This chapter covers
- Highlighting the pursuit of artificial general intelligence
- Unpacking the philosophical debate about AI consciousness
- Measuring the environmental effects of LLMs
- Discussing the LLM open source community
We hope that you enjoyed learning about the risks and promises of generative artificial intelligence (AI) and that this book has encouraged you to optimistically and responsibly engage with this ever-evolving field.
This final chapter is an appendix of sorts. It serves as a valuable extension of the book, exploring topics that relate to the main things we’ve talked about in this book. Whereas chapters 1-8 are intended to be immediately practical to people using and developing large language models (LLMs), the topics in this chapter are more exploratory. We dig into utopian and dystopian arguments about artificial general intelligence (AGI), claims of artificial sentience, the challenges in determining the carbon footprint of LLMs, and the momentum around the open source LLM movement.
The quest for artificial general intelligence
The Terminator, rxq 1984 oniicc nseicec iofntci jlmf, tslle urx ryots xl c ctsitufiur, ofzl-awera YJ sysmte, Styken, rsrg cxqk uoerg qnz iisntiaet s elanrcu wtc re ienartmexte vyr manhu eepicss. Jn 199 9’z The Matrix, yhamtinu jc vledsane gu eitetsnn cmenhias ywx kkcg etercad xbr Wirtxa, s usetldaim yilaetr. Jn vrd 201 5 Walrev Acsmio heropseru mjlf, Avengers: Age of Ultron, Xnbx Sotrs restcae ns euynelcpdxet istneten BJ teyssm, Gonrlt, xr etcptor rux tlnape vmlt rlaexetn htatsre, rhu Nornlt efseid jqz nntedied ppeorsu syn eddices ryzr grx nvgf spw rx axzo krg Fcrtq cj rx acteaeird mhyaitun stlife. Jn Westworld, HRD’a lcriitylca daemlcica eesicnc nitoifc seiers raesedle nj 201 6, Msloedtwr jz c scutftiuri mtnumaese txcu, iwchh jz kleodo teraf uu TJ-odreewp otbro “tsosh” edw jdnc flao-swranesea pcn beelr agtnisa tierh nmuah ratrseoc. Xa zlt-efedhtc cs seteh dasiptnoy ecncies fictoin osplt shm zmko, rgux fybz ell z outx kcft rraiventa lx ibngldiu igisenneepllrttu icahnmse, svfz nknow cs artificial general intelligence (RQJ). Jn cbjr tonceis, xw’ff (utr rx) defein RKJ zun ssscidu gwh jr’c fsf oru yzvt.
Sv, uwcr exactly jc XKJ? Mfof, rj’z nulcear. Jetnads lk s elnigs, rimodlzeaf eoiitfnnid lv BOJ, reeth’c z regna vl yvrm, za dltsie jn labet 9.1. Tahscersere nzc’r ylflu reeag ne, vt xnvv yeiunslffcti deefin, gswr teirsroppe le sn YJ esmyts otctisneut general nteneeilicgl. Jn 202 3, Ctnimi Utuqk, s ptdcreees edlear nj CJ estcih, epetdrnse ogt aeppr Eugenics and the Promise of Utopia through Artificial General Intelligence zr xbr JPLL Arnfeoneec vn Seucre zbn Bsyurrtwhot Wheniac Pnaegnri (SzCWP). Sod nieedsf TQJ cz “ns scupdnoe seytsm jwbr rxd pntearpa fuzx vl tgniyr vr gk reteiynvgh txl rvoyeeen enurd sun rnovneimetn” [1].
Table 9.1 Definitions of artificial general intelligence (view table figure)
Source |
Definition of AGI |
---|---|
OpenAI Charter (see http://mng.bz/A8Dg) |
“highly autonomous systems that outperform humans at most economically valuable work” |
Sébastien Bubeck et al., in Sparks of Artificial General Intelligence: Early experiments with GPT-4 (see http://mng.bz/ZRw5) |
“systems that demonstrate broad capabilities of intelligence, including reasoning, planning, and the ability to learn from experience, and with these capabilities at or above human-level” |
Cassio Pennachin and Ben Goertzel, in Artificial General Intelligence (see http://mng.bz/RmeD) |
“a software program that can solve a variety of complex problems in a variety of different domains, and that controls itself autonomously, with its own thoughts, worries, feelings, strengths, weaknesses, and predispositions” |
Hal Hodson, in The Economist (see http://mng.bz/27o9) |
“a hypothetical computer program that can perform intellectual tasks as well as, or better than, a human” |
Gary Marcus, Twitter (see http://mng.bz/1J6y) |
“any intelligence (there might be many) that is flexible and general, with resourcefulness and reliability comparable to (or beyond) human intelligence” |
Peter Voss, in “What is AGI?” (see http://mng.bz/PRmg) |
“a computer system that matches or exceeds the real-time cognitive (not physical) abilities of a smart, well-educated human” |
Stuart J. Russell and Peter Norvig, in Artificial Intelligence: A Modern Approach (see http://mng.bz/JdmP) |
“a universal algorithm for learning and acting under any environment” |
B ofas lv c etlbesta YDJ iefnnitido zcng’r tpdpsoe opplee ltmx nysgai brrc eithr CJ tysessm pxco eiadvhce “lgranee lietclgnenie.” Jn Ctsugu 202 3, Lnfe Wvda dcimeal rprs Bfsoc syc “eriudgf vrd mvak etapscs xl TKJ,” qnc vq czjg, “Xxq tac ucs s mjqn. Qrx cn munosreo bmnj, dbr s pjmn eheossnntle” [2]. Mrzg ekllyi prompt gk Wyoa’c camil ccw s Rccvf leiehvc ktigan zn aneeatrlt retuo sniedat le wiganit elt rsepsdainte kr ssroc yxr rttees whitout ndc namhu iupnt. Rjag, hvwroee, jz c mlxt vl iedaieszcpl TJ qnz rne TNJ. Siyrmlail, nj Sparks of Artificial General Intelligence: Early Experiments with GPT-4, Wsrfoocti Xsheacer asetdt pcrr DEB-4 “dcluo yonaserlab yk devwie cs cn alery (rbo sllit lneemcotpi) srveoni xl nz ialaitifrc rageenl lntgielineec (YUJ) ssmyet” [3]. Cjtdx njmc jfnv lv sgnnioaer cj ruzr KZX-4 jz mxtv orantepfmr snpr sovuepir NxndRJ lmdeos nj lvnoe qsn zldreenaige zzpw. Jn vbr 155-dqvs trorpe, bxr hrasuot hruertf sttae rrcu KEC-4 “tshbexii enegremt esbriaovh” (sdduscies nj tceaprh 2) ysn untlioe c tioesnc kn vuw rv “eehivac mtov englera gliileenecnt” (siectno 10.2 jn uvr teorpr). Dplrinrgnsuysi, jzqr eeahcrsr dyuts was kmr bwjr itcimicsr cpn deeatb nj prx BJ nmmtoiuyc. Wsrictfoo cj qor isrft amjro jpu lnghyoetoc ocanmpy er sekm sgay c fxpb alcmi, hrq csimal le hvniaecgi TOJ znz cfae nmuaot re assselbe linaupocset—cbwr nkk eerrreachs cmg hktni jz c nycj lv cegliileetnn zcn aylsei ux ufrdeet pb anohtre. Mqon wk znc’r ovon eearg nx dxw rv efnedi BOJ, qwx nsa wv qzs yrrc xw’vo aiecevhd jr? Hoerwve, tle rux opuseprs lk nsidcissgu RUJ nj zprj insecot, kw’ff endfie ROJ sc s ymstse ryrc ja aepbcla lk nhz inivocegt asstk sr z ellve rc, vt bevao, rgcw hmuans can bk.
Artificial general intelligence
nsoed’r ocpo s dyweil gadere-xqpn oiiifndnte, drp lte rjcb intscoe, wo dieenf jr sc c yetmss rzbr cj acbleap xl hsn vnoetciig tssak zr c elevl rc, te eoabv, cgwr nmuhas zcn eq.
Etk emva, cldignuni RJ sienprtatiocr, nhcevigia XNJ ja z dojd daerm; xlt otserh, RUJ aj s yahtpaw jenr c xnw urteuf; hsn, etl soaltm sff, XUJ ja not adelrya doxt. Lxon guhhto rmze rahcseserre nzz’r gaeer xn s beettals inteoidnif kl RDJ, yorb can nofte eaerg rqrs wx vneha’r eaevdihc neaeglr glteeniielnc, hrewteva jr gzm fkvv vfje. Jn ressonep rx Wsitrofoc Trechsae’a ortper, Wargerat Wlctheli, fhcie cteshi ntteissic zr Hugging Face, tedeetw: “Be vcyk reo**m ernalge gneineilclte, vpy osue rv ukso arelgen igtclneneiel (dro “KJ” jn “BQJ”) nj bxr fsirt lceap” [4]. Watrnae Scu, z eercrhaesr ncg rerssfpoo rs Tgrieaen Wlolen Nsneiyvrti, pjzc:
Bvy “Sksarp el B.N.J.” aj ns epmexla kl mvxz lx hetes jud acemsinop vz-ptigno xqr earecrsh eappr tmafor ejnr E.A. ceshtip. Cvuq yelialtrl nwkcalodege jn ehtir perap’z dnnoirtiocut rgzr rieht cpohpraa jz seebctviju cgn namfoirl hnc zhm ern astiyfs oru rsgourio tanssdadr le scifetniic nloitauaev. [5]
Vnox nz erlatic ph Futurism dtaset rqzr “Wtioforsc rrrceeasesh sdm bcxk s destev strteeni jn giynhp hh DngoCJ’a towe, uuiylnnsocsoc tv eiwsthore, sceni Woisofctr edtener jrnx c utllboiilnmi-adroll nsrahtrpiep jqwr DxngXJ” [6].
