TDT for Humans

Link post

This is part 19 of 30 of Ham­mer­time. Click here for the in­tro.

As is Ham­mer­time tra­di­tion, I’m mak­ing a slight change of plans right around the sched­uled time for Plan­ning. My ex­cuse this time:

Sev­eral com­menters pointed out se­ri­ous gaps in my knowl­edge of Fo­cus­ing. I will post­pone In­ter­nal Dou­ble Crux, an ad­vanced form of Fo­cus­ing, to the next cy­cle. In­stead, we will have two more posts on mak­ing and ex­e­cut­ing long-term plans.

Day 19: TDT for Humans

Pre­vi­ously on plan­ning: Day 8, Day 9, Day 10.

To­day I’d like to de­scribe two or­ders of ap­prox­i­ma­tion to a work­ing de­ci­sion the­ory for hu­mans.

TDT 101

Back­ground read­ing: How I Lost 100 Pounds Us­ing TDT.

Choose as though con­trol­ling the log­i­cal out­put of the ab­stract com­pu­ta­tion you im­ple­ment, in­clud­ing the out­put of all other in­stan­ti­a­tions and simu­la­tions of that com­pu­ta­tion.
~ Eliezer

In other words, ev­ery time you make a de­ci­sion, pre-com­mit to mak­ing the same de­ci­sion in all con­cep­tu­ally similar situ­a­tions in the fu­ture.

The strik­ing value of TDT is: make each de­ci­sion as if you would im­me­di­ately reap the long-term re­wards of mak­ing that same de­ci­sion re­peat­edly. And if it turns out you’re an up­date­less agent, this ac­tu­ally works! You ac­tu­ally lose 100 pounds by mak­ing one de­ci­sion.

I en­courage read­ers who have not tried to live by TDT to stop here and try it out for a week.

TDT 201

There are a num­ber of se­ri­ous differ­ences be­tween time­less agents and hu­man be­ings, so ap­ply­ing TDT as stated above re­quires an un­ac­cept­able (to me) level of self-de­cep­tion. My sec­ond or­der of ap­prox­i­ma­tion is to offer a prac­ti­cal and weak ver­sion of TDT based on the Soli­taire Prin­ci­ple and Magic Brain Juice.

Three ob­jec­tions to ap­ply­ing TDT in real life:


A hu­man is about halfway be­tween “one mono­lithic code­base” and “a loose con­fed­er­a­tion of spirits run­ning a ran­dom se­rial dic­ta­tor­ship.” Roughly speak­ing, each spirit is the piece of you built to satisfy one pri­mor­dial need: hunger, friend­ship, cu­ri­os­ity, jus­tice. At any given time, only one or two of these spirits is pre­sent and mak­ing de­ci­sions. As such, even if each in­di­vi­d­ual spirit as up­date­less and de­ter­minis­tic, you don’t get to make de­ci­sions for all the spirits cur­rently in­ac­tive. You don’t have as much con­trol over the other spirits as you would like.

Differ­ent spirits have ac­cess to differ­ent data and be­liefs. I’ve men­tioned, for ex­am­ple, that I have differ­ent per­son­al­ities speak­ing Chi­nese and English. You can ask me what my fa­vorite food is in English, and I’ll say dumplings, but the true an­swer 饺子 feels qual­i­ta­tively bet­ter than dumplings by a wide mar­gin.

Differ­ent spirits have differ­ent val­ues. I have two friends who re­li­ably pro­voke my “sadis­tic dick-mea­sur­ing ass­hole” spirit. If hu­man be­ings re­ally have util­ity func­tions this spirit has nega­tive signs in front of the terms for other peo­ple. It’s un­char­ac­ter­is­ti­cally happy to en­gage in nega­tive-sum games.

It’s al­most im­pos­si­ble to pre­dict when spirits will man­i­fest. Re­cently, I was on a 13-hour flight back from China. I started marathon­ing Game of Thrones af­ter ex­haust­ing the com­edy sec­tion, and a full sea­son of Cer­sei Lan­nister left me in “sadis­tic ass­hole” mode for a full day af­ter­wards. If Hainan Air­lines had stocked more com­edy movies this might not have oc­curred.

Spirits can lay dor­mant for months or years. Meet­ing up with high school friends this De­cem­ber, I fell into old roles and re­ceived effortless ac­cess to a large swathe of faded mem­o­ries.

Con­cep­tual Gerrymandering

Back­ground read­ing: con­cep­tual ger­ry­man­der­ing.

I can make a prob­lem look ei­ther big or small by draw­ing ei­ther a big or small con­cep­tual bound­ary around it, then iden­ti­fy­ing my prob­lem with the con­cep­tual bound­ary I’ve drawn.

TDT runs on an am­bigu­ous “con­cep­tual similar­ity” clause: you pre-com­mit to mak­ing the same de­ci­sion in con­cep­tu­ally similar situ­a­tions. Un­for­tu­nately, you will be prone to mo­ti­vated rea­son­ing and con­cep­tual ger­ry­man­der­ing to get out of time­less pre-com­mit­ments made in the past.

