Musings on Cumulative Cultural Evolution and AI

This post might be in­ter­est­ing to you if you want a con­cep­tual model of cu­mu­la­tive cul­tural evolu­tion and/​or you’re cu­ri­ous how cu­mu­la­tive cul­tural evolu­tion im­pacts AI fore­cast­ing and de­vel­op­ment.

In par­tic­u­lar, I’ll ar­gue that cu­mu­la­tive cul­tural evolu­tion makes one ar­gu­ment for the dis­con­ti­nu­ity of AI progress more plau­si­ble and sketches out at least two pos­si­ble paths of de­vel­op­ment that are worth fur­ther in­ves­ti­ga­tion.

Cu­mu­la­tive Cul­tural Evolution

Hu­mans have al­tered more than one-third of the earths’ land sur­face. We cy­cle more ni­tro­gen than all other ter­res­trial life forms com­bined and have now al­tered the flow of two-thirds of the earth’s rivers. Our species uses 100 times more bio­mass than any large species that has ever lived. If you in­clude our vast herds of do­mes­ti­cated an­i­mals, we ac­count for more than 98% of ter­res­trial ver­te­brate bio­mass. — Joseph Henrich

Cu­mu­la­tive cul­tural evolu­tion is of­ten framed as an an­swer to the gen­eral ques­tion: what makes hu­mans so suc­cess­ful rel­a­tive to other apes? Other apes do not oc­cupy as many con­ti­nents, are not as pop­u­lous, and do not shape or use en­vi­ron­ment to the ex­tent that hu­mans do.

There are a smor­gas­bord of differ­ent ac­counts of this, refer­enc­ing:

  • hu­man ca­pa­bil­ity for de­cep­tion ratch­eted up hu­man so­cial­ity and in­tel­li­gence via arms race like effects.

  • run­away sig­nal­ing.

  • hu­man’s gen­eral in­tel­li­gence or im­pro­vi­sa­tional intelligence

  • hu­mans’ prosociality

  • a pack­age of in­stinc­tual mod­ules, such as lan­guage, in­tel­li­gence, so­cial learn­ing, folk physics and biology

  • hu­man’s cul­tural learn­ing abilities

In this sec­tion, I’ll ex­plain the ac­count that refers to hu­man’s cul­tural learn­ing abil­ities. Cul­tural learn­ing typ­i­cally refers to a sub­set of so­cial learn­ing abil­ities such as min­dread­ing, imi­ta­tion, and teach­ing. I’ll use cul­tural and so­cial learn­ing in­ter­change­ably here.

Muthukr­ishna, Doe­beli, Chudek, and Hen­rich’s re­cent pa­per “The Cul­tural Brain Hy­poth­e­sis: How cul­ture drives brain ex­pan­sion, so­cial­ity, and life his­tory” con­tains one of the best con­cep­tual mod­els of the cu­mu­la­tive cul­tural evolu­tion story.

In the pa­per, Muthukr­ishna and co are in the busi­ness of mak­ing math­e­mat­i­cal mod­els which they can use to simu­late the im­pact of brain size, so­cial­ity, mat­ing struc­tures, and life his­tory. I’ll go the con­cep­tual fea­tures of the pri­mary model here and mo­ti­vate it’s plau­si­bil­ity.

The com­po­nents of the model are:

  • assumptions

  • lifecycle

  • parameters

Assumptions

Muthukr­ishna and co make the fol­low­ing as­sump­tions:

  • larger brains are more expensive

  • larger brains cor­re­sponds to an in­creased ca­pac­ity to store and man­age adap­tive information

  • adap­tive in­for­ma­tion re­duces the prob­a­bil­ity that its bearer will die or in­creases the prob­a­bil­ity that it will re­pro­duce.

One way to think about the role of adap­tive in­for­ma­tion is to think of hu­mans as oc­cu­py­ing “the cog­ni­tive niche.” Just as there are niches that par­tic­u­lar species may ex­ploit due to their biolog­i­cal fea­tures, a species may ex­ploit a cog­ni­tive niche through gath­er­ing, man­ag­ing, and ap­ply­ing in­for­ma­tion. Hu­mans are ap­par­ently unique with re­gards to the range of in­for­ma­tion we can gather, the abil­ity to ap­ply that in­for­ma­tion dur­ing de­vel­op­ment time, and the abil­ity to pass that on that in­for­ma­tion through gen­er­a­tions. Oc­cu­py­ing the cog­ni­tive niche al­lows a species like homo sapi­ens to in­no­vate new tools and ma­nipu­late the en­vi­ron­ment us­ing fire to smart phones.