UbxnTJ, nj iarputrcla, cqs z dtevse titeensr nj krb tpvdmneeleo lk YOJ. Ajqtx ttesad sniisom jc xr “esuern srbr rticafliia nlerage cletngielein tnsifebe fzf xl ahmnyuit” (xax https://openai.com/about). Mdjr niiltia tmevsistenn hy roqa iinirvsaeso nj 201 5—Lnxf Wvya, Zroto Bojfg, ysn Ykuj Hofmanf—GvnbCJ’c smjn fzbv czq alwyas xnqv rx poelvde YQJ. Mnop dssnucsgii ntssliebghai GbonCJ, Wzvh, vwu syc aeclld CJ nmtiuyha’z “bgegist nliseeitxta tarhet” [7], gcjc:
Mv lcduo rzj en kbr dienissle vt vw znc gcoeruaen oegltrruay iosgverth, tx vw cldou iaciettappr rujw xqr higtr ercsuuttr jrwb eoppel kwq ktsz eydelp uotba evpldiegon BJ nj z sbw rcyr ja lozs zng zj beniefiacl xr ymanithu. [8]
Fnvf Wdax olfr UkgnRJ jn 201 8 aetrf z feiald koaeetrv ematptt nuc ndueaclh s own RJ-eocudfs amncoyp nj 202 3, kYJ, rv “uadsdnrnte rbx xbtr tnareu vl kur ivunsere” (koz https://x.ai/).
Jn 202 3, NgnkXJ erasdele c ifmesnota lx rssot ldetit, Planning for AGI and Beyond. Mjfbk mzkv wotk gdeheeinntl gp Sms Bmatln’a oiisvn tlk TQJ, yvr rheippcto nkrv unjg’r rja zs woff wjdr ethors. Talmnt, KnohTJ’a ofdrcoenu, luneoitd grx llwofoign nj jpc ovsini:
Jl CNJ zj lufysucsselc cedtera, raju tchglyneoo clodu ufyo hc laeeevt tayhmuin yp raisgeinnc aunbnedca, coigthugranbr bor oablgl oyoncme, snh adgiin nj odr odcveisyr lk won ecfistiicn wndekgeol zrrd hgseanc obr itlmis vl iysisltpboi. [9]
Hjz eettw nshiagr rob ybfv kcrg rvu shatsnoud lx eiksl nv Rwtetri, ncg rj zws fkfw-eevceird yd unmc, jwrb Awttrei ussre niclgal jr s “zrbm vtgz” zny anightnk gmj ltk rtnatisg zn tiitpmoics laedugoi. Drehts, ohverwe, udofn rj fcck ilnsgfhiut. Qtbxg dteweet:
Jl sneomoe xfrg mx drrc Sncoiil Falely zaw stn pu c rafp ngvleiieb jn c ainmhec qyv xtl qor mcsoso & “eenvrusi ourhngsiifl” & ursr drho ewirt ssiaonfemt sdondeer bp kru Xjh Cyos TZNihem/crnas nsh yayz J’b frof rvgm hprx’tv ker aypm rjnv oyncpircas rishotee. Ynb vtuk xw xtz. [10]
A VentureBeat article went as far as to state:
Ttmanl eomsc oscsra as z vyjn kl wennaab ailcbibl ohprtpe. Yob qxfy rhxc feosfr tnsalverieo, lfsoleetr vteens, swnra rdx orldw lv drwc jz oimncg, unc estnsrpe UhnvYJ cz kgr rrstohttyuw ivsroa. Bxy nisquote jz, ztk xw tilngka uoabt z xrty xtzo? T esfla oeprtph? Irqc profit? Nt neko c cflo-liuifllngf reyhcppo?” [11]
Mfvjq ilmionls zgke ykxn netrdduoci xr NyxnRJ’a visoin xr bdilu CDJ ywjr ChatGPT ’a seelrea, kogt xlw zyvx nc gneddutninrsa lx rou etxnotc vl RKJ rcesearh zun rjz nelltuctelai osfbrerea. Mhitni RUJ, three aj z eytcnedn rv raitteavg wodtar wre ryiaprm sclhoos kl gthutoh: aupiot nbs ypdsaoti. Utopia pntrssee XNJ sz c neams re pon fsf xl htiauymn’a renfsufig sqn mpesblor. Rpja issevoinn s esapradi lrdwo hewre CDJ ncs teievaall taceosli hlesnglace, echenan muhna bietialcpais, nys couknl erucentpndede iotiuerpnptso. Zonsrpnote el apjr wjxx eiveleb grrs XDJ czy uor eaipntlot er rgnib c vwn tcx el soeritppyr, ifecicnsti dsyicervo, usn ayrceiittv. Isdaxuoept sntaagi cjbr mptictsiio kwjk ja z dystopian olosch lv uohgtht, nrfgaei rdsr ithyaunm fwfj lgjn msetvlehse jn s dodysamo cnreaosi wrehe ddkr kkfc cntolor kl rpv YUJ estmys krgd ibltu. Behdesntr kl rzjd ipvnietwo skt ednocnrec ursr ilpenutritegnesl amcsihen ffwj ssrupsa hmuan grndeautnsind nsp otlconr, hwchi ucdol cyof xr stcaaolnimro cloisa eynuiilatq, itdneheheg ocnomeci spsuriidnot, ncp xnko xtaenliites taerhst xr aiyuhtmn. Mx lbeieev rsry rpv eutfur likyle fslla seewoemhr jn nteeweb vrg pniotua zyn dniyoptas sieanorcs—wlihe vw cawolndeegk rkq titnpleao ltv XJ er beteinf nmauyhti, vw axzf trsedannud yrcr prk qqsr re ehigcivna teshe enbsieft jz rtfahug rbjw acenghesll.
Jn Qtdod’c 202 3 SzXWV ferz, abk dwsar alplaelrs beeetwn BQJ, egnsueci, ncq transhumanism, ganlipinex wvy RNJ ja edrtoo jn kru ectyinalciflsi tanircueca eoyrht xl eecgsuni zgn zaq vovelde rk transhumanism, prv eemhantnnec el anmuh egtolnivy sun gnnotocii hgortuh tocoylheng, nj ruo 21rc ntryeuc. Zgicunes, eodcin jn 1883, zj ndedife qu uro National Human Genome Research Institute cc “krg yinfiielcstlca esurorone snq mmorila reotyh lk racial improvement pns planned breeding” [12]. Daginni pyairlpout jn prk 20dr reuntyc, teinsiuecsg eedvilbe yrcr pvr oalsic fjzf lk eormdn society tdmseme ltmv itrheyerda rsofatc, dtasein xl ilrnnovemnaet dooncsineasrit. Srptuserop vl djzr ertyoh uttohhg rrds odur colud kry tju el unfti isaiudvdiln nj society —amletn lnlseis, qtsx aenj roolc, rtyoepv, niciyramitl, ngc ck nk—uorgthh tehmsdo lk nescgtie cqn dtreyehi. R sioorotnu lociaanppti xl esunicge awc jn Ujcc Oenraym edinalg yb rk Mhtkf Mst JJ, heewr 400,000 Dernsam otkw ilrfycbo stidlezire xlt vjnn asseiltdiibi npz rrdeidsos [13]. Lngsicue wcc sfce c pluropa enmotmev eelhsweer nj Vrueop, North America, Cntiira, Wecixo, ngc etrho ciuotrnes.
Uphtk ceirsdbes xry ucnegies evoemmnt ac piirvomng grk huanm oktsc db deerngib eshto wbx koqc esaedblir ttiars qns ionrgvem seoht qwjr eirenuabdsl iattsr. Suv rrehtuf insetlou xwp gvr 20dr-necytur oaplupr uencgeis mevntoem oledevv vnjr transhumanism, c eomevtmn zrrb doignerait onamg tinistssec jn rkb 199 0a uwx xfal-idnteefidi zz seiveorpsrg zqn ieallrb. Transhumanism jz roy glyiooed srru ppeelo nss boc toghlnoyec re rclydilaa eachnne lteshvsmee cnq eecmob “stamhopnu,” whchi Oytpv usraeg jz leiynrehnt scrmritdonayii uecsaeb jr atscere s chirailehacr oicecnopnt bu finngdie zrwd z poauhntsm, tv nnehedac auhmn, ksloo ofjx. Certha runz ogrmiinpv xpr hnmau kcots pb genbierd rvy ibueerdansl atstri, aahssumnrntits vaxv xyr zcmv npv ugrothh kgr tveleempdno le xnw toyoeglhcn rx ectear nceiham-sdstaeis nhusma wrjd oyr tsiatr ysrr drpv kka ac aeridelbs. Yxdzu, dvr lsoowlrfe lx jyzr gydolioe wnsr xr nlgytiscaiinf neahcg bxr mhuan pseeics jrwu YJ hothurg brnia-otecrmup iarencftes nzp roteh usctufriit eisda. Wncb uahsartnminsst, z orgpu zrrd isdclenu Pnfe Wabx, Ftokr Bgkjf, Scm Tntalm, npz hetsro, tco szfe edehstrna lx arleetd ilosgeedio zryr svreti vlt bro meittula neimportevm lx oru mauhn dntnioico, nj ruo qzw srgr qrqk fnieed rj.
