LevelsofAGI
Table1.Aleveled,matrixedapproachtowardclassifyingsystemsonthepathtoAGIbasedondepth(performance)andbreadth(gener-
ality)ofcapabilities.Theassignmentofexamplesystemstocellsisapproximate.UnambiguousclassificationofAIsystemswillrequire
astandardizedbenchmarkoftasks,aswediscussinSection5. NotethatgeneralsystemsthatbroadlyperformatalevelNmaybeable
toperformanarrowsubsetoftasksathigherlevels.The“CompetentAGI”level,whichhasnotbeenachievedbyanypublicsystemsat
thetimeofwriting,bestcorrespondstomanypriorconceptionsofAGI,andmayprecipitaterapidsocietalchangeonceachieved.
Performance(rows)x Narrow General
Generality(columns) clearlyscopedtaskorsetoftasks wide range of non-physical tasks, includ-
ingmetacognitivetaskslikelearningnew
skills
Level0:NoAI NarrowNon-AI GeneralNon-AI
calculatorsoftware;compiler human-in-the-loop computing, e.g.,Ama-
zonMechanicalTurk
Level1:Emerging EmergingNarrowAI EmergingAGI
equal to or somewhat better than an un- GOFAI(Boden,2014);simplerule-based ChatGPT (OpenAI, 2023), Bard
skilledhuman systems,e.g.,SHRDLU(Winograd,1971) (Aniletal., 2023), Llama 2
(Touvronetal., 2023), Gemini
(Pichai&Hassabis,2023)
Level2:Competent CompetentNarrowAI CompetentAGI
atleast50thpercentileofskilledadults toxicity detectors such as Jigsaw notyetachieved
(Dasetal., 2022); Smart Speakers
suchasSiri(Apple),Alexa(Amazon),or
GoogleAssistant(Google);VQAsystems
such as PaLI(Chenetal., 2023); Watson
(IBM);SOTALLMsforasubsetoftasks
(e.g.,shortessaywriting,simplecoding)
Level3:Expert ExpertNarrowAI ExpertAGI
atleast90thpercentileofskilledadults spelling & grammar checkers such as notyetachieved
Grammarly (Grammarly, 2023); gen-
erative image models such as Ima-
gen (Sahariaetal., 2022) or Dall-E 2
(Rameshetal.,2022)
Level4:Virtuoso VirtuosoNarrowAI VirtuosoAGI
atleast99thpercentileofskilledadults Deep Blue (Campbelletal., 2002), Al- notyetachieved
phaGo(Silveretal.,2016;2017)
Level5:Superhuman SuperhumanNarrowAI ArtificialSuperintelligence(ASI)
outperforms100%ofhumans AlphaFold (Jumperetal., 2021; notyetachieved
Varadietal., 2021), AlphaZero
(Silveretal.,2018),StockFish(Stockfish,
2023)
safety (e.g., acquiring strong knowledge of chemical en- qualitythanmostpeopleareabletodraw;however,thesys-
gineering before acquiring strong ethical reasoning skills tem has failuremodes(e.g., drawinghandswith incorrect
maybe a dangerouscombination). Note also thatthe rate numbers of digits, rendering nonsensical or illegible text)
of progressionbetween levels of performanceand/or gen- that prevent it from achieving a “Virtuoso” performance
eralitymaybenonlinear. Acquiringthecapabilitytolearn designation. Whiletheoreticallyan“Expert”levelsystem,
newskillsmayparticularlyaccelerateprogresstowardthe inpracticethesystemmayonlybe“Competent,”because
nextlevel. prompting interfaces are too complex for most end-users
toelicitoptimalperformance(asevidencedbyuserstudies
While this taxonomyrates systems according to their per-
(Zamfirescu-Pereiraetal., 2023) andthe existenceof mar-
formance, systems that are capable of achieving a cer-
ketplaces(e.g.,PromptBase)inwhichskilledpromptengi-
tainlevelofperformance(e.g.,againstagivenbenchmark)
neers sell prompts). This observation emphasizes the im-
may not match this level in practice when deployed. For
portanceofdesigningecologicallyvalidbenchmarks(that
instance, user interface limitations may reduce deployed
approximate deployed rather than idealized performance),
performance. Consider DALLE-2 (Rameshetal., 2022),
aswellastheimportanceofconsideringthehuman-AIin-
which we estimate as a Level 3 Narrow AI (“Expert Nar-
teractionparadigms.
rowAI”)inourtaxonomy. Weestimatethe“Expert”level
ofperformancesinceDALLE-2producesimagesofhigher Thehighestlevelinourmatrixintermsofcombinedperfor-
5