Manag­ing and stor­ing in­for­ma­tion re­quires large brains and large brains are ex­pen­sive. Large brains are harder to feed, birth, and de­velop. There are then two con­flict­ing se­lec­tion pres­sures: the costli­ness of large brains and the ad­van­ta­geous­ness of adap­tive in­for­ma­tion.

Lifecycle

To tease out the im­pact of these differ­ent pres­sures Muthukr­ishna and co run a move agents through a life­cy­cle of BIRTH, LEARNING, MIGRATION, and SELECTION. Th­ese steps are straight­for­ward, agents are born, spend time learn­ing aso­cially or so­cially, mi­grate be­tween groups, and then are se­lected ac­cord­ing to the amount of adap­tive knowl­edge they ac­quired, costli­ness of their brain size, and the en­vi­ron­men­tal pay­off.

Parameters

The model in­cludes the fol­low­ing pa­ram­e­ters:

    • Trans­mis­sion fidelity

      • how ac­cu­rately can an agent learn from oth­ers?

    • Aso­cial learning

      • how effi­cient is aso­cial learn­ing?

    • Ecolog­i­cal richness

      • what is the en­vi­ron­men­tal pay­off for adap­tive knowl­edge?

    • Re­pro­duc­tive skew

      • how much more do those with more adap­tive knowl­edge re­pro­duce? Groups with a pair bond­ing struc­ture will have less re­pro­duc­tive skew. Groups with a polyg­y­nous mat­ing struc­ture will have more re­pro­duc­tive skew.

Th­ese pa­ram­e­ters are mod­ified for differ­ent agents and groups.

Results

What is the re­sult of the simu­la­tion? The simu­la­tion out­puts out the fol­low­ing causal re­la­tion­ships:

  1. Larger brains al­low for more adap­tive knowl­edge. This cre­ates a se­lec­tion pres­sure for larger brains. This is true for both so­cial and aso­cial learn­ers.

  2. More adap­tive knowl­edge al­lows for a larger car­ry­ing ca­pac­ity. As an agents in­vest in more so­cial learn­ing, this cre­ates a se­lec­tion pres­sure for larger groups as larger groups are richer sources of adap­tive knowl­edge than smaller groups.

  3. Larger groups of in­di­vi­d­u­als will tend to have more adap­tive knowl­edge, this puts pres­sure on longer ju­ve­nile pe­ri­ods for so­cial learn­ers.

  4. Ex­tended ju­ve­nile pe­ri­ods cre­ates se­lec­tion pres­sures for bet­ter learn­ing bi­ases, in par­tic­u­lar bi­ases con­cern­ing who to learn from.

  5. Bet­ter learn­ing bi­ases and oblique learn­ing “lead to the realm of cu­mu­la­tive cul­tural evolu­tion”

Here’s a nifty pic­ture dis­play­ing these re­la­tion­ships:

Th­ese re­sults are for the most part, weakly, ver­ified in the em­piri­cal world.

  • Brain size and so­cial learning

    • This re­la­tion­ship ap­pears to hold for pri­mates and birds.

  • Larger groups & larger brains

    • This re­la­tion­ship holds for pri­mates, but not for other taxa.

  • Brain size and ju­ve­nile period

    • There is a pos­i­tive re­la­tion­ship for pri­mates be­tween brain size and ju­ve­nile pe­riod.

  • Group size and ju­ve­nile period

    • There is a pos­i­tive re­la­tion­ship for pri­mates be­tween groups size and ab­solute ju­ve­nile pe­riod.

See 19-22 for re­la­tion­ship strengths and caveats.

The gen­eral model pro­vides a rather neat pic­ture. Cru­cially, there are pos­i­tive feed­back loops be­tween larger brains and so­cial learn­ing in the right en­vi­ron­ment. This in turn pushes to­wards longer ju­ve­nile pe­ri­ods and larger groups.

Cu­mu­la­tive Cul­tural Evolution

Un­der the right pa­ram­e­ter val­ues, Muthukr­ishna and co saw that a species which un­der­goes some­thing like cu­mu­la­tive cul­tural evolu­tion can be gen­er­ated. Re­call the pa­ram­e­ters of the model:

  • trans­mis­sion fidelity

  • re­pro­duc­tive skew

  • aso­cial learning

  • ecolog­i­cal richness

Let a cu­mu­la­tive cul­tural be the phe­nomenon where a group con­tains far more adap­tive knowl­edge than could likely be gen­er­ated by all of its in­di­vi­d­ual mem­bers via aso­cial learn­ing. Ba­si­cally, a group that ex­hibits cu­mu­la­tive cul­ture would be a species where it is ex­ceed­ingly un­likely that any of its mem­bers could gen­er­ate its pool of adap­tive knowl­edge via aso­cial learn­ing.