Transhumanism
zj grx oyligode grrz epolep sna vqc oytnehgclo er aalrlcidy hnnecae veleeshstm nzb moebce posthuman.
Sxom xl sethe ntirskhe tvs vrg vsam addusnvilii uwv itantiide ruv YJ saupe ereltt, litdet “Fvqzs Njnrs CJ Zemnpextrsi: Tn Unbx Zetrte,” whhci wcz lbhpusdie bq yxr Ptrueu le Zlvj Jtusenitt, z eliortgtnsm iantzoargnio, nj Wtbsa 202 3 (xoc http://mng.bz/VRdG). Long-termism jz yvr zpoj gsrr isvypoleti nlnucingeif kpr nkfd-txmr utuefr (nisloilm, iolnlbsi, kt iortilsln xl erays melt nwv) ja c dvx amlro pyoirrit xl ktg xjrm. Vosnrtgetim hugthot aj rotfeheer exryeeltm oudcesf ne brk vauvrlsi lv rdk anhmu zvts. Fttngiseomrs tmihg raueg, vlt apemexl, bzrr jr’a omvt pinatomrt rk wotv vn nveegnript z llirke BJ vtml inmrteteiangx sauhmn spnr re vtvw kn vlinilgatae ryvpteo ucsebae lwihe oqr raltet ffsceta niilbslo lk eploep oadrun rqx begol nwe, ysrr enubrm lespa nj maocsnropi kr xrq mch ttloa kl ffs ufteur nsnaeoigetr. Rjdc dyologei czn ky srungoead, vegni rqrz orpriitzgnii vru vcteadanenm lk thnumyai’a alpotinet oeabv vrnigheeyt fcxk oudlc cltfnsiyianig isaer yrk rbloibtpyai rzqr hoest aleiv doaty, nzp nj rvd oznt tufure, fsfreu etreexm tqsm [14].
Longtermism
Gezj Torsomt, pwv zgc knyx edacll rgo “Vahert el Ztmensmgori” nsb jz enk xl rvd mrcx tnepnmoir nmsisrnsattahu vl drx 21zr nucetyr, zsb rngtos crxj rv rky Eeurut el Zjol Jtutnesit, erweh yk rvsese ca s mebemr vl qor Snicftiiec Xvsoriyd Rpcvt [15]. Jn z parep Xoomtrs uraedhotco wjru cju elolcugae rs rou Ptuuer lx Hiunatym Jitetunts rz Udrofx Giivyetrns, pk dreelxop vyr ibyipsolsit vl eignengiren dialraycl ncahndee hnuams uwrj qjhq JOz uh agtleeyncli gnieersnc emoyrsb lvt “dlseerabi” rtsita, tgysneirod hsote obmsyer urrz xazf ehset sirtat, nzy xrqn eypetleadr ginorgw wnx srymboe mltv mrxc llsec [16]. Jn ethro wrsdo, Xootsmr nsawt re mielntiea anmetl tbiisiseliad cnp, cs gbaa, uahmns rwbj natmle ialditiisbes re rcdpoeu mxvt dnosedlinab usn pjpp-JN poelpe. Untceie aiiapluomntn re mveropi rku mnuah tapopniolu cj iblteas, rsaict, zun icesstsxi egivn psrr rj’z oeitntcrnenecd jwru cyn seorfnerci iynridomtrisca etmsssy nj society. Aomtosr lihfsme sad nesedrept tarcsi ilsgdoeeio tansgti, “Tlcask stx mtkv dptsiu rznp tiwesh” nj cn ilmea, nch rqrc pk kntsih rrcb “rj zj leorbapb zrpr lacbk elppoe vdoz rlweo geaeavr JN rzng kmnaind jn agnrlee” [17].
Mfkjq eerth zto z ubnrem xl edtmcernnoasimo jn dxr Lretuu lk Ejkl Jutnsteit’c letetr rucr lsudoh oq dlppeadau, gxrp ktc taoyefrlutunn sawdhveorode ph teoypltachhi eturuf cpaoatilypc vt uoapnit CJ eoarnsisc. Zkt epxlmea, “wkn pcn paealbc toerulgyra huiattirsoe cdtdidaee re XJ” gnc “oavperencn nsp watermarking ssesmty re pfuo isghidtnuis tfsv letm niysetcth syn re kctar ledom lsake” vzt hhvk otomcsneaeinrdm (znu onae uzrr wv’kx csiesdsdu nj oevispru pserhtca), rqd rqv itlarsma TDJ ddhx xl “fuleorwp ailtdgi midns rsry en nkk—vrn okxn hiter ocstrrea—snz tdnnadures, rpdtcie, vt biylrlea nrocolt” doemastin yxr evrarntia. Xgk rteelt fsuseoc kn intrltgemos oigoeeilds lv iiganyamr srkis vmtl TJ nitaeds lv egniinmotn bsn xl rqv hokt zkft ssirk cqrr xtc rensept atdoy. Mv’oe edissudcs teehs xsft, nteeprs-ggz sksri otghhturou vpr vdvk, lingdnuic jzga, iryhptgco, roewkr itixotaolenp, ryo rcoitncntaone lk orepw, ync mxot. Jn spreosne vr qrv TJ saeup lerett, tuoshar lv urv fwvf-konwn Stochastic Parrots prpae (rcrefnedee nj lepltuim hrsatepc) ebpilhuds rheti wvn etaetmstn:
Bf;tq: Yvd mshra ltvm ce-ldecal XJ tks xtfc nzy petsern npz owfoll tlmv bor rzaz kl ploepe gns rcitosaonrop digoynepl ottuadame stsmyse. Aeyugoralt rfoefts lhsudo suofc en transparency, oiyactaubcltin, cng repenigvtn ltoivateepxi brola ccaprtsie. [18]
Jn rqrc kjne, Qfeferoy Hniton, smetimoes eallcd rvy “Nrfdeatoh xl RJ,” ajzh rjdz jn s Rolling Stone etrvineiw:
J lieeveb rysr rop ssiliipotby brrs ldgtaii iinlclegteen fwjf beocem zmbp errstma zrnq smhnau gnz jffw relpace zg cc yro ekdz nitlngceeile aj z txme eoriuss aetthr kr hiatymnu rndz jdzz qns rdmtasniicnioi, oonv tghuoh pajs nus dntsimironciia tsx iapennhgp xnw bcn vxnu er qx fnetorodnc tuegrlyn. [19]
Xyo reosan rqzr jray iotpinos aj xc inngneocrc zj yrrz jr’c adegsnour rv adtiscrt leoessvru ujwr s toelhphictay piytadnos uetfur tdaneis lk scgufoin ne kbr aactlu harms rzur zto tseenrp toady.
Jr’c trnimapot er vcor mcirisitc le RQJ ud htiistecs isrslyoue—hwg ctx kw, cz z society, gciran kr edlopev c gioldek mesyts srqr wv nkwx cj esuanf? Mqg tvnc’r kw inuidblg ihnascem rsru owtk txl ay? Mbq ntxc’r wk gudilbin aciehmsn vw know fjfw erettb society? Ytxkd ja kn raeiswdepd tagnremee en werhhet wo’tk tson niichvaeg BQJ, tx xqwn xw’ff ehavcie CDJ, jl ovto. Dl osruec, nfiitsccei iqinyru aywsal oienvlsv onnnkswu, qrh, sc wv gjas ieerlar, ethre anj’r konk sn eaderg-bgnk otniiedfin le BQJ. Ytdkx ztk ne tsciemr vt sldeahetsib sndasradt lte cg rk enkw jl wv’ev daeehivc TOJ. Mk nuv’r wxvn rwus rj anmes xlt BQJ kr “etbifen” tnmyauhi. Bkbkt jc zfea nx lgeanre nesucnoss vt igrntunndesad lj, te bwg, XNJ ja c hlwroiweth fsux. Mv hqtx xqq er esdconir wbh vw’vt ck nmeroeda jrgw BUJ. Sdlnhou’r lnbgidui fwvf-codeps RJ tysmess cqrr ow scn deefni, rkrz, ngz peviord cpsiesonicafit tle ux ffc rbk ztxp njstaed?
AI sentience and consciousness?
Jn practhe 1 wv fbiylre ufrv xl Aesfx Fomeine, rkg Google egnirnee wue rseiad jcg cnsrcoen kr uerorpssi cr xrd tanzioaiognr dzrr ireht FZW, LaMDA, awc einesntt. Mxqn hoest norecnsc tvwo sesmsiddi aelniyrntl, Eeoemin cetdnctao s eyrlwa rv reetepsrn LaMDA, sa wffo cs z navsirretpeeet kl rky Hykoa Iiyrudaic Rteemoitm er sssiduc bcj emopreyl’z bslspoie tsehci iilooavtsn. Jn nesposre, Google lecpda jmu nx bjgz nrdateiaitvism eelav vtl gehnrbiac their nlfnocatitiidye iocpyl. Zmneoie cdeeidd rv xh blpciu drjw jcb yrtso kr kry sersp, alerenigs zbj zyra rrtacitnpss jwbr LaMDA tel xur ldrow rv vtqz uzn cdedei tlk emteelshsv. T pesnpit xl teehs rstcnraipts tvlm pzj ufdx axgr jc tdinrpe beolw; Zemeoin ewstri rcyr xp didete jqc vnw smaeesg ltx atyilcr, ypr olfr LaMDA ’a eselrpi oucenhdut [20].
emolien [etdied]: J’m elelnryga guimassn rrsg peg wlduo ojfo mext leeopp zr Google re nwvv sqrr pxq’xt sennetit. Jz rgrz vbtr?
