抓住人工智能科學(xué)機(jī)遇 A new golden age of discovery Seizing the AI for Science opportunity 2024

November2024

Anew

goldenageofdiscovery

SeizingtheAIforScienceopportunity

ConorGriffin|DonWallace|JuanMateos-Garcia|HannaSchieve|PushmeetKohli

Acknowledgements

ThankyoutoLouisaBartolo,Zo?BrammerandNickSwansonforresearchsupport,andtothefollowingindividualswhosharedinsightswithusthroughinterviewsand/orfeedbackonthedraft.Allviews,andanymistakes,belongsolelytotheauthors.

?igaAvsec,NicklasLundblad,JohnJumper,MattClifford,BenSouthwood,CraigDonner,Jo?lleBarral,TomZahavy,BeenKim,SebastianNowozin,MattClancy,MatejBalog,JenniferBeroshi,NitarshanRajkumar,BrendanTracey,YannisAssael,MassimilianoCiaramita,MichaelWebb,AgnieszkaGrabska-Barwinska,

AlessandroPau,TomLue,AgataLaydon,AnnaKoivuniemi,AbhishekNagaraj,HarryLaw,TomWestgarth,GuyWard-Jackson,AriannaManzini,StefanoBianchini,SameerVelankar,AnkurVora,SébastienKrier,

JoelZLeibo,ElisaLaiH.Wong,BenJohnson,DavidOsimo,AndreaHuber,DipanjanDas,EkinDogusCubuk,JacklynnStott,KelvinGuu,KiranVodrahalli,SanilJain,TrieuTrinh,RebecaSantamaria-Fernandez,

RemiLam,VictorMartin,NeelNanda,NenadTomasev,ObumEkeke,UchechiOkereke,FrancescaPietra,RishabhAgarwal,PeterBattaglia,AnilDoshi,YianYin.

GoogleDeepMind2

Introduction

GoogleDeepMind3

Introduction

PartA:Theopportunities

PartB:Theingredients

PartC:Therisks

PartD:Thepolicyresponse

Introduction

Aquietrevolutionisbrewinginlabsaroundtheworld,wherescientists’useofAIis

growing

exponentially

.

Oneinthreepostdocs

nowuselargelanguagemodelstohelpcarryoutliterature

reviews,coding,andediting.InOctober,thecreatorsofour

AlphaFold2

system,DemisHassabisandJohnJumperbecame

Nobel

LaureatesinChemistryforusingAItopredictthestructureofproteins,alongsidethescientistDavidBaker,forhisworktodesignnewproteins.Societywillsoonstarttofeelthesebenefitsmoredirectly,with

drugs

and

materials

designedwiththehelpofAIcurrentlymakingtheirwaythroughdevelopment.

Inthisessay,wetakeatourofhowAIistransformingscientificdisciplinesfromgenomicstocomputersciencetoweatherforecasting.SomescientistsaretrainingtheirownAImodels,whileothersarefine-tuningexistingAImodels,orusingthesemodels’predictionstoacceleratetheirresearch.Scientists

areusingAIasascientificinstrumenttohelptackleimportantproblems,suchas

designingproteinsthat

bindmoretightlytodiseasetargets

,butarealsograduallytransforminghowscienceitselfispractised.

Thereisagrowingimperativebehindscientists’embraceofAI.Inrecentdecades,scientistshave

continuedtodeliverconsequentialadvances,fromCovid-19vaccinestorenewableenergy.Butit

takes

aneverlargernumberofresearcherstomakethesebreakthroughs

,andto

transformtheminto

downstreamapplications

.Asaresult,eventhoughthescientificworkforcehasgrownsignificantly

overthepasthalf-century,

risingmorethan

sevenfold

intheUSalone,thesocietalprogressthat

wewouldexpecttofollow,hasslowed.Forinstance,muchoftheworldhaswitnesseda

sustained

slowdown

inproductivitygrowththatisunderminingthequalityofpublicservices.Progresstowardsthe2030SustainableDevelopmentGoals,whichcapturethebiggestchallengesinhealth,the

environment,andbeyond,is

stalling

.

Inparticular,scientistslookingtomakebreakthroughstodayincreasinglyrunintochallengesrelatingtoscaleandcomplexity,fromtheever-growingliteraturebasetheyneedtomaster,totheincreasinglycomplexexperimentstheywanttorun.

