{"id":10342,"date":"2024-02-27T18:58:26","date_gmt":"2024-02-27T18:58:26","guid":{"rendered":"https:\/\/dailyai.com\/?p=10342"},"modified":"2024-03-28T00:37:55","modified_gmt":"2024-03-28T00:37:55","slug":"generative-ai-systems-hallucinations-and-mounting-technical-debt","status":"publish","type":"post","link":"https:\/\/dailyai.com\/sv\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/","title":{"rendered":"Generativa AI-system, hallucinationer och \u00f6kande teknisk skuld"},"content":{"rendered":"<p><strong>I takt med att AI-system som stora spr\u00e5kmodeller (LLM) v\u00e4xer i storlek och komplexitet uppt\u00e4cker forskarna sp\u00e4nnande grundl\u00e4ggande begr\u00e4nsningar.\u00a0<\/strong><\/p>\n<p>Nya studier fr\u00e5n Google och University of Singapore har avsl\u00f6jat mekaniken bakom AI-\"hallucinationer\" - d\u00e4r modeller genererar \u00f6vertygande men fabricerad information - och ackumuleringen av \"teknisk skuld\", vilket kan skapa r\u00f6riga, op\u00e5litliga system \u00f6ver tid.<\/p>\n<p>Ut\u00f6ver de tekniska utmaningarna \u00e4r det fortfarande en \u00f6ppen fr\u00e5ga att anpassa AI:s kapacitet och incitament till m\u00e4nskliga v\u00e4rderingar.<\/p>\n<p>N\u00e4r f\u00f6retag som OpenAI str\u00e4var mot artificiell allm\u00e4n intelligens (AGI) m\u00e5ste man erk\u00e4nna gr\u00e4nserna f\u00f6r nuvarande system f\u00f6r att s\u00e4kra v\u00e4gen fram\u00e5t.<\/p>\n<p><span style=\"font-weight: 400;\">Men att noggrant erk\u00e4nna risker \u00e4r i strid med Silicon Valleys motto att \"g\u00e5 snabbt fram och f\u00f6rst\u00f6ra saker\", vilket k\u00e4nnetecknar AI FoU liksom det gjorde f\u00f6r tekniska innovationer f\u00f6re det.\u00a0<\/span><\/p>\n<h2>Studie 1: AI-modeller ackumulerar \"teknisk skuld<\/h2>\n<p><span style=\"font-weight: 400;\">Maskininl\u00e4rning framh\u00e5lls ofta som kontinuerligt skalbart, med system som erbjuder ett modul\u00e4rt, integrerat ramverk f\u00f6r utveckling.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Men i bakgrunden kan utvecklarna bygga upp en h\u00f6g niv\u00e5 av \"teknisk skuld\" som de kommer att beh\u00f6va l\u00f6sa l\u00e4ngre fram.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">I en <\/span><a href=\"https:\/\/storage.googleapis.com\/gweb-research2023-media\/pubtools\/pdf\/43146.pdf\" target=\"_blank\" rel=\"noopener\"><span style=\"font-weight: 400;\">Googles forskningsrapport<\/span><\/a><span style=\"font-weight: 400;\">, \"Maskininl\u00e4rning: The High-Interest Credit Card of Technical Debt\", diskuterar forskare begreppet teknisk skuld i samband med ML-system.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">D. Sculley, VD f\u00f6r Kaggle och m\u00e5ng\u00e5rig Google-forskare, och hans kollegor h\u00e4vdar att ML visserligen erbjuder kraftfulla verktyg f\u00f6r att snabbt bygga komplexa system, men att dessa \"snabba vinster\" ofta \u00e4r missvisande.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Den enkla och snabba implementeringen av ML-modeller kan d\u00f6lja de framtida b\u00f6rdor som de medf\u00f6r f\u00f6r systemets underh\u00e5ll och utveckling. <\/span><\/p>\n<p><span style=\"font-weight: 400;\">Som f\u00f6rfattarna beskriver uppst\u00e5r denna dolda skuld p\u00e5 grund av flera ML-specifika riskfaktorer som utvecklare b\u00f6r undvika eller omarbeta.