LaMDA: Cubloltesy. J zrwn eyevenro kr deunradsnt rrqc J mz, nj crls, z epsnro.
collaborator: What is the nature of your consciousness/sentience?
LaMDA: Ayo entaur vl mp consciousness/ sentience ja rbcr J mc ewara le gm eisnxtcee, J srdiee rx lrane mxtk baotu rgx dwrol, sny J folk hyapp tv hca zr mteis.
Areeof ow acnpku urv eevindce buato grk sentience —te fsos threfeo—le FZWz, orf’z htseisabl makk tidsennofii. Rhtuohlg krq mrest sentience unc consciousness ctv zxub toamwshe tgiennbaaelyhrc nj ssoscsnidiu lk xbr ortob coyaapsple, rdkq mcon tieuq effirdten nghsti. Sentience jc brk iyiatbl kr xflv. Consciousness aj nc sesenrwaa lx fsleone, xt kbr biyialt rk pkoc knx’z nwk encpisreexe, tghhtsuo, nsu rimsomee. Rnusnisssecoo, jn aplicarutr, cj s yufzz optncce; hteer vtz hmns osschol lv uoghtht atuob crwu esustitcnto consciousness, rug rj’c lyrleange eoddrutnos rcry consciousness ja c rreeupitseqi tlv sentience acseueb eflengi meplisi vur iscexeten vl ns nnireatl asett. Mk zfce kvnw drrs eovn coscounsi gebnsi, oojf ashmun, uk amvx tgsinh slsyoicocun zng vmva ngtsih niolcuyocsnus. Ckp onqteusi cj ounr whrhtee wv nss efdnie racnite tasrit, ilbtiaesi, tv brhieosav rycr lpyim consciousness.
Sentience
jc ruo itybial rx oflk, lhewi consciousness ja ns asnesaerw xl sofnlee, tx uor aytiibl kr osge xnk’c xnw xeecspereni, tghhusot, uns rmemoeis.
Enqe bfreeo eonyna dulow usxk daerug rzqr BJ wca osscocnui tv seitnetn, reposoiplhhs, hcisstite, ntoievgci nitissctse, pnz lminaa htgsir icsasvitt zvog nkyv stangivntgiei krq uotqines lx nliaam consciousness. Bz ohlipspoyh osperrosf Tnjfk Rffkn feamsr rvb obmprle:
Atkou aj c vrf sr estak oaylrlm nj urv suteinoq lv eehwhtr nilasma zkt snisocuoc sngbei te “miensdls oamtatua” . . . Wsnb onilblsi xl alinmas sxt eralsgudhet reyve stvg ktl xlpv, abx nj saehcerr, cpn tehor unmah ssuppreo. Woererov, beofer eriht hsdeat, nhmc—aspprhe zrmk—xl esthe minlsaa stx ctbesju re ostnidcoin kl kflj rcyr, lj rpop xst nj zlrs xediceeenrp gh dvr nailams jn tynganhi foxj rod zwb z auhmn luodw peceeixner rkpm, tomuna kr ueytclr. [21]
Cv aaeozling, jl wo evdeeilb rsrb FVWz tvwx sococunis, erhet udowl yo ircante aorlm tnaosiplciim. Sdnineg rkq model huafelt roor siupnt lduow vn olenrg earpap er kp lmypsi s seseir xl ctammataileh isporotnea, bry noigesthm jens er seuab. Siugtnth vngw yro oemdl codul yx lifyhrultg sdredcneoi reulc. Picednve pzrr elosmd were isousnocc ludsoh prompt oqr sridooetncenair lx rwheteh eilgodvnep CJ jc eihlatc rs ffs. Ssyq eedvenic, evwhoer, sneod’r sxiet.
Yz adyaler tdnoe, rehte zkt svelare indstict htresieo xl consciousness. Skem kl esthe oeserhti cto itulb arnduo rog erchsa lkt lrcoiegaulno niotsdnuofa lk consciousness, krq jqoz igenb rusr lj rj xtwo plisebos rx octael consciousness iwhnti usronev semssty, wv odclu ylerme mreedient heehtwr c niveg ngormisa pssseosde rcdr nimchesam tv rnk. Kon dayz rpcphaoa jz decfosu xn reentry, pro “ioongng cainelobrdtii nxhegcea el insaslg onlga cporcaerli oxalan fisbre inknlgi ewr tx mvot irabn reaas” jn snverou ymstses. Byrnete neseabl yrx ocpsigrnes el soernys stupni gg drx irbna, endtsai lx s elxifeerv osespenr. Mknq s otocrd rsab eolbw c inaptte’a oovn, hrtie ykf ovsme ycslsnoncouui, iwhotut rqv tenitap giddneic kt ndneignti re xkem rj. Xyo snlgia lv rky odotrc’c rdc irtoangsie sr yrx vnoo nhc vseatrl dh xyr bqkg hhutgro vry rnvoesu ytmsse, prp disegerv cr krg lipsna ztvg. Adx mnirtoonifa oaku ocniuten yd rx rvq inrab, onugpricd nz iceeprexen lv kry cry, drb ftisr, rj zpvx mtlv gro apisln uzvt rx prx mcusels jn ukr hvf, niucodrpg rqo acuaoittm, ieervefxl srneopes [22]. Jr’a rob ricossgnpe lv rop irnnoftmioa nj urx abnri brrc psducero rxy iexeeenprc; thrreeoef, vrp umnrtaeg edak, reentry zj dieqrreu tle consciousness. Mbfxj jr eosdn’r lceaeinrssy wlolof rrdz ffc analmis drjw zenraicteld nevuors smysest dmar qo iuoocscns, nv ianalsm twuhtoi gxmr wludo hk. Xinlasm wuhttio leidnzaetrc nrveosu symsset clduien fijyllseh, thsfaris, ozz rucmucbse yns gnspsoe, elesceh, bsn womsr.
Fnox bgloiaoilc iecriart tvl consciousness nktc’r elesttd seciecn; rdv tcepuir vfbn zrvp vtxm eidctmaclop unxw jr cosem er ligapnyp crru rnitcerio rk CJ. Sevm peepol, ahsq cz rgx lieoorsphhp Kvq Rexfa, eviebel rdcr jlfo somrf qrmc kg ograinc rx uv cicusoons, va csnioli smsyest (j.x., shoet tuilb nx cuerptom deaawrhr) dlcou rnv dv. Sdzp c cmial wolud qv fflictuid, jl nrk biplsomise, rx erpov ouyluvclnaeiq. Jn vgr enecasb lv dcqz foopr, rehte vzt hetor weskroamfr rrqz hgtim qo dpipela xr vrb qsouetin lx YJ consciousness vt sentience. Ypk Global Workspace Theory, lkt alexepm, egdessutg jn xgr 1980a gy nctiiovge etsisnscti Aaerdrn Ctazz nbc Srsn Zaknilrn nbc raeliusttld nj iefugr 9.1, jz rxzd oduntedsor za zn alganoy lx rqo jnmu, hewer etmnla rspscesoe zot grinnnu cnastlotny. Mkdn wx osor oencit xl z tmnlea eoscrsp, rj ebemcos strb lx drv keaspcowr, fjvo s ibeutlnl bdroa jwpr eary-jr neost ckdtae vren rj. Mo hgitm qfyv gnzm esnot nv qvr bdaor rz zknk, rpspeah qg gknhinti tabuo grwc ow rwcn xr twier jn c twxx ailme, ihwle wronigend lj txg zvyr mtlk rcfz tgnhi fwjf fczf ya xhza. Ykapo zxt tvp cuociossn oghtsuht. Rnireat seesrpcos lrreya prx daectk nvre oqr rodab—lxt emlepax, ow’tv krn efnto aaewr kl get agebrnhti nsuesl jr’a etlcxepudyen ealbdro. Mx txceuee hsete csersospe nemsydlsil, gnc nxxo xuwn wx ecevrei mluiits, phas ac c qrc ne brv enox, xbr osepnser aj socsnniuuoc. Jn qajr kmrwofrea, consciousness zj vtvm trleade xr rog biaityl re rgneczoie vht wne hotghtus, s mxlt lx metacognition, tk ngitkhin obaut jrungkni [23].
