst part of the cycle, and lowest at the best parts:

In this model, immigration is a tool of the plutocracy. High supply of labor (relative to demand) drives down wages, increases inequality, and lowers workers’ bargaining power. If the labor supply is poorly organized, comes from places that don’t understand the concept of “union”, don’t know their rights, and have racial and linguistic barriers preventing them from cooperating with the rest of the working class, well, even better. Thus, periods when the plutocracy is successfully squeezing the working class are marked by high immigration. Periods when the plutocracy fears the working class and feels compelled to be nice to them are marked by low immigration.

This position makes some sense and is loosely supported by the long-term data above. But isn’t this one of the most-studied topics in the history of economics? Hasn’t it been proven almost beyond doubt that immigrants don’t steal jobs from American workers, and that since they consume products themselves (and thus increase the demand for labor) they don’t affect the supply/demand balance that sets wages?

It appears I might just be totally miscalibrated on this topic. I checked the IGM Economic Experts Panel. Although most of the expert economists surveyed believed immigration was a net good for America, they did say (50% agree to only 9% disagree) that “unless they were compensated by others, many low-skilled American workers would be substantially worse off if a larger number of low-skilled foreign workers were legally allowed to enter the US each year”. I’m having trouble seeing the difference between this statement (which economists seem very convinced is true) and “you should worry about immigrants stealing your job” (which everyone seems very convinced is false). It might be something like – immigration generally makes “the economy better”, but there’s no guarantee that these gains are evently distributed, and so it can be bad for low-skilled workers in particular? I don’t know, this would still represent a pretty big update, but given that I was told all top economists think one thing, and now I have a survey of all top economists saying the other, I guess big updates are unavoidable. Interested in hearing from someone who knows more about this.

Even if it’s true that immigration can hurt low-skilled workers, Turchin’s position – which is that increased immigration is responsible for a very large portion of post-1973 wage stagnation and the recent trend toward rising inequality – sounds shocking to current political sensibilities. But all Turchin has to say is:

An imbalance between labor supply and demand clearly played an important role in driving real wages down after 1978. As Harvard economist George J. Borjas recently wrote, “The best empirical research that tries to examine what has actually happened in the US labor market aligns well with economic theory: An increase in the number of workers leads to lower wages.”

My impression was that Borjas was an increasingly isolated contrarian voice, so once again, I just don’t know what to do here.

Second, the plutocratic oppression story relies pretty heavily on the idea that inequality is a unique bad. This fits the zeitgeist pretty well, but it’s a little confusing. Why should commoners care about their wages relative to elites, as opposed to their absolute wages? Although median-wage-relative-to-GDP has gone down over the past few decades, absolute median wage has gone up – just a little, slowly enough that it’s rightly considered a problem – but it has gone up. Since modern wages are well above 1950s wages, in what sense should modern people feel like they are economically bad off in a way 1950s people didn’t? This isn’t a problem for Turchin’s theory so much as a general mystery, but it’s a general mystery I care about a lot. One answer is that the cost disease is fueled by a Baumol effect pegged to per capital income (see part 3 here), and this is a way that increasing elite wealth can absolutely (not relatively) immiserate the lower classes.

Likewise, what about The Spirit Level Delusion and other resources showing that, across countries, inequality is not particularly correlated with social bads? Does this challenge Turchin’s America-centric findings that everything gets worse along with inequality levels?

Third, the plutocratic oppression story meshes poorly with the elite overproduction story. In elite overproduction, united elites are a sign of good times to come; divided elites means dysfunctional government and potential violence. But as Pseudoerasmus points out, united elites are often united against the commoners, and we should expect inequality to be highest at times when the elites are able to work together to fight for a larger share of the pie. But I think this is the opposite of Turchin’s story, where elites unite only to make concessions, and elite unity equals popular prosperity.