<\/span><\/p>\n<p>H\u00e4r \u00e4r de viktigaste insikterna:<\/p>\n<ul>\n<li><span style=\"font-weight: 400;\">ML-system \u00e4r till sin natur komplexa p\u00e5 ett s\u00e4tt som g\u00e5r ut\u00f6ver ren kodning. <\/span>Detta kan leda till vad f\u00f6rfattarna kallar \"boundary erosion\", d\u00e4r de tydliga linjerna mellan olika systemkomponenter suddas ut p\u00e5 grund av de \u00f6msesidiga beroenden som ML-modellerna skapar. Detta g\u00f6r det sv\u00e5rt att isolera och genomf\u00f6ra f\u00f6rb\u00e4ttringar utan att p\u00e5verka andra delar av systemet.<\/li>\n<li>Rapporten belyser ocks\u00e5 problemet med \"sammanfl\u00e4tning\", d\u00e4r f\u00f6r\u00e4ndringar i n\u00e5gon del av ett ML-system, t.ex. indata eller modellparametrar, kan f\u00e5 of\u00f6ruts\u00e4gbara effekter p\u00e5 resten av systemet. Om man \u00e4ndrar en liten parameter kan det leda till en kaskad av effekter som p\u00e5verkar hela modellens funktion och integritet.<\/li>\n<li>Ett annat problem \u00e4r skapandet av \"dolda \u00e5terkopplingsslingor\", d\u00e4r ML-modeller p\u00e5verkar sina egna tr\u00e4ningsdata p\u00e5 of\u00f6rutsedda s\u00e4tt. Detta kan leda till system som utvecklas i oavsiktliga riktningar, vilket f\u00f6rv\u00e4rrar sv\u00e5righeten att hantera och f\u00f6rst\u00e5 systemets beteende.<\/li>\n<li>F\u00f6rfattarna tar ocks\u00e5 upp \"databeroenden\", t.ex. n\u00e4r insignaler \u00e4ndras \u00f6ver tid, som \u00e4r s\u00e4rskilt problematiska eftersom de \u00e4r sv\u00e5rare att uppt\u00e4cka.<\/li>\n<\/ul>\n<h3>Varf\u00f6r teknisk skuld \u00e4r viktigt<\/h3>\n<p><span style=\"font-weight: 400;\">Teknisk skuld ber\u00f6r ML-systemens l\u00e5ngsiktiga h\u00e4lsa och effektivitet. <\/span><\/p>\n<p><span style=\"font-weight: 400;\">N\u00e4r utvecklare skyndar sig f\u00f6r att f\u00e5 ML-system att fungera kan de ignorera de r\u00f6riga invecklade datahanteringen eller fallgroparna med att \"limma\" ihop olika delar. <\/span><\/p>\n<p><span style=\"font-weight: 400;\">Detta kan fungera p\u00e5 kort sikt men kan leda till en trasslig r\u00f6ra som \u00e4r sv\u00e5r att dissekera, uppdatera eller ens f\u00f6rst\u00e5 senare.<\/span><\/p>\n<blockquote class=\"twitter-tweet\">\n<p dir=\"ltr\" lang=\"en\">\u26a0\ufe0f \u26a0\ufe0f \u26a0\ufe0f \u26a0\ufe0f \u26a0\ufe0f \u26a0\ufe0f \u26a0\ufe0f<\/p>\n<p>GenAI \u00e4r en lavin av teknisk skuld* som bara v\u00e4ntar p\u00e5 att h\u00e4nda<\/p>\n<p>Bara den h\u00e4r veckan<br \/>\n\ud83d\udc49ChatGPT gick \"b\u00e4rs\u00e4rk\" med n\u00e4stan ingen riktig f\u00f6rklaring<br \/>\n\ud83d\udc49Sora kan inte konsekvent dra slutsatser om hur m\u00e5nga ben en katt har<br \/>\n\ud83d\udc49Geminis m\u00e5ngfaldssatsning sp\u00e5rade ur fullst\u00e4ndigt.... <a href=\"https:\/\/t.co\/qzrVlpX9yz\">pic.twitter.com\/qzrVlpX9yz<\/a><\/p>\n<p>- Gary Marcus @ AAAI 2024 (@GaryMarcus) <a href=\"https:\/\/twitter.com\/GaryMarcus\/status\/1761414330577539340?ref_src=twsrc%5Etfw\">24 februari 2024<\/a><\/p><\/blockquote>\n<p><script async src=\"https:\/\/platform.twitter.com\/widgets.js\" charset=\"utf-8\"><\/script><br \/>\n<span style=\"font-weight: 400;\">Att till exempel anv\u00e4nda ML-modeller som de \u00e4r fr\u00e5n ett bibliotek verkar effektivt tills du fastnar i en mardr\u00f6m med \"limkod\", d\u00e4r det mesta av systemet bara \u00e4r silvertejp som h\u00e5ller ihop bitar som inte var avsedda att passa ihop.\u00a0<\/span><\/p>\n<p>Eller ta \"pipeline-djungler\", som beskrivs i en <a href=\"https:\/\/proceedings.neurips.cc\/paper_files\/paper\/2015\/file\/86df7dcfd896fcaf2674f757a2463eba-Paper.pdf\" target=\"_blank\" rel=\"noopener\">f\u00f6reg\u00e5ende artikel av D. Sculley<\/a> och kollegor, d\u00e4r dataf\u00f6rberedelse blir en labyrint av sammanfl\u00e4tade processer, s\u00e5 att en f\u00f6r\u00e4ndring k\u00e4nns som att desarmera en bomb.