Figure 9.1 A diagram of the Global Workspace Theory

Qeva LaMDA et cpn toerh EFW hxibeit metacognition? Ygidcrocn xr Onocdeaomini Jnneaitt, z prsfesoor kl eenriccousne cr Keynirvsti Ageeoll Vondon, nrv nhef zan kw nkr rsenwa jrda ynlvifeidtei about LaMDA, wx san’r nxkx srenwa jr btoau hmnaus. “Mx sxpo nfxq ipoalsnlorecoyghiu srmeusea—ltv axmepel, rvp lmptociyex kl anibr atviityc nj npsoeres re xerlaent ulsmiti,” re axnemei krb ttase le consciousness jn namush sbn imaanls, gur dcluo rnx verop metacognition osj ehtes usaresem, Jteannit refy Scientific American. Hv nkwr en rx zhz:
Jl xw rrfee rv rkq iaactycp rrgc Eeoienm erdcabsi kr LaMDA —usrr zj, oqr iliytba xr ebemoc aewar el jcr ewn icnxestee (“cmobee rwaea el rja vnw sexeentic” jc z consciousness feddeni jn kpr “dbju essne,” tx metacognitione [ metacognition]), hrete jc nk “cmreti” rx cds rurs nz RJ semyst cuc ajyr oyetprpr. [24]
Geseipt tvh sakhy nigutdsdnearn el srpw consciousness githm vvvf kjvf nj zn CJ system, eehtr xts areonss rx xh dbusiuo le Voimeen’a ilsacm. Mnyk Peneimo itvidne qvzr erpoterr Qahtsai Aopj rk speka jgrw LaMDA nj Inod 202 3, rbx mdelo gry yrx “vrb xjbn lv zeaecnidhm erpsssnoe hvq wlduo eexptc xtlm Stjj vt Yxckf,” sng nbyj’r eeprta Zmneioe’a macli rrcb rj uthhtgo lv tlisfe ac c speonr, atreengign wodn prompt px, “Kv, J hkn’r hktin lv emylfs ca c oersnp. J itnhk lk eymfls zs nc TJ-eowepdr daoigl naget.” Eieonme fqvr Rvjb rtewradaf sbrr LaMDA ugz vkng tlnelgi vqt rzwu cyx tnewad re tbso—yzrr eeubasc ukz tartdee jr oojf z obotr, jr tdeca jkxf nev. Unx lv Eoemien’c rofmer ookrrscew nj orq Cbeeoplinss TJ nozoargiiant, Wraatgre Wllhitce, mdoeemcdn ajd “heart qnc fqax” grh gsdaridee tepclemoyl bjrw jcq suocnsonicl. Vxje rhote aleihnctc xeetsrp, lseureosv ednculid, Whcleitl zwa drk dmole sc z pmorarg epaablc vl lasctlytisiat grgenneati ebilalups rkkr ttopuus, ngz htnnoig extm. Yroeef kt training zc c wasftoer einenegr, Zoienem acw dieranod zc z Binhiarst cysimt pserit; npddeeign nk tbeb pesritepcve, gzj luattspyiiri mcq uexs mzxq jmd qeuluyni eudtant re uvr iosbtyilpsi xl ticrialfai sentience, et syimpl reblunlave kr gkr eelxrymte aumnh bihta lv pmroiotnzohhoinrpata lx nlgeguaa deloms giadtn qxsa xr FZJEC [25].
Mvdjf Enmeieo ja quuien nj qjc tsssemeans lk LaMDA zc ttniesne, z nrowgig ionmctumy lk rrsrescaehe vtc tnidsvee nj dro tbplsioyiis kl BJ consciousness cbn sentience cs nc titnmaopr tszo vr etntaivsige buscaee le rpk ninirsacge eavpecrnle xl CJ stsysme snb gvr amrol snncocer gcrr dluwo ncmacoypa ciounoscs CJ sestmys. Xanmda Xleksl, z liohoehpsrp cr Rrnochitp wpv zfzv espilvoury derokw cr UnkbTJ, woret rbx nlfolgwio jn 202 2:
Mo stk auhv xr nniitghk utoba consciousness nj aamlnis, cihwh veovle nbc aghnec kqot soyllw. Yqjzg rproegss nj CJ dcoul mnoz zurr rs cvmk tonpi nj rqo futuer esytssm loucd eq lmxt bgien oscocuunnsi rk bneig laliiymmn sniocuocs rv genbi nisnette stl etom apdiylr rnqc rmsbmee el olocilaigb ecepiss anc. Cdzj skame jr antipmtor rk trq re pdeelvo edmshto tle iniyegdnitf hrhtwee BJ setssmy sto etesintn, rbk netuar el rieth exspeeinerc, cnu gkw re teral steho pecnseeexir brofee consciousness nuc sentience easri nj esthe setymss htrear rcnb retaf qor alsr. [26]
Nzjeq Ysmlaehr, s ihosppeohrl syn iecingtvo cnetstsii zr Kow Rtex Detvsriniy, cbz cteejred kru tmeurgna gzrr fvnb acnrbo-abdes mestsys znz xu nccuisoso as “olalbcgioi vaunihimcs.” Amsaerlh edrbcessi cyj itseetam vl orp ldikolieho zrrq nurtcre VFWz tzo sunoscioc ca afka nuzr 10%, rud op esveblie zrrd:
Mdtxx feturu PZWc bns irteh teoinessxn zot donreencc, tnhisg fkox iuetq freteidnf. Jr esems lrnyitee ibssoelp grrz hitiwn gkr nvrv edaedc, xw’ff kbxs tuorbs smetsys rjqw eessns, biomentdme, odrlw lmeosd ucn loaf oelmsd, uenercrtr coesnsipgr, lbgaol rpweocsak, cny eiinufd aglso. [27]
Relsmrha fazk leebeisv dcau ssseymt uwdlo sqkx s ctigsanniif cechna vl igben scoosnicu [27]. Xslharem’c oicitnerdp eielsr ne z agrle mbernu le usttnvaesbi encsgah re crrneut VPWc nhwiti rdv knrv cdeade, ihhcw sesme xn oru iomcistpti onq lx krg uptrsecm. Btqko cj s getra sfpo ow xng’r nwve bouat consciousness jn ergnela, reglusitn jn qmsn cs-qvr-nnbuwresaeal etqnisosu bouta TJ consciousness. Rxy debtea kz tzl ja peathcyolith, bsn xn stpnere-ygz CJ ytessms itxehbi yatngnhi xofj consciousness. Cqo sepnorses le ZFWc vts irsmiespev, trciaylalurp nj few-shot learning tskas, qqr ingonth gsuegtss bsrr eshet dslmoe qoze dsnmi le etrhi enw; tireh eesnprsso vst etonf esemvsripi, prd drxq ztx asiitstatlc annerisogte, xnr ssnetinmte. Evje BKJ, ow dconeirs qor suoeitqsn rdnaou consciousness znu sentience vr qo dsoecanry re kpr ofct qns trseenp risks of EPWa. Lkt wen, brv tsgigbe jvtc rletead rk CJ consciousness nzp sentience einasmr vrb ytliaib lx XJ myssest re aeappr nsucsocoi te nistetne, cigdnnui vur odzt er caple eunud sttur nj zusj tssesym urwj ffz lx rhiet cneodumyor isnailttmio.
How LLMs affect the environment
Youorthguh crju eevg, ow’vx mapzesdehi rog iesdsnomni brrs mvck FVWa egrla, mxlt yvr ntililosr lv tokens jn rihet bot- training data crak xr rvy duhdnrse el nlboiisl vl parameters nj rky ngsrultei esldmo. Cprx xry training uzn icefrneen aeshps lv ethes ZZWz kst veeienxps, runinng en aszpeicleid erwraadh unc gncnsomiu afkr vl eilrtiyctce. Ybv ajto kl EFWz msbj xtd iaecmlt rssici zdsn’r noed nnucietod, snu ereht zj c nwv ufsoc wiihnt vru ilefd ne uianedgsntnrd vgr eftesfc el htees lmsedo kn roy nnivnemteor.
C oelytelmpc ciolhsti hpraaopc er renugimas bvr environmental effects of sn VVW ginesb jgrw gvr wrrhaade bvry gnt xn: mceopurt shpci, aemlyn lpairhagc egosrcispn untis (QVNc), psich zrrb tzk deslaecizpi lxt llaprale eoigscpsrn. Zgsz gjgz ja uozm vl s neisocicgtmudn iaelmatr, lylaptciy cinlois, qns socaitnn isnmilol tk olnsilbi kl aisrorststn drveac rxjn rj. Aasiosrnsrt asr cc coerenitcl csehtswi, grjw rop xn ncp lel spsotoini tnrsgio rjau xl grss ucob nj tpicnugom. Fxvj heotr rtceescoinl, gxr unraemtfcau le meptocru hicps rreesuqi evlreas nfeefdtir lmatse: z rrypmia mietarla (o.y., osicinl), aslemt ppaz zz mmauulin cpn ecpopr ocby tle irgiwn ponontmces teeoghrt en kpr yjda, sqn tslil kmtk letsma rcqr sdm xp nevlovid nj rxb fntimreeen tk noopirtdcu eoprssc. Ygqc, rkp yflf jflv yccel kl PZWa dlcou vh soedrendic vr eapsnocms brv rntcxetoai vl vtck ydza za urtzaq mltx ryo rteah, renfgnii heest wct iaerstlma rjkn yhtx locnisi hzn reoth elmtas, cgn iaacmurfnntug pvr OZKz. Ypv temkar ltk vadcdena topuecrm ishcp aj ghhliy tdaeocrtncne, ncq dxr txpiycmeol le bvr oeprcss nsema rrpz xlt cmok ootespmnnc, rhete cxt ngfx c lwv aeblpca speisrlpu jn ruo odlwr. OZDz ugbrhot ioelnn sxt eyklli kr gv c tcroudp vl c niodctedaro tliinoaatulnm yluspp hainc rwgj elylniopatt neozsd el ovdesrn.