This prob­lem can be re­duced but not solved by clearly stat­ing bound­aries. Life is too high-di­men­sional to even figure out what vari­ables to care about, let alone where to draw the line for each of them. What in­for­ma­tion be­comes salient is a func­tion of your at­ten­tion and notic­ing skills as much as of re­al­ity it­self. Th­ese days, it’s al­most a rou­tine ex­pe­rience to read an ar­ti­cle that suffi­ciently al­ters my ca­pac­i­ties for at­ten­tion as to ren­der situ­a­tions I would pre­vi­ously have con­sid­ered “con­cep­tu­ally similar” al­to­gether dis­tinct.

Magic Brain Juice

Back­ground read­ing: Magic Brain Juice.

Every ac­tion you take is ac­com­panied by an un­in­ten­tional self-mod­ifi­ca­tion.

The hu­man brain is finicky code that self-mod­ifies ev­ery time it takes an ac­tion. The situ­a­tion is even worse than this: your ac­tions can shift your very val­ues in sur­pris­ing and illeg­ible ways. This bug is an in­her­ent con­tra­dic­tion to ap­ply­ing TDT as a hu­man.

Self-mod­ifi­ca­tion hap­pens in mul­ti­ple ways. When I wrote Magic Brain Juice, I was refer­ring to the im­me­di­ate strength­en­ing of neu­ral path­ways that are ac­ti­vated, and the cor­re­spond­ing de­cay through time of all path­ways not ac­ti­vated. But other things hap­pen un­der the hood. You get at­tached to a cer­tain iden­tity. You get sucked into the near­est at­trac­tor in the so­cial web. And also:

Ex­po­sure ther­apy is a pow­er­ful and in­dis­crim­i­nate tool. You can re­duce any aver­sion to al­most zero just by vol­un­tar­ily con­fronting it re­peat­edly. But you have fears and aver­sions in ev­ery di­rec­tion!

Every move you make is ex­po­sure ther­apy in that di­rec­tion.

That’s right.

Every vol­un­tary de­ci­sion nudges your com­fort zone in that di­rec­tion, squash­ing aver­sions (en­dorsed or oth­er­wise) in its path.



I hope I’ve con­vinced you that the hu­man brain is suffi­ciently bro­ken that our in­tu­ition about “up­date­less source code” don’t ap­ply, and try­ing to make de­ci­sions from TDT will be harder (and may have se­ri­ous un­in­tended side effects) as a re­sult. What can be done?

First, I think it’s worth di­rectly in­vest­ing in TDT-like be­hav­iors. Make con­scious de­ci­sions to re­in­force the spirits that are amenable to mak­ing and keep­ing pre-com­mit­ments. Make more leg­ible de­ci­sions and clearly state con­cep­tual bound­aries. Ex­plore virtue ethics or de­on­tol­ogy. Zvi’s blog is a good a place.

In the same vein, prac­tice pre­dict­ing your fu­ture be­hav­ior. If you can be­come your own Omega, prob­lems you face start look­ing New­comb-like. Then you’ll be forced to give up CDT and the failures it en­tail.

Se­cond, I once pro­posed a model called the “Ten Per­cent Shift”:

The Ten Per­cent Shift is a thought ex­per­i­ment I’ve suc­cess­fully pushed to Sys­tem 1 that helps build long-term habits like blog­ging ev­ery day. It makes the as­sump­tion that each time you make a choice, it gets 10% eas­ier.
Sup­pose there is a habit you want to build such as go­ing to the gym. You’ve drawn the pen­ta­grams, sprin­kled the pixie dust, and done the proper rit­u­als to de­cide that the benefits clearly out­weigh the costs and there’s no su­pe­rior al­ter­na­tives. Nev­er­the­less, the effort to make your­self go ev­ery day seems in­sur­mountable.
You spend 100 units of willpower drag­ging your­self there on Day 1. Now, no­tice that you have magic brain juice on your side. On Day 2, it gets a lit­tle bit eas­ier. You spend 90 units. On Day 3, it only costs 80.
A bit of math and a lot of magic brain juice later, you spend 500 units of willpower in the first 10 days, and the habit is free for the rest of time.

The ex­act num­ber is ir­rele­vant, but I stand by this model as the proper weak­en­ing of TDT: act as if each sin­gle de­ci­sion re­wards you with 10% of the value of mak­ing that same de­ci­sion in­definitely. One de­ci­sion only loses you 10 pounds, and you need to make 10 con­sec­u­tive de­ci­sions be­fore you get to reap the full re­wards.

The Ten Per­cent Shift guards against spirits. Once you make the same de­ci­sion 10 times in a row, you’ll have made it from a wide range of states of mind, and the ex­act con­text will have differed in ev­ery situ­a­tion. You’ll prob­a­bly have to con­vince a ma­jor­ity of spirits to agree with mak­ing the de­ci­sion.

The Ten Per­cent Shift also guards against con­cep­tual ger­ry­man­der­ing. Hav­ing made the same de­ci­sion from a bunch of differ­ent situ­a­tions, the con­vex hull of these data points is a 10-di­men­sional con­vex re­gion that you can un­am­bigu­ously stake out as a time­less pre-com­mit­ment.

Daily Challenge

This post is ex­tremely ten­ta­tive and the­o­ret­i­cal, so I’ll just open up the floor for dis­cus­sion.

No nominations.
No reviews.