M

oderndeeplearningmethods

areparticularlywell-suited

tothese

scaleandcomplexitychallenges

andcancompressthetimethatfuturescientificprogresswouldotherwiserequire.Forinstance,instructuralbiology,asinglex-raycrystallographyexperimenttodeterminethestructureofaprotein

cantakeyearsofworkandcostapproximately$100,000

,

dependingontheprotein.The

AlphaFoldProteinStructureDatabase

nowprovidesinstantaccessto200millionpredictedproteinstructuresforfree.

ThepotentialbenefitsofAItosciencearenotguaranteed.AsignificantshareofscientistsalreadyuseLLM-basedtoolstoassistwitheverydaytasks,suchascodingandediting,buttheshareofscientistsusingAI-centricresearchapproaches

ismuchlower,albeit

risingrapidly

.IntherushtouseAI,some

earlyscientificusecaseshavehad

questionableimpact

.PolicymakerscanhelpaccelerateAI’suseandsteerittowardshigher-impactareas.The

USDepartmentofEnergy,

the

EuropeanCommission

,theUK’s

Royal

Society,

andthe

USNationalAcademies

,amongothers,haverecentlyrecognisedtheAIforScienceopportunity.Butnocountryhasyetputacomprehensivestrategyinplacetoenableit.

GoogleDeepMind4

Introduction

PartA:Theopportunities

PartB:Theingredients

PartC:Therisks

PartD:Thepolicyresponse

Wehopeouressaycaninformsuchastrategy.Itisaimedatthosewhomakeandinfluencesciencepolicy,andfundingdecisions.Wefirstidentify5opportunitieswherethereisagrowingimperativetouseAIinscienceandexaminetheprimaryingredientsneededtomakebreakthroughsinthese

areas.Wethenexplorethemostcommonly-citedrisksfromusingAIinscience,suchastoscientificcreativityandreliability,andarguethatAIcanultimatelybenetbeneficialineacharea.WeconcludewithfourpublicpolicyideastohelpusherinanewgoldenageofAI-enabledscience.

ThroughouttheessaywedrawoninsightsfromovertwodozeninterviewswithexpertsfromourownAIforScienceprojects,aswellasexternalexperts.Theessaynaturallyreflectsourvantagepointasaprivatesectorlab,butwebelievethecasewemakeisrelevantforthewholeofscience.WehopethatreaderswillrespondbysharingtheirtakeonthemostimportantAIforScienceopportunities,

ingredients,risksandpolicyideas.

PartA:Theopportunities06

PartB:Theingredients13

PartC:Therisks25

PartD:Thepolicyresponse32

GoogleDeepMinds

PartA

The

opportunities

6

Introduction

PartA:Theopportunities

PartB:Theingredients

PartC:Therisks

PartD:Thepolicyresponse

Theopportunities

Scientistsaimtounderstand,predict,andinfluencehowthenaturalandsocialworldswork,toinspireandsatisfycuriosity,andtotackleimportantproblemsfacingsociety.

Technologiesandmethods

,

likethemicroscope,x-raydiffraction,andstatistics,arebothproductsofscienceandenablersof

it.Overthepastcentury,scientistshaveincreasinglyreliedontheseinstrumentstocarryouttheir

experimentsandadvancetheirtheories.Computationaltoolsandlarge-scaledataanalysishave

becomeparticularlyimportant,enablingeverythingfromthediscoveryoftheHiggsbosontothe

mappingofthehumangenome.Fromoneview,scientists’growinguseofAIisalogicalextensionofthislong-runningtrend.Butitmayalsosignalsomethingmuchmoreprofound-adiscontinuousleapinthelimitsofwhatscienceiscapableof.

RatherthanlistingallareaswhereitispossibletouseAI,wehighlightfiveopportunitieswherewethinkthereisanimperativetouseit.Theseopportunitiesapplyacrossdisciplinesandaddressaspecific

bottleneck,relatedtoscaleandcomplexity,thatscientistsincreasinglyfaceatdifferentpointsinthescientificprocess,fromgeneratingpowerfulnovelhypothesestosharingtheirworkwiththeworld.