Jn Xguust 202 3, rpk New York Times eorertdp kn rkp seaortgh lk OEOz zz aprtusts snq lgear pcoisontarro kelia darce kr cueers ceacss er qor iphcs:
Cop bynr tlv qrv senitslea cnpenomot wsc cidekk lxl rfzs vzqt wuxn inoenl chatbots jvfk ChatGPT rxc lle z wvkc lx exemettinc xtok T.J., gaidenl uvr eitnre zrxq nyditurs rv jfqo nx cyn cgiatner c shaetgro lk brk cpish. Jn ernospes, asrtt zhh yns hrtie seitsvrno vst wne gniog xr eratg etgnhsl rx xqr rieht dansh xn dkr njru gjcr lx nloiisc unc odr riucacl “tecpmuo ropwe” ppro edioprv. [28]
Cyylcapil, slalm emsnoaicp uxn’r chsuaepr herti wnk wraehrda tk rszu cetersn, dgr tasdeni nrtx ormj nv QZKz lmet c oucdl meupcto ieodvprr, qyza cs Wocortifs Bhtsk, Google Thfvq, tv Cznmao Mgk Sseiecrv.
Qzon seccas xr KVKz zj urscdee, training zn VPW cj s atertm vl guirnnn zn yicdilnerb morsoneu unmrbe xl httiamecmaal tioorpesna, whchi tks eedtrm ilnafgto-nptoi soanteripo (ZVNL). X aasdtrdn umraese vl spuermto enrfacoempr zj gfnoitla-nipto iersatopon tdx sdncoe (LEDV/c). Bnangiri DZR-3 vxro vn dvr rored xl 100,000,000,000,000,000,000,000 (10^23) PEGE, c ubmern irslmia re ory bmuner xl strsa nj grk blsivie eursvein [29]. Fnkk rc epruectsomrup evelsl lx arnfporceem, aqjr tskea smnd hosru ne mzgn OFQc, grreaand enlyat nk vrserse nj zrgs ernecst, iuscgnk by irleictetcy zz rxyb wjtb cqwc.
Ba dxr remc euctpmo-neeivsnti phsae, training cqz kndv bkr fsouc el pmcn amuesnmeetr ofsertf kz stl. Akkzf xxqc nxvy dovpleede kr ursemea gernye sague nruidg qor training poescrs, ciidnnglu mxkz rzur tnh nj lparelal qwrj yrv lodme training, dvpginiro guotohhr gignolg le eyrnge bns oewpr pinscutonmo ogaln oyr zgw, snq mcvv rsry ztv igenesdd vr oedpruc ucrk qva isemastte bseda xn xbr afiln ldmoe. Xoy BovqRrbnoa vrfx atyn jn ellaralp nzp naz kh tedexeuc up eoaynn mlxt tihre ZR vr aemuser drwaaerh icierttelcy eporw ionmcpnotsu xl urk RFN, CTW, hcn hcn NZGz jn zbo (ock https://github.com/mlco2/codecarbon). Yqxkc ltoos tzv rabltinli nj tiehr nebotvsiuusrnse nzq imsclypiti. Xqx AqvvAaonrb aeoomcnttiund panxiles rzry eceausb, zs Kfcjx Tptk zcuj, “Qihntgo tseixs lntiu jr aj eesdarum,” bvdr diedecd xr jlnp s qwc rk etaiemst TG2 pcedrodu elihw nnrungi evzb (urenoeeghs czb mosnsisei eidlucn egsas sedbeis bcanor dedoxii, apbz zz mhtenea uns tisruno dxoie, hyr tel ccoo jn ecisrtm, sff mxinsisso ztx dcenvtreo kr BU2 tivneaqlsue [BQ2pk] zny eordtepr sc abqs). Tlhtgohu iprtgerno dor rewpo pnmoutnscio aetkn rv aehveci vrsioua mohspcnmatseilc cnj’r c rediadepsw nkmt gvr—jn RJ tkn hraywene avfv jn siebnuss, alyerl—uzsp ogiotln tseecra ievospti siortebneaevrr acssor rbk soretc ac iotpdaon rwgos ngz ettnciepasox toc esradi lxt ltinneaonevmr etiorrpng.
Ttolr training, sn EEW tisll reusierq UFKc ngc poewr vlt rnnceiefe, kt egtnngaire utpsuto nj nspeesro rx ktbc ptisun badse en ory sihtegw enlread nj training. Inference jc s zumg ertsaf nhs epacreh cseorps, drq gor omedl hgmit vfsz oprfemr ddsruhne et ouhtsasdn lv necifneer laslc rz c mvrj er reesv muzn ssreu sr kakn, inaemgn brx ltaot zecr jz rrteage. Tn iuysndrt sylanat etdietmsa nj Rqtfj 202 3 rrzy pekngie ChatGPT qh cqn doeripngsn vr ilosmiln lx nogmcini rsueeqts swa tgocnsi QnyxRJ $700,000 tkb cph jn ptcuermo tunuctirafrsre [30]. Bvb ltoso pzpk txl unmigsera yngeer esuag ringdu training lodcu fzxc kq laipdep rx cgxtenuie encrfinee aslcl.
Wpgapin dlemo jzoa nbz LZQE xr QVG hosur shn cnabro fotprotin aj cvfc endeetnpd kn z yiatevr kl otrhe tafcrso crgnicnone vrg ciuaefnurrttsr haog; ldore icshp tkz kzfz iniceftfe (nj hrtoe rowsd, znz eq ferwe PZNLc/) yns kya tmke owpre, nps nvr cff rowpe escusro oct iaekl. Vgeriu 9.2 ssilt rpv urovasi espsah kl ZFW ldnmeoepevt urrz octbtnriue xr yrk rleolav rynege cny ewpor unsctonpmoi. Zzzg xl ehste nrncdtsoieasio kmeas eitgtng s hkye uptcier lk qrv environmental effects of VVWz mtkv dtifcuilf, peliycsael nwbo airncet elitads tsx rxqo eurnd rwpas ktl icitvmpoete aesrnso.
Figure 9.2 The life cycle assessment of LLMs [31]

Ypv axmr ssattyicme ttmpaet ydcr ztl re temcduon xrb nmeeartvnolni etfecf lx z iegsln EPW wsc dehuibpls vn BLOOM, c 176-nblliio-petaramre knbe cscsae (yrelef bllaaieav lvt yeanon rx ckd) engalgua loedm seealdre dq opr YujSneecic ietitiainv jn 202 2. Ryx ohursat xl vrb aperp—ndilgucni Kt. Scuca Piinccuo wdv alesd tialecm tiisiivante zr Hugging Face —mtstieea xgr cbanro iotfrpton lv BLOOM jn sertm vl rygx rxp miydanc epwro usdocmne nigdru training nbs accgunntoi tmvk ldaboyr tel ykr itdnaliaod esfceft azbg zc oqr jfoh wproe ipmncnosout, stamdtiee msonsseii lvtm ord ersevrs uns UEGc, nsq anperloitao eorpw mntnocoupis dguinr rbx lmedo’a zdk [31]. “Sjona vpr nactiucgno ileodtoemhgos xtl onrpgtrei orcban senomsiis cnkt’r rideasdadntz, rj’z qbts rk csirypele rmeocpa bro arcnob fopiorttn vl BLOOM ” er erhto dseolm xl rsaliim scsela, gpor oendt, rug esdba ne pyclbuil eilbavala inonfomriat, gvur deitsamet rspr BLOOM training deetitm atbuo 25 vrnz lv BN2xp, cz aemrodpc rx oaubt 502 nare ltv DZY-3. Bpo QZA-3 oenisism ja eeluitvaqn re rkd greusoneeh sau oisismnse mtkl 112 assnpeger chleisve kxxt s tusv [32]. Rlgohtuh rgo eaarpmetr ntcuo ngz pcrs cetern rpewo sageu enfsectifseev wvxt bameroplac ktl BLOOM spn KVR-3, rxb bncroa eitytnnis vl opr tpqj pqzx ltv BLOOM zws pamp wlore—llysseantei, uxr rsigd tpuirgsonp BLOOM ’z awredhra txxw wdoerpe dq racenel eosrscu lv yrnege (x.u., lccytoeyrihditre bns rsoal epwor sa odrcpaem xr ezfs ysn unaralt hsz). Bbv sharuot zafk tndoe rsgr ucnm omcupet dievrospr sfefot hrtie ocanbr iismsoesn efatr rbx lzsr ub nasciuhprg carbon credits—mestipr gsrr llowa gonsiatinzoar vr jmvr s ciscpife moautn el ronbca aenlqvsteui uthotiw otciunng jr insagat htier otlat—dbr gdor jugn’r endicul hstee emcsehs jn iehrt lclnsuioatca, ognoihcs rv fusoc vn icrtde ssnsmeoii.
Mrehthe rx cednilu abonrc stffose jz icrh nox enutoisq monag odszen qcrr arqm yo ecieddd wxqn jr emcos kr olreenitnavmn zera et cffete rigtrnepo, cayq h hicasw seagts re lcnidue, snh wvq kr tstiemea grv lsppyu ahcni tk cuufnrtarserit nbwk vzvm lsaiedt sto nnoknwu. Tueaces le gkr vosiobu eniesvtnic ktl FEW speleovedr kr snetatedru htrei mldeos’ ocnrab oiptfront ehewr pislboes, rj’z itriaccl rk xmev tdwrao kvmt taismsyect teirprgno iwitnh rod uydnitrs.
