Implications of Quantum Computing for Artificial Intelligence alignment research (ABRIDGED)

I re­cently co-au­thored a pa­per with Pablo Moreno, Im­pli­ca­tions of Quan­tum Com­put­ing for Ar­tifi­cial In­tel­li­gence al­ign­ment re­search, which can be ac­cessed through arXiv.

Our pa­per fo­cuses on an­a­lyz­ing the in­ter­ac­tion be­tween Quan­tum Com­put­ing (QC) and the cur­rent land­scape of re­search in Ar­tifi­cial In­tel­li­gence (AI) al­ign­ment, and we weakly con­clude that knowl­edge of QC is un­likely to be helpful to ad­dress cur­rent bot­tle­necks in AI al­ign­ment.

In this post I in­tend to very briefly sum­ma­rize the gen­er­a­tor of the main ar­gu­ments of the pa­per, con­vey the main con­clu­sions and in­vite the reader to read the full re­port if they wish to get a deeper in­tu­ition or see our list of open ques­tions.

It might be tempt­ing to con­clude that QC has im­por­tant im­pli­ca­tions for AI al­ign­ment since there are some promis­ing av­enues of re­search in Quan­tum Ma­chine Learn­ing, so QC might end up be­ing an in­te­gral com­po­nent of fu­ture AI sys­tems.

How­ever, we ar­gue that for the most part QC can be sim­plified away as a black box ac­cel­er­a­tor that lets you ex­po­nen­tially speed up cer­tain com­pu­ta­tions—the so-called quan­tum speedup. This is rele­vant be­cause we be­lieve that cur­rent re­search in al­ign­ment should feel free to use in­vo­ca­tions to that kind of or­a­cles to dis­cuss for­mal solu­tions for the differ­ent prob­lems of the field, and worry about its con­crete effi­cient im­ple­men­ta­tion later down the line.

The biggest challenge that QC sup­poses for AI Align­ment pur­poses is what we called quan­tum obfus­ca­tion—the fact that read­ing the con­tents of a quan­tum com­put­ing is hard to do clas­si­cally, which may ren­der some over­sight mechanisms we might de­sign use­less.

How­ever most re­search agen­das and prob­lems AI al­ign­ment re­searchers are work­ing on have lit­tle to do with the ac­tual im­ple­men­ta­tion of low-level over­sight mechanisms, and fo­cus rather on al­ign­ing the in­cen­tives of AI sys­tems to co­op­er­a­tively send in­for­ma­tion to its op­er­a­tors in an in­ter­pretable way.

Fur­ther­more, there might be di­rect analogues of clas­si­cal over­sight in the quan­tum realm, so re­search con­ducted in this stage may be res­cued later in­stead of wasted.

We have also looked into rea­sons why QC might be a good tool to solve some AI al­ign­ment sub­prob­lems, and iden­ti­fied a cou­ple of cases. They are how­ever not es­pe­cially promis­ing.

First, we iden­ti­fied the pos­si­bil­ity of us­ing ac­cess to quan­tum com­put­ing as an am­plifi­ca­tion of an over­seer that ver­ifies or pro­vides the re­ward in a way hard to un­der­stand by an agent be­ing ver­ified—we call this ex­ploit­ing quan­tum asym­me­try.

Se­cond, we might be able to ex­ploit quan­tum iso­la­tion to mon­i­tor quan­tum agents—the fact that a quan­tum com­puter has to re­main iso­lated to be able to achieve quan­tum speedups. This might point in the di­rec­tion of a trip­wire that would al­low us to de­tect whether a sys­tem has in­ter­acted with the out­side world with­out our con­sent. Albeit we have not looked into this in-depth, we weakly ar­gue against the pos­si­bil­ity of an effi­cient schema of this type.

Long story short, we do not be­lieve that QC is a crit­i­cal area of knowl­edge for ad­vanc­ing cur­rent re­search agen­das of tech­ni­cal AI al­ign­ment, and I would weakly recom­mend against pur­su­ing a ca­reer in it for this pur­pose or fund­ing re­search in this in­ter­sec­tion.

For the full dis­cus­sion of our rea­son­ing and a list of open ques­tions, I re­fer the reader to our pa­per.

This post was writ­ten by Jaime Sevilla, sum­mer fel­low at the Fu­ture of Hu­man­ity In­sti­tute. I want to thank Pablo Moreno for work­ing with me on this topic and his feed­back on this sum­mary.