Fourth, everything about the elite overproduction story confuses me. Who are “elites”? This category made sense in Secular Cycles, which discussed agrarian societies with a distinct titled nobility. But Turchin wants to define US elites in terms of wealth, which follows a continuous distribution. And if you’re defining elites by wealth, it doesn’t make sense to talk about “not enough high-status positions for all elites”; if you’re elite (by virtue of your great wealth), by definition you already have what you need to maintain your elite status. Turchin seems aware of this issue, and sometimes talks about “elite aspirants” – some kind of upper class who expect to be wealthy, but might or might not get that aspiration fulfilled. But then understanding elite overproduction hinges on what makes one non-rich-person person a commoner vs. another non-rich-person an “elite aspirant”, and I don’t remember any clear discussion of this in the book.

Fifth, what drives elite overproduction? Why do elites (as a percent of the population) increase during some periods and decrease during others? Why should this be a cycle rather than a random walk?

My guess is that Ages of Discord contains answers to some of these questions and I just missed them. But I missed them after reading the book pretty closely to try to find them, and I didn’t feel like there were any similar holes in Secular Cycles. As a result, although the book had some fascinating data, I felt like it lacked a clear and lucid thesis about exactly what was going on.


Accepting the data as basically right, do we have to try to wring some sense out of the theory?

The data cover a cycle and a half. That means we only sort of barely get to see the cycle “repeat”. The conclusion that it is a cycle and not some disconnected trends is based only on the single coincidence that it was 70ish years from the first turning point (1820) to the second (1890), and also 70ish years from the second to the third (1960).

A parsimonious explanation would be “for some reason things were going unusually well around 1820, unusually badly around 1890, and unusually well around 1960 again.” This is actually really interesting – I didn’t know it was true before reading this book, and it changes my conception of American history a lot. But it’s a lot less interesting than the discovery of a secular cycle.

I think the parsimonious explanation is close to what Thomas Piketty argued in his Capital In The Twenty-First Century. Inequality was rising until the World Wars, because that’s what inequality naturally does given reasonable assumptions about growth rates. Then the Depression and World Wars wiped out a lot of existing money and power structures and made things equal again for a little while. Then inequality started rising again, because that’s what inequality naturally does given reasonable assumptions about growth rates. Add in a pinch of The Spirit Level – inequality is a mysterious magic poison that somehow makes everything else worse – and there’s not much left to be explained.

(some exceptions: why was inequality decreasing until 1820? Does inequality really drive political polarization? When immigration corresponds to periods of high inequality, is the immigration causing the inequality? And what about the 50 year cycle of violence? That’s another coincidence we didn’t include in the coincidence list!)

So what can we get from Ages of Discord that we can’t get from Piketty?

First, the concept of “elite overproduction” is one that worms its way into your head. It’s the sort of thing that was constantly in the background of Increasingly Competitive College Admissions: Much More Than You Wanted To Know. It’s the sort of thing you think about when a million fresh-faced college graduates want to become Journalists and Shape The Conversation and Fight For Justice and realistically just end up getting ground up and spit out by clickbait websites. Ages of Discord didn’t do a great job breaking down its exact dynamics, but I’m grateful for its work bringing it from a sort of shared unconscious assumption into the light where we can talk about it.

Second, the idea of a deep link between various indicators of goodness and badness – like wages and partisan polarization – is an important one. It forces me to reevaluate things I had considered settled, like that immigration doesn’t worsen inequality, or that inequality is not a magical curse that poisons everything.

Third, historians have to choose what events to focus on. Normal historians usually focus on the same normal events. Unusual historians sometimes focus on neglected events that support their unusual theses, so reading someone like Turchin is a good way to learn parts of history you’d never encounter otherwise. Some of these I was able to mention above – like the Mine War of 1920 or the cholera epidemic of 1849; I might make another post for some of the others.

Fourth, it tries to link events most people would consider separate – wage stagnation since 1973, the Great Stagnation in technology, the decline of Peter Thiel’s “definite optimism”, the rise of partisan polarization. I’m not sure exactly how it links them or what it has to stay about the link, but link them it does.

But the most important thing about this book is that Turchin claims to be able to predict the future. The book (written just before Trump was elected in 2016) ends by saying that “we live in times of intensifying structural-demographic pressures for instability”. The next bigenerational burst of violence is scheduled for about 2020 (realistically +/- a few years). It’s at a low point in the grand cycle, so it should be a doozy.

