OpenAI Codex
OpenAICodexisadescendantofGPT-3;itstrainingdatacontainsbothnaturallanguageandbillionsoflinesofsourcecodefrompubliclyavailablesources,includingcodeinpublicGitHubrepositories.OpenAICodexismostcapableinPython,butitisalsoproficientinoveradozenlanguagesincludingJavaScript,Go,Perl,PHP,Ruby,SwiftandTypeScript,andevenShell.Ithasamemoryof14KBforPythoncode,comparedtoGPT-3whichhasonly4KB—soitcantakeintoaccountover3xasmuchcontextualinformationwhileperformingany task.
GPT-3’smainskillisgeneratingnaturallanguageinresponsetoanaturallanguageprompt,meaningtheonlywayitaffectstheworldisthroughthemindofthereader.OpenAICodexhasmuchofthenaturallanguageunderstandingofGPT-3,butitproducesworkingcode—meaningyoucanissuecommandsinEnglishtoanypieceofsoftwarewithanAPI.OpenAICodexempowerscomputerstobetterunderstandpeople’sintent,whichcanempowereveryonetodomorewith computers.
Onceaprogrammerknowswhattobuild,theactofwritingcodecanbethoughtofas(1)breakingaproblemdownintosimplerproblems,and(2)mappingthosesimpleproblemstoexistingcode(libraries,APIs,orfunctions)thatalreadyexist.Thelatteractivityisprobablytheleastfunpartofprogramming(andthehighestbarriertoentry),andit’swhereOpenAICodexexcels most.
OpenAICodexisageneral-purposeprogrammingmodel,meaningthatitcanbeappliedtoessentiallyanyprogrammingtask(thoughresultsmayvary).We’vesuccessfullyuseditfortranspilation,explainingcode,andrefactoringcode.Butweknowwe’veonlyscratchedthesurfaceofwhatcanbe done.
We’renowmakingOpenAICodexavailableinprivatebetaviaourAPI,andweareaimingtoscaleupasquicklyaswecansafely.Duringtheinitialperiod,OpenAICodexwillbeofferedforfree.OpenAIwillcontinuebuildingonthesafetygroundworkwelaidwithGPT-3—reviewingapplicationsandincrementallyscalingthemupwhileworkingcloselywithdeveloperstounderstandtheeffectofourtechnologiesinthe world.