What val­ues of the model pa­ram­e­ters would pro­duce cu­mu­la­tive cul­ture?

Be­fore read­ing on, an­swer this ques­tion on your own.









  • Very high trans­mis­sion fidelity.

    • In or­der for adap­tive knowl­edge to ac­cu­mu­late in a species that species needs to be able to ac­cu­rately trans­mit that knowl­edge through gen­er­a­tions.

  • Low re­pro­duc­tive skew.

    • What would hap­pen if there were a high re­pro­duc­tive skew? Then in­di­vi­d­u­als who learn aso­cially and have es­pe­cially large brains would be re­warded and sur­vive. How­ever, pop­u­la­tions with large brained aso­cial learn­ers even­tu­ally go ex­tinct. This is be­cause as large brained aso­cial learn­ers sur­vive var­i­ance in so­cial learn­ing abil­ity de­creases. As var­i­ance in so­cial learn­ing abil­ity de­creases the pop­u­la­tion is: (i) un­able to cheaply ac­cu­mu­late knowl­edge via so­cial learn­ing and (ii) un­able to tran­si­tion to smaller brained so­cial learn­ers (re­mem­ber, brains are ex­pen­sive!).

  • Moder­ate aso­cial learn­ing.

    • So­cial learn­ers face a boot­strap­ping prob­lem—the adap­tive knowl­edge must come from some­where. This means that the species needs to be able to gen­er­ate in­no­va­tions aso­cially. How­ever, the species can­not in­vest in aso­cial learn­ing too much, oth­er­wise aso­cial learn­ing may be too effi­cient and so­cial learn­ing will not take off.

  • Ecolog­i­cal rich­ness.

    • Fi­nally, adap­tive knowl­edge must pay­off. Brains are too ex­pen­sive to grow, un­less there is a sig­nifi­cant benefit to their be­com­ing larger. The ecol­ogy can ac­count for that benefit.

Th­ese pa­ram­e­ter val­ues offer an ex­pla­na­tion of hu­man brain size, so­cial learn­ing ca­pa­bil­ities, and gen­eral suc­cess. More­over, they may also ex­plain why species like hu­mans are very low in num­ber. In or­der for there to be a species that in­vests in so­cial learn­ing at a very high rate:

  • Ecolo­gies must be suffi­ciently rich.

  • There must be a low re­pro­duc­tive skew

  • A species must be in the goldilocks zone with re­spect to aso­cial learning

  • Trans­mis­sion fidelity must be very high

In this en­vi­ron­ment there is sig­nifi­cant pres­sure to:

  • in­crease brain size

  • in­crease so­cial learn­ing

  • in­crease so­cial learn­ing efficiency

I per­son­ally think this stuff is very ex­cit­ing. We have a model for how brain sizes could 3x and how hu­mans can emerge as a uniquely so­cial species. There are ad­di­tional in­sights about the value of in­for­ma­tion and im­por­tance of mat­ing struc­ture.

AI Fore­cast­ing and Development

Now, what, if any­thing, is the im­port of this for AI fore­cast­ing and de­vel­op­ment?

There’s an ar­gu­ment that’s been float­ing awhile for some­time that goes some­thing like this:

Hu­mans are vastly more suc­cess­ful in cer­tain ways than other ho­minids, yet in evolu­tion­ary time, the dis­tance be­tween them is small. This sug­gests that evolu­tion in­duced dis­con­tin­u­ous progress in re­turns, if not in in­tel­li­gence it­self, some­where ap­proach­ing hu­man-level in­tel­li­gence. If evolu­tion ex­pe­rienced this, this sug­gests that ar­tifi­cial in­tel­li­gence re­search may do also. — AI Impacts

One can push back on the ar­gu­ment above with the claims that:

  • Evolu­tion wasn’t op­ti­miz­ing for im­prov­ing hu­man intelligence

  • Hu­man evolu­tion­ary suc­cess is not due to hu­man intelligence

In re­sponse to the first claim, there’s good rea­son to be­lieve, from the model above, that there are feed­back loops that en­able a species’ brain size and so­cial learn­ing ca­pa­bil­ities to take off. More­over, evolu­tion was “tar­get­ing” this take­off due the se­lec­tion pres­sures for adap­tive in­for­ma­tion and against large brains. Large brain so­cial learn­ers were sig­nifi­cantly “re­warded” in the right ecosys­tems. In the rele­vant sense, evolu­tion­ary dy­nam­ics op­ti­mized for and pushed hu­mans’ aso­cial and so­cial learn­ing ca­pa­bil­ities up­ward.