What about beyond that? It’s unclear exactly where he thinks we are right now in the grand cycle. If the current cycle lasts exactly as long as the last one, we would expect it to bottom out in 2030, but Turchin never claims every cycle is exactly as long. A few of his graphs suggest a hint of curvature, suggesting we might currently be in the worst of it. The socialists seem to have gotten their act together and become an important political force, which the theory predicts is a necessary precursor to change.

I think we can count the book as having made correct predictions if violence spikes in the very near future (are the current number of mass shootings enough to satisfy this requirement? I would have to see it graphed using the same measurements as past spikes), and if sometime in the next decade or so things start looking like there’s a ray of light at the end of the tunnel.

I am pretty interested in finding other ways to test Turchin’s theories. I’m going to ask some of my math genius friends to see if the dynamic feedback models check out; if anyone wants to help, let me know how I can help you (if money is an issue, I can send you a copy of the book, and I will definitely publish anything you find on this blog). If anyone has any other ideas for to indicators that should be correlated with the secular cycle, and ideas about how to find them, I’m intereted in that too. And if anyone thinks they can explain the elite overproduction issue, please enlighten me.

I ended my review of Secular Cycles by saying:

One thing that strikes me about [Turchin]’s cycles is the ideological component. They describe how, during a growth phase, everyone is optimistic and patriotic, secure in the knowledge that there is enough for everybody. During the stagflation phase, inequality increases, but concern about inequality increases even more, zero-sum thinking predominates, and social trust craters (both because people are actually defecting, and because it’s in lots of people’s interest to play up the degree to which people are defecting). By the crisis phase, partisanship is much stronger than patriotism and radicals are talking openly about how violence is ethically obligatory.

And then, eventually, things get better. There is a new Augustan Age of virtue and the reestablishment of all good things. This is a really interesting claim. Western philosophy tends to think in terms of trends, not cycles. We see everything going on around us, and we think this is some endless trend towards more partisanship, more inequality, more hatred, and more state dysfunction. But Secular Cycles offers a narrative where endless trends can end, and things can get better after all.

This is still the hope, I guess. I don’t have a lot of faith in human effort to restore niceness, community, and civilization. All I can do is pray the Vast Formless Things accomplish it for us without asking us first.

Meetups Everywhere 2019

Last autumn we organized meetups in 85 different cities (and one ship!) around the world. Some of the meetup groups stuck around or reported permanent spikes in membership, which sounds like a success, so let’s do it again.

For most cities: If you’re willing to host a meetup for your city, then decide on a place, date, and time, and post it in the comments here, along with an email address where people can contact you. Then please watch the comments in case I need to ask you any questions. If you’re not sure whether your city has enough SSC readers to support a meetup, see the list of people by city at the bottom of this post. There may be more of us than you think – last year we were able to support meetups in such great megalopolises as Norman, Oklahoma and Wellington, New Zealand. But I would prefer people not split things up too much – if you’re very close to a bigger city, consider going there instead of hosting your own.

If you want a meetup for your city, please err in favor of volunteering to organize – the difficulty level is basically “pick a coffee shop you like, tell me the address, and give me a time”; it would be dumb if nobody got to go to meetups because everyone felt too awkward and low-status to volunteer.

For especially promising cities in the US: I am going to try to attend your meetups. My very tentative schedule looks like this:

Friday 9/20: Boston
Saturday 9/21: NYC
Sunday 9/22: Philly
Monday 9/23: DC
Thursday 9/26: Ann Arbor
Saturday 9/28: Chicago
Sunday 9/29: Austin
Tuesday 10/1: Portland
Wednesday 10/2: Seattle
Friday 10/4: Fairbanks
Thursday 10/10: Berkeley
Friday 10/11: Orange County

If you are in one of these cities and want to host a meetup, please schedule it for the evening of the relevant day. If that’s impossible, let me know and I might be able to reschedule. I will announce these ones on the blog, and in the past that’s meant they can get very big (100+ people in the biggest cities) – you might want to hold it in a house, park, or classroom (not a cafe or small apartment). If you have a great location but need money, email me and I might be able to help.

Small-print rules for organizers

1. In a week or so, I’ll make another post listing the details for each city so people know where to go.I don’t guarantee I’ll have the post with times and addresses up until September 9, so please choose a day after that. The weekend of September 21st and 22nd might be one good choice.