5opportunitiestoacceleratesciencewithAl

1.Knowledge2.Data3.Experiments

TransformhowscientistsdigestGenerate,extract,andannotateSimulate,accelerateandinform

andcommunicateknowledgelargescientificdatasetscomplexexperiments

Q

4.Models5.Solutions

ModelcomplexsystemsandhowIdentifynovelsolutionstoproblems

theircomponentsinteractwithlargesearchspaces

GoogleDeepMind7

Introduction

PartA:Theopportunities

PartB:Theingredients

PartC:Therisks

PartD:Thepolicyresponse

1.Knowledge

Transformhowscientistsdigestandcommunicateknowledge

Tomakenewdiscoveries,scientistsneedtomasterapre-existingbodyofknowledgethatcontinuestogrow

exponentially

andbecomeevermorespecialised.This

‘burdenofknowledge

’helpsexplainwhyscientists

making

transformative

discoveries

areincreasinglyolder,interdisciplinary,andlocatedateliteuniversities,andwhythe

shareofpapers

authoredbyindividuals,orsmallteams,isdeclining,eventhough

smallteamsareoftenbetter-placedtoadvancedisruptivescientificideas

.Whenit

comestosharingtheirresearchtherehavebeenwelcomeinnovationssuchaspreprintserversandcoderepositories,butmostscientistsstillsharetheirfindingsin

dense,jargon-heavy,

English-only

papers

.Thiscanimpederatherthanigniteinterestinscientists’work,includingfrompolicymakers,businesses,andthepublic.

ScientistsarealreadyusingLLMs,andearlyscientificassistantsbasedonLLMs,tohelpaddress

thesechallenges,suchasby

synthesisingthemostrelevantinsights

fromtheliterature.Inanearly

demonstration

,ourScienceteam

usedour

Gemini

LLMtofind,extract,andpopulatespecificdatafromthemostrelevantsubsetof200,000papers,withinaday.Upcominginnovations,suchasfine-tuningLLMsonmorescientificdataandadvancesinlongcontextwindowsandcitationuse,will

steadilyimprovethesecapabilities.Asweexpandonbelow,theseopportunitiesarenotwithoutrisk.Buttheyprovideawindowtofundamentallyrethinkcertainscientifictasks,suchaswhatitmeansto‘read’or‘write’ascientificpaperinaworldwhereascientistcanuseanLLMtohelpcritiqueit,tailoritsimplicationsfordifferentaudiences,ortransformitintoan‘interactivepaper’or

audioguide

.

2.Data

Generate,extract,andannotatelargescientificdatasets

Despitepopularnarrativesaboutaneraofdataabundance,thereisachroniclackofscientificdataonmostofthenaturalandsocialworld,fromthesoil,deepoceanandatmosphere,totheinformaleconomy.AIcouldhelpindifferentways.Itcouldmakeexistingdatacollectionmoreaccurate,forexampleby

reducingthenoiseanderrorsthatcanoccurwhen

sequencingDNA

,

detectingcelltypesinasample,

or

capturinganimalsounds

.ScientistscanalsoexploitLLMs’growingabilitytooperateacrossimages,videoandaudio,toextracttheunstructuredscientificdatathatisburiedinscientificpublications,

archives,andlessobviousresourcessuchasinstructionalvideos,andconvertitintostructureddatasets.

AIcanalsohelptoannotatescientificdatawiththesupportinginformationthatscientistsneedinordertouseit.Forexample,atleastone-thirdofmicrobialproteins

arenotreliablyannotated

withdetailsaboutthefunction(s)thattheyarethoughttoperform.In2022,ourresearchers

usedAIto

predictthefunctionofproteins

,leadingtonewentriesinthe

UniProt

,

Pfam

and

InterPro

databases.

GoogleDeepMind8

Introduction

PartA:Theopportunities

PartB:Theingredients

PartC:Therisks

PartD:Thepolicyresponse

AImodels,oncevalidated,canalsoserveasnewsourcesofsyntheticscientificdata.Forexample,our

AlphaProteo

proteindesignmodelistrainedonmorethan100millionAI-generatedprotein

structuresfromAlphaFold2,alongwithexperimentalstructuresfromthe

ProteinDataBank

.TheseAIopportunitiescancomplementandincreasethereturnonothermuch-neededeffortstogeneratescientificdata,suchas

digitisingarchives

,orfundingnewdatacapturetechnologiesandmethods,likeeffortsunderwayinsinglecellgenomicstocreatepowerfuldatasetsofindividualcellsin

unprecedenteddetail.

3.Experiments

Simulate,accelerateandinformcomplexexperiments

Manyscientificexperimentsareexpensive,complex,andslow.Somedonothappenatallbecauseresearcherscannotaccessthefacilities,participantsorinputsthattheyneed.Fusionisacase

inpoint.Itpromisesanenergysourcethatispracticallylimitless,emission-freeandcouldenable

thescalingofenergy-intensiveinnovations,likedesalination.Torealisefusion,scientistsneedto

createandcontrolplasma-afourthfundamentalstateofmatter.However,thefacilitiesneededarehugelycomplextobuild.

