4202
nuJ
5
]IA.sc[
4v26420.1132:viXra
Position: Levels of AGI for Operationalizing Progress on the Path to AGI
MeredithRingelMorris1 JaschaSohl-Dickstein2 NoahFiedel2 TrisWarkentin2 AllanDafoe3
AleksandraFaust2 ClementFarabet3 ShaneLegg3
Abstract language models (LLMs); some predict AI will broadly
outperform humans within about a decade (Bengioetal.,
We propose a framework for classifying the ca-
2023); some even assert that current LLMs are AGIs
pabilities and behavior of Artificial General In-
(Agu¨erayArcas&Norvig,2023).
telligence (AGI) models and their precursors.
This framework introduces levels of AGI per- TheconceptofAGIisimportantasitmapsontogoalsfor,
formance, generality, and autonomy, providing predictionsabout,andrisksofAI:
a common language to compare models, assess
Goals:Achievinghuman-level“intelligence”isanimplicit
risks, and measure progress along the path to
or explicit north-star goal for many in our field, from the
AGI.To developourframework,we analyzeex-
1956 Dartmouth AI Conference (McCarthyetal., 1955)
isting definitions of AGI, and distill six princi-
thatkick-startedthemodernfieldofAI,totoday’sleading
ples that a useful ontology for AGI should sat-
AIresearchfirms,whosemissionstatementsincludegoals
isfy. With these principlesin mind, we propose
such as “ensure transformative AI helps people and soci-
“Levels of AGI” based on depth (performance)
ety” (Anthropic, 2023a) and“ensurethatartificialgeneral
and breadth (generality) of capabilities, and re-
intelligencebenefitsallofhumanity”(OpenAI,2023).
flect on how current systems fit into this ontol-
ogy. We discuss the challenging requirements Predictions: The concept of AGI is related to a predic-
forfuturebenchmarksthatquantifythebehavior tion aboutprogressin AI,namelythatit is towardgreater
andcapabilitiesofAGImodelsagainsttheselev- generality, approaching and exceeding human generality.
els. Finally,wediscusshowtheselevelsofAGI Additionally, AGI is typically intertwined with a notion
interactwith deploymentconsiderationssuch as of “emergent” properties (Weietal., 2022), i.e. capabili-
autonomy and risk, and emphasize the impor- ties not explicitly anticipated by the developer. Such ca-
tance of carefully selecting Human-AI Interac- pabilitiesofferpromise,perhapsincludingabilitiesthatare
tion paradigms for responsible and safe deploy- complementarytotypicalhumanskills,enablingnewtypes
mentofhighlycapableAIsystems. of interaction or novel industries. Such predictions about
AGI’s capabilities in turn predict likely societal impacts;
AGI may have significant economic implications, i.e.,
1. Introduction
reaching the necessary criteria for widespread labor sub-
stitution (Ellingrudetal., 2023; Dell’Acquaetal., 2023;
Artificial General Intelligence (AGI) is an important and
Eloundouetal.,2023),aswellasgeo-politicalimplications
sometimes controversial concept in computing research,
relating not only to the economic advantages AGI may
used to describe an AI system that is at least as capa-
confer,butalsotomilitaryconsiderations(Kissingeretal.,
ble as a human at most tasks. Given the rapid advance-
2022).
ment of Machine Learning (ML) models, the concept
of AGI has grown from a subject of philosophical de- Risks: Lastly, AGI is viewed by some as a concept
bate, to one which also has near-term practical relevance. for identifying the point when there are extreme risks
Some expertsbelieve that “sparks” of AGI (Bubecketal., (Shevlaneetal., 2023; Bengioetal., 2023), as some spec-
2023) are already present in the latest generation of large ulatethatAGIsystemsmightbeabletodeceiveandmanip-
ulate, accumulate resources, advance goals, behave agen-
1Google DeepMind, Seattle, WA, USA 2Google Deep-
tically,outwithumansinbroaddomains,displacehumans
Mind, Mountain View, CA, USA 3Google DeepMind, Lon-
don, UK. Correspondence to: Meredith Ringel Morris <mer- fromkeyroles,and/orrecursivelyself-improve.
rie@google.com>.
Inthispositionpaper,wearguethatitiscriticalforthe
Proceedings of the 41st International Conference on Machine AIresearchcommunitytoexplicitlyreflectonwhatwe
Learning, Vienna, Austria. PMLR 235, 2024. Copyright 2024 mean by “AGI,” and aspire to quantify attributes like
bytheauthor(s).
1