2. In the past, the best venues have been ones that are quiet(ish) and have lots of mobility for people to arrange themselves into circles or subgroups as desired. Private houses have been pretty good. Same with food courts. Cafes and restaurants have gone okay, as have empty fields (really). Bars don’t seem to have worked very well at all.

3. Usually only about a quarter of people who express interest actually attend. If your city has fewer than 20 people on the big list, don’t offer to organize unless you’re okay with a good chance of only one or two other people showing up.

4. If more than one person volunteers to organize, I’ll pick one of them. Priority will be given to people I know well, people who have organized meetups before, and (especially) an existing SSC/LW/EA meetup group in the city. If you run an existing SSC/LW/EA meetup group and you want to organize your city’s SSC meetup, please mention that in the post so I can give you precedence.

5. If you have an existing meetup group, you can just tell me what you’re already doing and when your next meetup is. But try to have the one you list here be some kind of “welcome, SSC people” meetup or otherwise low-barrier-to-entry. And please give me a firm date and time commitment instead of “tell people to check our mailing list to find out where the meeting will be that week”.

6. If you’re formally volunteering to organize a meetup, please respond with an unambiguous statement to this effect, the exact address, the exact time, and the date (+ contact details if possible), preferably in bold. I’m not going to count someone as offering to organize a meetup unless they do this. Please don’t post “I hope someone agrees to organize a meetup in my city”. Just offer to organize the meetup! Again, please include an exact time, exact date, and exact address with your offer to host. Please don’t post vague speculation about how you might want to host at some point – just offer to host and give me the information I need. If it turns out there’s someone better, don’t worry, they’ll also offer and I’ll choose them.

7. Mingyuan is Director Of Meetups and might be asking you some questions; I vouch for her and you should give her any information she needs.

Thanks, and see (some of) you soon!

Posted in Uncategorized | Tagged , | 235 Comments

Book Review: Reframing Superintelligence

Ten years ago, everyone was talking about superintelligence, the singularity, the robot apocalypse. What happened?

I think the main answer is: the field matured. Why isn’t everyone talking about nuclear security, biodefense, or counterterrorism? Because there are already competent institutions working on those problems, and people who are worried about them don’t feel the need to take their case directly to the public. The past ten years have seen AI goal alignment reach that level of maturity too. There are all sorts of new research labs, think tanks, and companies working on it – the Center For Human-Compatible AI at UC Berkeley, OpenAI, Ought, the Center For The Governance Of AI at Oxford, the Leverhulme Center For The Future Of Intelligence at Cambridge, etc. Like every field, it could still use more funding and talent. But it’s at a point where academic respectability trades off against public awareness at a rate where webzine articles saying CARE ABOUT THIS OR YOU WILL DEFINITELY DIE are less helpful.

One unhappy consequence of this happy state of affairs is that it’s harder to keep up with the field. In 2014, Nick Bostrom wrote Superintelligence: Paths, Dangers, Strategies, giving a readable overview of what everyone was thinking up to that point. Since then, things have been less public-facing, less readable, and more likely to be published in dense papers with a lot of mathematical notation. They’ve also been – no offense to everyone working on this – less revolutionary and less interesting.

This is one reason I was glad to come across Reframing Superintelligence: Comprehensive AI Services As General Intelligence by Eric Drexler, a researcher who works alongside Bostrom at Oxford’s Future of Humanity Institute. This 200 page report is not quite as readable as Superintelligence; its highly-structured outline form belies the fact that all of its claims start sounding the same after a while. But it’s five years more recent, and presents a very different vision of how future AI might look.

Drexler asks: what if future AI looks a lot like current AI, but better?

For example, take Google Translate. A future superintelligent Google Translate would be able to translate texts faster and better than any human translator, capturing subtleties of language beyond what even a native speaker could pick up. It might be able to understand hundreds of languages, handle complicated multilingual puns with ease, do all sorts of amazing things. But in the end, it would just be a translation app. It wouldn’t want to take over the world. It wouldn’t even “want” to become better at translating than it was already. It would just translate stuff really well.