ITER

’sprototypetokamakreactorbeganconstructionin2013,butplasmaexperimentsare

notsettobegin

untilthemid-2030sattheearliest,althoughothershopetobuildsmallerreactorsonshortertimelines.

AIcouldhelptosimulatefusionexperimentsandenablemuchmoreefficientuseofsubsequentexperimenttime.Oneapproachistorunreinforcementlearningagentsonsimulationsofphysicalsystems.Between2019and2021,ourresearcherspartneredwiththeSwissFederalInstituteof

TechnologyLausanneto

demonstrate

howtouseRLtocontroltheshapeofplasmainasimulationofatokamakreactor.Theseapproachescouldbeextendedtootherexperimentalfacilities,suchas

particleaccelerators

,

telescopearrays

,or

gravitationalwavedetectors

.

UsingAItosimulateexperimentswilllookverydifferentacrossdisciplines,butacommonthreadis

thatthesimulationswillofteninformandguidephysicalexperiments,ratherthansubstituteforthem.Forexample,theaverageperson

hasmorethan9,000

missensevariants,orsinglelettersubstitutionsintheirDNA.Mostofthesegeneticvariantsarebenignbutsomecandisruptthefunctionsthat

proteinsperform,contributingtoraregeneticdiseaseslikecysticfibrosisaswellascommondiseaseslikecancer.Physicalexperimentstotesttheeffectsofthesevariantsareoftenlimitedtoasingle

protein.Our

AlphaMissensemodel

classifies89%ofthe71millionpotentialhumanmissensevariants

aslikelyharmfulorbenign,enablingscientiststofocustheirphysicalexperimentsonthemostlikelycontributorstodisease.

GoogleDeepMind9

Introduction

PartA:Theopportunities

PartB:Theingredients

PartC:Therisks

PartD:Thepolicyresponse

Allpossible71millionhumanmissensevariants

AlphaMissensepredictions:

Likelybenign57%

Likelypathogenic

32%

.Uncertain

11%

Humanannotations:

Seeninhumans~6%

Confirmedbyhumanexperts~0.1%

AlphaMissensepredictedthepathogenicityofallpossible71millionmissensevariants.Itclassified89%

-predicting57%werelikelybenignand32%werelikelypathogenic.

4.Models

Modelcomplexsystemsandhowtheircomponentsinteract

Ina

1960paper,

theNobelPrizewinningphysicistEugeneWignermarvelledatthe“unreasonable

effectiveness”ofmathematicalequationsformodellingimportantnaturalphenomena,suchas

planetarymotion.However,overthepasthalfcentury,modelsthatrelyonsetsofequationsor

otherdeterministicassumptionshavestruggledtocapture

thefullcomplexityofsystemsinbiology,

economics,weather,andelsewhere

.Thisreflectsthesheernumberofinteractingpartsthatmakeupthesesystems,aswellastheirdynamismandpotentialforemergent,randomorchaoticbehaviour.Thechallengesinmodellingthesesystemsimpedesscientists’abilitytopredictorcontrolhowtheywillbehave,includingduringshocksorinterventions,suchasrisingtemperatures,anewdrug,ortheintroductionofataxchange.

GoogleDeepMind10

Introduction

PartA:Theopportunities

PartB:Theingredients

PartC:Therisks

PartD:Thepolicyresponse

AIcouldmoreaccuratelymodelthesecomplexsystemsbyingestingmoredataaboutthem,andlearningmorepowerfulpatternsandregularitieswithinthisdata.Forexample,modernweather

forecastingisatriumphofscienceandengineering.Forgovernmentsandindustry,itinforms

everythingfromrenewableenergyplanningtopreparingfor

hurricanes

and

floods

.Forthepublic,

theweatheristhemostpopularnon-brandedqueryonGoogleSearch.Traditional

numeralprediction

methods

arebasedoncarefully-definedphysicsequationsthatprovideaveryuseful,yetimperfect,approximationoftheatmosphere’scomplexdynamics.Theyarealsocomputationallyexpensiveto

run.In2023,wereleaseda

deeplearningsystem

thatpredictsweatherconditionsupto10daysinadvance,whichoutperformedtraditionalmodelsonaccuracyandpredictionspeed.Asweexpandonbelow,usingAItoforecastweathervariablescouldalsohelptomitigateandrespondtoclimatechange.Forinstance,whenpilotsflythroughhumidregionsitcancausecondensationtrailsthat

contributeto

aviation’sglobalwarmingimpact.Googlescientists

recentlyused

AItopredictwhenandwherehumidregionsmayarisetohelppilotsavoidflyingthroughthem.