In re­sponse to the sec­ond claim, whether or not hu­mans are sig­nifi­cantly bet­ter aso­cial learn­ers than our pri­mate rel­a­tives is un­clear. Ob­vi­ously adult hu­mans have sig­nifi­cantly higher in­tel­li­gence than our pri­mate rel­a­tives, how­ever it’s un­clear to what ex­tent this higher in­tel­li­gence is a re­sult of aso­cial learn­ing rather than so­cial learn­ing. What is clear is that, from a very early age, hu­mans are vastly bet­ter so­cial learn­ers than our pri­mate an­ces­tors. Given this, it’s plau­si­ble to claim that hu­man suc­cess, at least rel­a­tive to other apes, de­rived from our aso­cial and so­cial learn­ing abil­ities. Th­ese abil­ities en­abled hu­mans to oc­cupy the cog­ni­tive niche.

Cu­mu­la­tive cul­tural evolu­tion ren­ders this ar­gu­ment more plau­si­ble. How­ever, be­fore get­ting too ex­cited, it’s worth not­ing that cu­mu­la­tive cul­tural evolu­tion is one plau­si­ble ac­count of hu­man suc­cess and in­tel­li­gence among an ar­ray of plau­si­ble ac­counts.

So much for fore­cast­ing, what of ar­tifi­cial in­tel­li­gence de­vel­op­ment? The de­vel­op­ment of hu­man in­tel­li­gence sug­gested by the cu­mu­la­tive cul­tural evolu­tion story, fol­lowed the chain be­low:

mod­er­ate aso­cial learn­ing ⇒ so­cial learn­ing.

One can imag­ine ma­chine learn­ing work de­vel­op­ing suffi­cient aso­cial learn­ing tech­niques and then ratch­et­ing ca­pa­bil­ities for­ward by com­bin­ing pre­vi­ous work through so­cial learn­ing tech­niques (po­ten­tially via imi­ta­tion learn­ing or model trans­fer, but likely much more so­phis­ti­cated tech­niques). On this model aso­cial learn­ing (prob­a­bly made up by a num­ber of differ­ent mod­ules and tech­niques), en­ables so­cial learn­ing to be­come a win­ning strat­egy. How­ever, so­cial learn­ing is an in­de­pen­dent thing, it is not built on top of aso­cial learn­ing mod­ules.

Re­lated to this, more work needs to be done de­ter­min­ing ca­pac­i­ties pri­mar­ily drive hu­man so­cial learn­ing. Hen­rich sug­gests that it is both min­dread­ing and imi­ta­tion. To­masello’s work stresses min­dread­ing, in par­tic­u­lar the abil­ity for hu­mans to de­velop joint at­ten­tion . An­swers to this is­sue would provide at least some ev­i­dence about what ma­chine learn­ing al­gorithms are likely to be more suc­cess­ful.

Another is­sue brought up by this work is whether so­cial learn­ing is an in­stinct, that is (roughly) whether it is en­coded by genes and not brought into ex­is­tence via cul­ture, or whether it is a gad­get. A gad­get is not en­coded by ge­netic in­for­ma­tion, but is in­stead de­vel­oped by cul­tural means. Sugges­tive work by Heyes ar­gues that so­cial learn­ing ca­pa­bil­ities could be de­vel­oped from hu­mans’ tem­per­a­ment, high work­ing mem­ory, and at­ten­tion ca­pa­bil­ities. If this is so, then the de­vel­op­ment of ar­tifi­cial in­tel­li­gence sketched above is likely flawed. It may in­stead look like:

tem­per­a­ment + com­pu­ta­tion power + work­ing mem­ory + at­ten­tion ⇒ so­cial learning

On this model, not only does aso­cial learn­ing en­able so­cial learn­ing to be a win­ning strat­egy, aso­cial learn­ing ca­pa­bil­ities com­pose so­cial learn­ing abil­ities. So­cial learn­ing is re­ally not that differ­ent from aso­cial learn­ing, it just a layer built on top of lower level in­tel­li­gence sys­tems.

Both of these mod­els sug­gest that AI de­vel­op­ment is cur­rently bot­tle­necked on the aso­cial learn­ing step. How­ever, once a thresh­old for aso­cial learn­ing is reached, in­tel­li­gence will in­crease at a vastly quick rate.

There’s a lot more to do here. I hope to have per­sua­sively mo­ti­vated the cu­mu­la­tive cul­tural evolu­tion story and the idea that it has im­por­tant up­shots for AI de­vel­op­ment and fore­cast­ing.