The future could contain a vast ecosystem of these superintelligent services before any superintelligent agents arrive. It could have media services that can write books or generate movies to fit your personal tastes. It could have invention services that can design faster cars, safer rockets, and environmentally friendly power plants. It could have strategy services that can run presidential campaigns, steer Fortune 500 companies, and advise governments. All of them would be far more effective than any human at performing their given task. But you couldn’t ask the presidential-campaign-running service to design a rocket any more than you could ask Photoshop to run a spreadsheet.

In this future, our AI technology would have taken the same path as our physical technology. The human body can run fast, lift weights, and fight off enemies. But the automobile, crane, and gun are three different machines. Evolution had to cram running-ability, lifting-ability, and fighting-ability into the same body, but humans had more options and were able to do better by separating them out. In the same way, evolution had to cram book-writing, technology-inventing, and strategic-planning into the same kind of intelligence – an intelligence that also has associated goals and drives. But humans don’t have to do that, and we probably won’t. We’re not doing it today in 2019, when Google Translate and AlphaGo are two different AIs; there’s no reason to write a single AI that both translates languages and plays Go. And we probably won’t do it in the superintelligent future either. Any assumption that we will is based more on anthropomorphism than on a true understanding of intelligence.

These superintelligent services would be safer than general-purpose superintelligent agents. General-purpose superintelligent agents (from here on: agents) would need a human-like structure of goals and desires to operate independently in the world; Bostrom has explained ways this is likely to go wrong. AI services would just sit around algorithmically mapping inputs to outputs in a specific domain.

Superintelligent services would not self-improve. You could build an AI researching service – or, more likely, several different services to help with several different aspects of AI research – but each of them would just be good at solving certain AI research problems. It would still take human researchers to apply their insights and actually build something new. In theory you might be able to automate every single part of AI research, but it would be a weird idiosyncratic project that wouldn’t be anybody’s first choice.

Most important, superintelligent services could help keep the world safe from less benevolent AIs. Drexler agrees that a self-improving general purpose AI agent is possible, and assumes someone will build one eventually, if only for the lulz. He agrees this could go about the way Bostrom expects it to go, ie very badly. But he hopes that there will be a robust ecosystem of AI services active by then, giving humans superintelligent help in containing rogue AIs. Superintelligent anomaly detectors might be able to notice rogue agents causing trouble, superintelligent strategic planners might be able to develop plans for getting rid of them, and superintelligent military research AIs might be able to create weapons capable of fighting them off.

Drexler therefore does not completely dismiss Bostromian disaster scenarios, but thinks we should concentrate on the relatively mild failure modes of superintelligent AI services. These may involve normal bugs, where the AI has aberrant behaviors that don’t get caught in testing and cause a plane crash or something, but not the unsolveable catastrophes of the Bostromian paradigm. Drexler is more concerned about potential misuse by human actors – either illegal use by criminals and enemy militaries, or antisocial use to create things like an infinitely-addictive super-Facebook. He doesn’t devote a lot of space to these, and it looks like he hopes these can be dealt with through the usual processes, or by prosocial actors with superintelligent services on their side (thirty years from now, maybe people will say “it takes a good guy with an AI to stop a bad guy with an AI”).

This segues nicely into some similar concerns that OpenAI researcher Paul Christiano has brought up. He worries that AI services will be naturally better at satisfying objective criteria than at “making the world better” in some vague sense. Tasks like “maximize clicks to this site” or “maximize profits from this corporation” are objective criteria; tasks like “provide real value to users of this site instead of just clickbait” or “have this corporation act in a socially responsible way” are vague. That means AI may asymmetrically empower some of the worst tedencies in our society without giving a corresponding power increase to normal people just trying to live enjoyable lives. In his model, one of the tasks of AI safety research is to get AIs to be as good at optimizing vague prosocial tasks as they will naturally be at optimizing the bottom line. Drexler doesn’t specifically discuss this in Reframing Superintelligence, but it seems to fit the spirit of the kind of thing he’s concerned about.