Inmanycases,AIwillenrichtraditionalapproachestomodellingcomplexsystemsratherthanreplacethem.Forexample,agent-basedmodellingsimulatesinteractionsbetweenindividualactors,like

firmsandconsumers,tounderstandhowtheseinteractionsmightaffectalargermorecomplex

systemliketheeconomy.Traditionalapproachesrequirescientiststospecifybeforehandhowthesecomputationalagentsshouldbehave.Ourresearchteams

recentlyoutlined

howscientistscoulduseLLMstocreatemoreflexiblegenerativeagentsthatcommunicateandtakeactions,suchassearchingforinformationormakingpurchases,whilealsoreasoningaboutandrememberingtheseactions.

Scientistscouldalsousereinforcementlearningtostudyhowtheseagentslearnandadapttheirbehaviourin

moredynamicsimulations

,forexampleinresponsetotheintroductionofnewenergypricesorpandemicresponsepolicies.

5.Solutions

Identifynovelsolutionstoproblemswithlargesearchspaces

Manyimportantscientificproblemscomewithapracticallyincomprehensiblenumberofpotential

solutions.Forexample,biologistsandchemistsaimtodeterminethestructure,characteristics,andfunction(s)ofmoleculessuchasproteins.Onegoalofsuchworkistohelpdesignnovelversionsofthesemoleculestoserveasantibodydrugs,plastic-degradingenzymesornewmaterials.However,todesignasmallmoleculedrug,scientistsface

morethan10

60

potentialoptions.Todesignaproteinwith400standardaminoacids,theyface20400options.Theselargesearchspacesarenotlimitedtomoleculesbutarecommonplaceformanyscientificproblems,suchasfindingthebestprooffora

mathsproblem,themostefficientalgorithmforacomputersciencetask,orthe

bestarchitecturefor

acomputerchip

.

GoogleDeepMind11

Traditionally,scientistsrelyonsomecombinationofintuition,trialanderror,iteration,orbruteforce

computingtofindthebestmolecule,proof,oralgorithm.However,thesemethodsstruggletoexploitthehugespaceofpotentialsolutions,leavingbetteronesundiscovered.AIcan

openupnewpartsof

thesesearchspaces

whilealsohominginmorequicklyonthesolutionsthataremostlikelytobeviableanduseful-adelicatebalancingact.Forexample,inJuly,our

AlphaProofandAlphaGeometry2

systemscorrectlysolvedfouroutofsixproblemsfromthe

InternationalMathematicalOlympiad

,

anelitehighschoolcompetition.ThesystemsmakeuseofourGeminiLLMarchitecturetogeneratealargenumberofnovelideasandpotentialsolutionsforagivenmathsproblem,andcombinethiswithsystemsgroundedinmathematicallogicthatcaniterativelyworktowardsthecandidatesolutionsthataremostlikelytobecorrect.

AIscientistsorAI-empoweredscientists?

ThisgrowinguseofAIinscience,andtheemergenceofearlyAIscientificassistants,raisesquestionsabouthowfastandhowfarthecapabilitiesofAImayadvanceandwhatthiswillmeanforhuman

scientists.CurrentLLM-basedAIscientificassistantsmakearelativelysmallcontributiontoa

relativelynarrowrangeoftasks,suchassupportingliteraturereviews.Thereareplausiblenear-termscenariosinwhichtheybecomebetteratthesetasksandbecomecapableofmoreimpactfulones,suchashelpingtogeneratepowerfulhypotheses,orhelpingtopredicttheoutcomesofexperiments.

However,currentsystemsstillstrugglewiththedeepercreativityand

reasoning

thathumanscientistsrelyonforsuchtasks.

EffortsareunderwaytoimprovetheseAIcapabilities

,forexamplebycombiningLLMswithlogicaldeductionengines,asinour

AlphaProofandAlphaGeometry2

examples,butfurtherbreakthroughsareneeded.Theabilitytoaccelerateorautomateexperimentswillalsobeharderforthosethatrequirecomplicatedactionsinwetlabs,interactingwithhumanparticipants,orlengthy

processes,suchasmonitoringdiseaseprogression.Althoughagain,workisunderwayinsomeoftheseareas,suchasnewtypesoflaboratoryroboticsand

automatedlabs

.