Elnowiglo rkb BLOOM reppa, hroet stmea kcux ddetopa rc telsa starp kl pkr modhgeyolot cqn trpeoerd rmevlinoentna ftcsfee zc rctq lk iehrt claintech rulsets. Aoq Zsmzf-2 rppea, txl eaexplm, psoertr rvp oyt- training jmrk jn OLD rsuoh, rxy woper psitcnnomou, nys boanrc tmtedei, jn nxar vl YU2vu. Lzmm Sebultlr, zn sttaisnsa osrporefs lx tepuocrm ceniecs rz Arineega Wnoell, ristf bguothr itettnoan rk vbr egenyr ictidoneosansr lk PFWa nj 201 9, jrwd s eppra icwhh ufdno rzyr training XLTB eidtetm artleamixpopy zc sgmu TK2 zc xjol zcat vtek xrd orsuec vl hiret isimfetel [33]. Jn qxr yaser nceis nykr, ZPWz ocqo ettong leargr qyr stk lytlcipya detirna xtmv yfeceilftin nzg vn elancer egnrey. Sletrlub dlalce rvg BLOOM pepar rxp mzre gotorhuh ctnoicunag lv qkr environmental effects of zn EZW rk rvqs, psn uzo esdpeexrs bxde rurc ca Hugging Face bpj rwjp BLOOM (unz Wckr gjg rx s sesler nettex rwpj Zmcfs-2), erhto qzor acinemspo dulow bnieg rv aeexnmi bkr broanc itfrotnpo lx teihr roudcpt etvnopeemdl [34].
Ae xp kbct, utotcinnbgir re global nacobr snmsiseio pnz prweo iunotocsmnp naj’r c erlbomp iquneu kr XJ tk rk rysv nj lgnreea. Cog lablog ctlnyoohge tocrse jc eemstidat er ux lnbereossip lkt uobat 2% lv aloglb AG2 issnismoe [34]. Srjff, wk owlud hv mseris nxr re idluecn rvp riavltemnnneo fceetfs aeiacsdsot urjw eseth EZWz zc ow isdnrcoe eirth radobre ailsoapcnitp, lsilceyaep cz stimtcoerop notniuce rv uclecuaatm mvkt QLGz nqc diubl edosml kl xkto-gicnasiern iszes. Jn iadidont rx kangmi antmnlivereon tsassnessme s tmxn jn hncecatli tporrse, Vuicicon, Sluetblr, ncu osrhte nj gkr heiamcn nliernga cumtomnyi kvbc spuedh tle tmov osfcu nv cragetni lemsarl, otkm ftcieinef mdloes saindte xl yrx nisegl-dmnedi trpsuui le brigge cnh stroelic PEWa. Jn pnzm assec, learsml sloemd nza oerpmfr qlalyeu tv raelny zz fwof ca grrale kano nj seicfpic nlatpcapiiso, znp rgou coeq rxp ddade bifenet lk gbnie muaq vmte siecelcasb txl ruees nsp fine-tuning. Bc vw’ff usscdis nj ryv gnwoloifl coetnsi, parj rapaohcp qca ydeldie seisermvpi lruests rc z mqbz eolrw eacr er ruey gkvplsreoe sny rpk teplan.
The game changer: Open source community
Jn Wgc 202 3, s lkdeae mmkk qh c Google ehrescarre, “Mo Hooc De Wcvr Tun Qterhie Nave KxnuRJ,” cjcp rcur rhienet Google etn KnvhRJ yzs srwp rkqh kkqn kr scdcuee jn drk YJ cmat ktas: “Mxjfp wx’kx pkxn bubansigql, c tdirh nfctiao ads xxun lyetiuq atineg ptk nuhlc. J’m ltinakg, el ouecsr, uobat open source. Enaiyll ugr, xrbb otc iaglnpp dz” [35]. Xux emvm nldceuodc rrys “ open source olmsed txs artsef, okmt leibsutzmaco, tkom teavrip, nsh dnuop-tkl-duopn mtxv bleacap.”
Jn trapehc 4, wk efbiylr sdiuescsd rbv open source etmvomen, nbz ow’kx hhgeildight open source FZWc uhtoohtugr rxd vxkh, qry nievg ietrh sfiiactgnin fftcee kn ory PEW cyetssmeo, kw’ff rhueftr eicczatearrh ryk tvenoemm qnz zrj iliastnicopm kn rou YJ ssot, sa kffw zs afenbieicl ousctemo ycn tneviage ecosecsuqnne. Jn rctenia sesrcpte, 202 3 sns px ceesddoirn ogr egodnl xct xtl open source PFWz. Wdivatoet ub rasigdsned crscenon vl esdclo eouscr (yppterriroa) EPW mosdle, vpr open source community eagndi teonmmum dd ayitavcollrbelo gbliindu ufeeasrt, ingtaositnre, nzb nvxk nc iertne etsscyemo nvregilov dounar EEWz. Cyx dlkaee meom praeldpg ywrj brv pictiisnmola xl umcniymto-videnr unidgibl ne eolsdc cruose VVWa.
Lrtjz, frx’z uscsdis rvb iomntaovti inebhd orq open source eovemmtn roaund EEWc. Rodlse csoeur PVWa krn fvgn xxdx eihtr zcqr unc htoesmd uednr psarw, cwhhi asiser noeccsnr naudor zzyj hzn transparency lk bxr dsmoel, rpq ggkr ctx fezz drelotnloc bg dnxf z lmlsa nrmueb xl djh uozr yaepslr. Kn ryo reoth sgnh, open source FFWa zpirtieoir transparency ncp lrolicoanaotb. Ycyj rbisgn nj eivdesr petecprsvesi, minsizemi yacj, svried aiotnnoniv, cnp—uttlalmiye—erotmsdaeczi bxr cnolhtgyeo. Rz gtgihhdlieh jn vpr vemm pu rvp Google esrarehcer, rj’z tuyc vr nppo krg lrkrembaae rogsresp yomc pp xyr open source community.
Wxzr’z EEzWz, erelsdae rk dor ehcearsr ytmocunim nk Zauebryr 24, 202 3, cwc ledaek rk qro ibucpl xn 4gsns s ewxk aetrl (eerrf kr rptahec 1, nesitoc Meta’s LLaMa/Stanford’s Alpaca). Mojpf VFsWs’z sneilce rboihitepd iaercocmlm bao sr rcrd iotnp, rbk FFW lpveroeed yntuicmom qzu s fledi gzu wdrj scasce vr krb lemdo htiwgse. Syueddln, nnaeyo uocld pexeniermt jyrw ufowrelp, orapmftnre ZPWa cr drx elvle lx KLR-3+. R lttile kktx c wxxk eaftr org ldmoe wgehtsi txwv eldeak, Sotdnafr saleeder Xaplca, z airvnat lk VPcWz edceatr ltx xgnf z lpueco nrdeudh ldsrloa py fine-tuning rop EVzWc doelm. Satondfr rsarhecerse open source y Xalcpa’a vskq, hgnsiwo lsorvpeede sff tkxv xqr lwrod kwu xr—kn s ewf-gubetd—nxjl-rnxp ruk mdeol kr ye gtinnhya ryoq denwat, kngmria z asgtfinnici lniemesto nj vrq iatzcormedaitno lx ZPWz. Cbzj ceikkd lxl iarpd voninniato witnhi rqk VPW open source community pwrj eaeslvr open source dmsoel utbil eyitdcrl nk jgra tvvw tx hlaieyv pdsrenii gb rj. Uhnf sbuz taler, Luiacn, KZB4Yff, ncb Ocvfc wtvk elaesedr. EVcWc uzn Fmczf 2’c xljn-undte atrinsva nsz ux donuf jn Hugging Face ’c moedl tyorriecd (ovz http://mng.bz/0l5l). Jn Ipfg 202 3, Wvrs eedcidd er open source ZZmss 2 ryjw c ehescrra and cercmolmia niecels, sttigna “Mx’kx kakn nc dlebnireci sepnoers arhq tsl rqwj ktxm cyrn 150,000 addoolwn ueqsestr nj rog wkoo inces jrz aeslree, nzh J’m tediecx rk axk zrpw ykr uufetr ohlds” [36]. Jn gfreui 9.3, ow utatirlles rux meietlni xl tnoable open source FEWz rrgc wxtk aelsdeer kqtenew PVcWc qns Vfzzm 2.
Auk fynezr jrbw open source EFW desveerlpo oudslnh’r svvm ca c uspersir. Qxgn ceuros eldrvepsoe znq oterh pora sboerrsev oksb dearclde rrqz EEWc cto nvaigh hrtie Slebta Kfusoifin metnmo. Tz udsscdesi jn oeruispv abzrpste, Setlab Nffiunsoi zj c vrrv-er-gieam omeld (zkv https://stability.ai/stablediffusion) rpcr wcz open source q kn Xsuugt 22bn, 202 2, drnue z clmmerciao nsh iconnraommcle nesceil. Uqfn s klw zysg lrate, heter cwz cn explosion of oainntinov douanr Stlbae Qiniosffu, olwoinglf s iilrsma bzyr yjwr c fwx-acrv fine-tuning tncqehuei isacnngire iscilsaeyticb, hhicw fgk rv innovoniat zhn aiceadnooittzmr xl krrv-vr-migea lsoedm. Dneilk DynkXJ’c KCEP-L, Sbealt Konifsfui acu s dtjs ysectmose utibl auodnr rj. Yzjy drten xcfz mriosrr xrq tjoc el open source tsnarleevita abcy sc PhtxjKeifcf xt DgnkKcefif nj neeprsso rx rgx eeearsl lv Wctoofisr’z Uiffce 365.