I’m not sure how much of the AI alignment community is thinking in a Drexlerian vs. a Bostromian way, or whether that is even a real dichotomy that a knowledgeable person would talk about. I know there are still some people who are very concerned that even programs that seem to be innocent superintelligent services will be able to self-improve, develop misaligned goals, and cause catastrophes. I got to talk to Dr. Drexler a few years ago about some of this (although I hadn’t read the book at the time, didn’t understand the ideas very well, and probably made a fool of myself); at the time, he said that his work was getting a mixed reception. And there are still a few issues that confuse me.

First, many tasks require general intelligence. For example, an AI operating in a domain with few past examples (eg planning defense against a nuclear attack) will not be able to use modern training paradigms. When humans work on these domains, they use something like common sense, which is presumably the sort of thing we have because we understand thousands of different domains from gardening to ballistics and this gives us a basic sense of how the world works in general. Drexler agrees that we will want AIs with domain-general knowledge that cannot be instilled by training, but he argues that this is still “a service”. He agrees these tasks may require AI architectures different from any that currently exist, with relatively complete world-models, multi-domain reasoning abilities, and the ability to learn “on the fly” – but he doesn’t believe those architectures will need to be agents. Is he right?

Second, is it easier to train services or agents? Suppose you want a good multi-domain reasoner that can help you navigate a complex world. One proposal is to create AIs that train themselves to excel in world simulations the same way AlphaGo trained itself to excel in simulated games of Go against itself. This sounds a little like the evolutionary process that created humans, and agent-like drives might be a natural thing to come out of this process. If agents were easier to “evolve” than services, agentic AI might arise at an earlier stage, either because designers don’t see a problem with it or because they don’t realize it is agentic in the relevant sese.

Third, how difficult is it to separate agency from cognition? Natural intelligences use “active sampling” strategies at levels as basic as sensory perception, deciding how to direct attention in order to best achieve their goals. At higher levels, they decide things like which books to read, whose advice to seek out, or what subdomain of the problem to evaluate first. So far AIs have managed to address even very difficult problems without doing this in an agentic way. Can this continue forever? Or will there be some point at which intelligences with this ability outperform those without it.

I think Drexler’s basic insight is that Bostromian agents need to be really different from our current paradigm to do any of the things Bostrom predicts. A paperclip maximizer built on current technology would have to eat gigabytes of training data about various ways people have tried to get paperclips in the past so it can build a model that lets it predict what works. It would build the model on its actually-existing hardware (not an agent that could adapt to much better hardware or change its hardware whenever convenient). The model would have a superintelligent understanding of the principles that had guided some things to succeed or fail in the training data, but wouldn’t be able to go far beyond them into completely new out-of-the-box strategies. It would then output some of those plans to a human, who would look them over and make paperclips 10% more effectively.

The very fact that this is less effective than the Bostromian agent suggests there will be pressure to build the Bostromian agent eventually (Drexler disagrees with this, but I don’t understand why). But this will be a very different project from AI the way it currently exists, and if AI the way it currently exists can be extended all the way to superintelligence, that would give us a way to deal with hostile superintelligences in the future.


All of this seems kind of common sense to me now. This is worrying, because I didn’t think of any of it when I read Superintelligence in 2014.

I asked readers to tell me if there was any past discussion of this. Many people brought up Robin Hanson’s arguments, which match the “ecosystem of many AIs” part of Drexler’s criticisms but don’t focus as much on services vs. agents. Other people brought up discussion under the heading of Tool AI. Combine those two strains of thought, and you more or less have Drexler’s thesis, minus some polish. I read some of these discussions, but I think I failed to really understand them at the time. Maybe I failed to combine them, focused too much on the idea of an Oracle AI, and missed the idea of an ecosystem of services. Or maybe it all just seemed too abstract and arbitrary when I had fewer examples of real AI systems to think about.

I’ve sent this post by a couple of other people, who push back against it. They say they still think Bostrom was right on the merits and superintelligent agents are more likely than superintelligent services. Many brought up Gwern’s essay on why tool AIs are likely to turn into agent AIs and this post by Eliezer Yudkowsky on the same topic – I should probably reread these, reread Drexler’s counterarguments, and get a better understanding. For now I don’t think I have much of a conclusion either way. But I think I made a mistake of creativity in not generating or understanding Drexler’s position earlier, which makes me more concerned about how many other things I might be missing.