EvenasAIsystems’capabilitiesimprove,thegreatestmarginalbenefitwillcomefromdeploying

theminusecasesthatplaytotheirrelativestrengths-suchastheabilitytorapidlyextractinformationfromhugedatasets-andwhichhelpaddressgenuinebottleneckstoscientificprogresssuchasthefiveopportunitiesoutlinedabove,ratherthanautomatingtasksthathumanscientistsalreadydowell.AsAIenablescheaperandmorepowerfulscience,demandforscienceandscientistswillalsogrow.

Forexample,recentbreakthroughshavealreadyledtoaslewofnewstartupsinareaslike

protein

design

,

materialscience

and

weatherforecasting

.Unlikeothersectors,anddespite

pastclaimstothe

contrary,

futuredemandforscienceappearspracticallylimitless.Newadvances

havealwaysopened

upnew,unpredictableregions

inthescientificmapofknowledge,andAIwilldosimilar.As

envisioned

byHerbertSimon,AIsystemswillalsobecomeobjectsofscienceresearch,withscientistssettoplayaleadingroleinevaluatingandexplainingtheirscientificcapabilities,aswellasindevelopingnew

typesofhuman-AIscientificsystems.

GoogleDeepMind12

PartB

The

ingredients

13

Introduction

PartA:Theopportunities

PartB:Theingredients

PartC:Therisks

PartD:Thepolicyresponse

Theingredients

WeareinterestedintheingredientsthatambitiousAIforScienceeffortsneedtosucceed-bothat

theindividualresearcheffortlevelandatthelevelofthescienceecosystem,wherepolicymakershavemorescopetoshapethem.Theexpertsthatweinterviewedroutinelycitedseveralingredientsthat

weorganisedintoatoymodel,whichwecalltheAIforScienceproductionfunction.Thisproductionfunctionisnotmeanttobeexhaustive,prescriptive,oraneatlinearprocess.Theingredientswillbeintuitivetomany,butourinterviewsrevealedanumberoflessonsaboutwhattheylooklikeinpracticewhichwesharebelow.

TheAIforscienceproductionfunction

Startingpoint

InfrastructureDoingtheresearch

Drivingimpact

Adoption

Partnerships

Safety&responsibility

Organisationaldesign

Problemselection

Interdisciplinarity

Evaluations

Compute

Data

GoogleDeepMind14

Introduction

PartA:Theopportunities

PartB:Theingredients

PartC:Therisks

PartD:Thepolicyresponse

1.Problemselection

Pursueambitious,AI-shapedproblems

Scientificprogressrestsonbeingabletoidentifyanimportantproblemandasktherightquestion

abouthowtosolveit.In

theirexploration

intothegenesisofscientificbreakthroughs,Venkatesh

NarayanamurtiandJeffreyY.Tsaodocumenthowimportantthereciprocalandrecursiverelationshipbetweenquestionsandanswersis,includingtheimportanceofaskingambitiousnewquestions.

OurScienceteamstartsbythinkingaboutwhetherapotentialresearchproblemissignificantenoughtojustifyasubstantialinvestmentoftimeandresources.OurCEODemisHassabishasamentalmodeltoguidethisassessment:thinkingaboutallofscienceasatreeofknowledge.Weareparticularly

interestedintheroots-fundamental‘rootnodeproblems’like

proteinstructureprediction

or

quantumchemistry

that,ifsolved,couldunlockentirelynewbranchesofresearchandapplications.

ToassesswhetherAIwillbesuitableandadditive,welookforproblemswithcertaincharacteristics,

suchashugecombinatorialsearchspaces,largeamountsofdata,andaclearobjectivefunctionto

benchmarkperformanceagainst.OftenaproblemissuitableforAIinprinciple,buttheinputsaren’t

yetinplaceanditneedstobestoredforlater.Oneoftheoriginal

inspirations

forAlphaFoldwas

conversationsthatDemishadmanyyearspriorasastudentwithafriendwhowasobsessedwith

theproteinfoldingproblem.Manyrecentbreakthroughsalsofeaturethiscomingtogetherofan

importantscientificproblemandanAIapproachthathasreachedapointofmaturity.Forexample,

our

fusion

effortwasaidedbyanovel

reinforcementlearningalgorithm

calledmaximumaposteriori

policyoptimization,whichhadonlyjustbeenreleas