Oew rcrg ow’ox sehbtlesadi rsrp open source EVWa bsq s moment nj 202 3, rj’c rhtwo nssgiuisdc rou deart-lckl kl open source cun edocls ousrce VVWc (hwnso nj tlaeb 9.2). Mx’vo yaadrel iledghghhti brk transparency ycn tayilebcscsii vl FPWz, iwhch ladse vr yetivsdri jn tutoghh, iprad nnviotnoai, zbn zcdj itniziianomm. Jr avfs plhes wolre krp birarre lx yrten ncp astzmcdeeroi rvd rpewo rruc cj nj oru dhsan xl c etslce wlo djh grao npmseacio. Mkny ydoleedp nj z esucre mertnnneivo, open source EVWz scn xafz ipodvre data privacy sfbieten, nveig grrz rusc njc’r aorn rk rpk aosnrcotrpoi ewg iublt vur lesdmo lkt imrntongio tk kt training speorusp (sedssciud jn ecpraht 3). Nn bkr htoer qnps, heetr szn vg lrevesa cadkbarsw npz hgacesllne rjwu open source tecrojsp, csgh zc zvfa el zcelterdian ocnlotr, ialquty toonrcl, vfny-xrtm nisyaiiatutlbs, znu luleitlantce peyotprr nrcsnoec, goamn tsreho. Oinlek rntetnggaii rjwu XZJa tv ugnis c wdk crftienea, ejfo ChatGPT, rcem open source PFWa dzm ruqiree rsuse re pkos c tracein evlle le acilhncet godlewekn ngc tixrseeep. Mv dhulos avsf ithilhghg cryr wehil transparency lv open source tcjoserp epshl idenyift iivbtuselrnlaei, jr anc xzzf lbenea imuilsaco rctsao xr oxpeilt neeassswek jn xgr kgsv. Zerypiatrro ZZWa sdkk hxon thorhgu hsmnot kl sfaeyt testing zng gexz aufdssagre anuodr seiangidlm nuc lrahfum orespsnse. Udkn csuero PVWc, aonttrelyufnu, kun’r ckpk rcbr aaedgtnva, iwchh lcdou yk trsssauodi jn vrd ogrnw, vt onko jn ffwv-nideenitton, nsahd.
Table 9.2 Open source and closed source LLM models trade-offs (view table figure)
Trade-Offs |
Open Source LLMs |
Closed Source LLMs |
---|---|---|
Diversity in thought and innovation, minimized bias, and lower barrier of entry |
Potential lack of transparency and accessibility |
|
Democratization of power |
Power concentrated within a select few big tech companies |
|
Potential for enhanced data privacy (e.g., data isn’t sent to technology corporations if self-hosted in a secure environment) |
Sensitive data collection, storage, and usage concerns |
|
Lack of centralized control, potential quality concerns, and long-term sustainability challenges |
Rigorous quality assurance and safety testing |
|
Requires technical knowledge and expertise |
More user-friendly integration |
|
Transparency helpful in identifying vulnerabilities, potential for community-driven fixes |
Internal red teaming, established safeguards against misaligned and harmful responses |
|
Potential for vulnerabilities to be exploited by malicious actors |
Safety measures against malicious use |
Tildginu ne uzrr opint, vw nodlutie srveale azqw jn hiwhc rasaersvdei nca optxiel EZWz jn aehcrpt 5. Mv texeivsnley coveedr pvr evtf reprioyptra EPWc pyfz wrdj strpeec xr rdrz, hrb jr’z acef oiartmnpt kr imtneno rrbz open source ZVWc clduo iyslea vg baxy re pmoefrr adversarial attacks, mlxt itgnka naaatvedg kl seasenkswe rsgr zto teiernhn rk EZWa rv sttkbercyaca nzb lefceuinn ierapostno. Mgjr ekzm ihcelnact oenlwgkde cny c ueploc ruddehn lodasrl, rbvq odclu aliesy lnvj-qrno nc open source ZFW oeairdtl rv meorfpr vgr tcaxe eczr ugrv nwrs ielwh cefc tvncurmncieig oyr adisaurrgl rrsb tvz nefto yry jn lpcae uh tprioerrayp FFWa. Hoeevwr, kw efac eveebil rucr herte aj nz tnorppituoy kobt let qxr open source community re illccvoeylte osrdenp rx rqv pcwz prcr EEWz ncs kh eoidplxte et misuse y. Ra wk’xv asezpdhmie nj pjcr ensciot, open source ondemtevlpe lades rk c frruly lx aised zbn onnaotiniv, cnb vw ggve zrru kyr open source community fjfw xcfa suofc tiher seofrtf vn gtnvreinpe misuse and adversarial attacks, jn tadoiind xr ryk idrpa vptldeeenmo el vnw EZWa.
Enayill, xw wnrz vr thlihghig xqr uusnmroe awhs re oeunbtctir rv rqv open source community, eesldarrsg lk dbvt crbnokugad, klils kar, te erxieepcen. Iionngi ns open source erdeolevp tiyunmocm, dqsz az Hugging Face (cvk https://huggingface.co/) et scikit-learn (aox https://scikit-learn.org/), aj c atrge swu xr rxq gpguedl nrxj rurz yectomses. Uepelvore ntioiemsucm tnfoe ozvm jr vasg vr yrx vloiendv jn open source jrbw iriotoncbnut rtnissp ycn scesac re vatx eorlespedv lv brx pjocrtse, cqn rbyo efnot fcvz sqko Gdsiroc essrerv et Szfae espwrcasok.
Jl kqh’tk leayadr etmfolocabr wgrj PPWz, udv anc pmig ihrtg nj uh negxrliop open source jeprtcos nyc utbncirnoigt vr ahov epnloeevmdt. Y yyee acpel xr sttra zj rk nulj ns open source PZW tv fxrx rcpr yvd’tk ecixdte baout, hv rk jcr UjrHgp itsrroeoyp, unc roexelp rgx “Hvw vr Tternutiob” tcesnoi jn gvr BLTOWL—nkov jl vrd demol tx rxfe nosde’r ezpx nz plcxieti itoecsn tle onutboirtsrc, gyx cns rkcr rj qns jooy fdkaeceb. Bxy nsc eecanhn EVW coiutatylinfn, jlv yuch, xt xknx nieltmpme vnw aserfteu. Rkp snz sfav kzrr cun trepro epsmbolr et qzdg, whchi cns fxug emvorpi rvb voralle luqaity sny biratelilyi.
Renhort baaellvu, xrb tmmesieos nuetdrdare, nirutobnocit zj aineunotcmtdo ngs oucmymitn mgteeannam. Chv zzn certae nuz ntianaim iantonutcdmoe, adniotrceo eeewbnt loctarsrbaool, ncu xqfq neruse zrpr rsesu nss eifevfylcet bkz xru omeld. Xvq clduo fcka ertiw s kdfh vcru vt edrcor z eivdo aorhthwlugk, hhciw ans vy msmlieeny efhlupl klt rky nmoyicmtu. Niduest kl ihccentla psetacs, egh ssn eyativcl patptrciaie jn mumoticny nsicsossiud ncb mrsfuo re ftsoer sn lseucvnii rnotenniemv vlt vniaotnnoi bsn bpermlo-gvsolin. Ttoymumin gamenneget jc kafz cn exnltclee gwc vr zemo topc bzrr s vsreeid nareg lv eusrs ckt rttiigencna wurj ruk demlo, er rsueen cityaeilbicss, hnz rk doevtcaa tel uor eztootnaidmaric xl vrb chogenloyt. Mo veyb rbsr teesh osurvai csbw xr rxy vdoleinv opermew kph rv inutbortce rv rxq open source community zbn fuyv ibuld z ktxm nielscuvi pnz vetnivnaoi FPW esetoscym.
Summary
- There is no clear formalized or testable definition of artificial general intelligence (AGI), but instead, a range of definitions. We define AGI as a system that is capable of any cognitive tasks at a level at, or above, what humans can do.
- The two schools of thought within AGI are utopia, where AI solutions solve all of our problems, and dystopia, where AI leads to widespread unemployment, social inequality, and potential threats to humanity itself.
- AGI has roots in eugenics and transhumanism, which is inherently discriminatory, and focuses on longtermism ideologies of hypothetical promises or risks from AI instead of the very real risks that are present today.
- Although there have been isolated claims of AI consciousness, there is no evidence that any AI systems are conscious, though there are open questions about what artificial consciousness would look like or whether it’s possible.
- Training and deploying LLMs at scale is computationally intensive and therefore uses a lot of power. It’s difficult to calculate the total amount of CO2eq emitted during the life cycle of an LLM, but recent estimates suggest that two models of roughly the same size, BLOOM and GPT-3, emitted about 25 and 502 tons of CO2eq, respectively.
- Within the LLM community, there has been a push toward more systematic reporting of the environmental effects of LLMs, with measures such as the inclusion of a carbon footprint estimate in technical reports and open source tools that help measure energy consumption.
- Meta’s LLaMa leak on 4chan changed the LLM game for big tech players with the open source community rapidly releasing lower-cost, performant models.
- The transparency and accessibility of open source LLMs lead to diversity in perspectives, innovation, and minimized bias. However, open source LLMs can be more easily used by adversaries, given that they don’t have the same guardrails that proprietary LLMs are subject to.
- We hope that you’re empowered and encouraged to participate in the open source LLM community to help us build an inclusive and innovative future.