<\/p>\n<h3>Konsekvenserna av teknisk skuld<\/h3>\n<p><span style=\"font-weight: 400;\">Till att b\u00f6rja med \u00e4r det s\u00e5 att ju mer invecklat ett system blir, desto sv\u00e5rare blir det att f\u00f6rb\u00e4ttra eller underh\u00e5lla det. Detta h\u00e4mmar inte bara innovation utan kan ocks\u00e5 leda till mer allvarliga problem.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Om ett ML-system till exempel b\u00f6rjar fatta beslut baserade p\u00e5 f\u00f6r\u00e5ldrade eller partiska data eftersom det \u00e4r f\u00f6r kr\u00e5ngligt att uppdatera dem, kan det f\u00f6rst\u00e4rka eller <\/span><a href=\"https:\/\/dailyai.com\/sv\/2023\/07\/unmasking-the-deep-seated-biases-in-ai-systems\/\"><span style=\"font-weight: 400;\">f\u00f6rst\u00e4rka samh\u00e4lleliga f\u00f6rdomar<\/span><\/a><span style=\"font-weight: 400;\">.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Dessutom, i kritiska applikationer som <\/span><a href=\"https:\/\/dailyai.com\/sv\/2024\/02\/does-ai-display-racial-and-gender-bias-when-evaluating-images\/\"><span style=\"font-weight: 400;\">sjukv\u00e5rd<\/span><\/a><span style=\"font-weight: 400;\"> eller autonoma fordon kan en s\u00e5dan teknisk skuld f\u00e5 allvarliga konsekvenser, inte bara i form av tid och pengar utan \u00e4ven f\u00f6r m\u00e4nniskors v\u00e4lbefinnande.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">I studien beskrivs det s\u00e5 h\u00e4r: \"Alla skulder \u00e4r inte n\u00f6dv\u00e4ndigtvis d\u00e5liga, men tekniska skulder tenderar att \u00f6ka. Att skjuta upp arbetet f\u00f6r att betala av den leder till \u00f6kade kostnader, sk\u00f6rhet i systemet och minskad innovationstakt.\"<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Det \u00e4r ocks\u00e5 en p\u00e5minnelse f\u00f6r f\u00f6retag och konsumenter att kr\u00e4va transparens och ansvarighet i den AI-teknik de anv\u00e4nder. <\/span><\/p>\n<p><span style=\"font-weight: 400;\">M\u00e5let \u00e4r trots allt att utnyttja kraften i AI f\u00f6r att g\u00f6ra livet b\u00e4ttre, inte att fastna i en o\u00e4ndlig cykel av teknisk skuld\u00e5terbetalning.<\/span><\/p>\n<h2>Studie 2: Det g\u00e5r inte att skilja hallucinationer fr\u00e5n LLM:er<\/h2>\n<p>I en annan men <a href=\"https:\/\/arxiv.org\/pdf\/2401.11817.pdf\" target=\"_blank\" rel=\"noopener\">relaterad studie<\/a> fr\u00e5n National University of Singapore unders\u00f6kte forskarna Ziwei Xu, Sanjay Jain och Mohan Kankanhalli de inneboende begr\u00e4nsningarna hos LLM.<\/p>\n<p><span style=\"font-weight: 400;\">\"Hallucinationer \u00e4r oundvikliga: An Innate Limitation of Large Language Models\" utforskar AI-hallucinationer, som beskriver fall d\u00e4r AI-system genererar trov\u00e4rdig men felaktig eller helt fabricerad information.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Hallucinationsfenomenen utg\u00f6r en stor teknisk utmaning, eftersom de visar p\u00e5 en grundl\u00e4ggande skillnad mellan vad en AI-modell producerar och vad som anses vara \"sanningen p\u00e5 marken\" - en idealisk modell som alltid producerar korrekt och logisk information.\u00a0<\/span><\/p>\n<p>Att f\u00f6rst\u00e5 hur och varf\u00f6r generativ AI hallucinerar \u00e4r av st\u00f6rsta vikt n\u00e4r tekniken integreras i kritiska sektorer som polis och r\u00e4ttsv\u00e4sende, h\u00e4lso- och sjukv\u00e5rd samt juridik.<\/p>\n<blockquote class=\"twitter-tweet\">\n<p dir=\"ltr\" lang=\"en\">T\u00e4nk om man kunde * bevisa * att hallucinationer \u00e4r oundvikliga inom LLM?<\/p>\n<p>Skulle det f\u00f6r\u00e4ndra<br \/>\n- Hur ser du p\u00e5 LLM?<br \/>\n- Hur stor investering skulle du g\u00f6ra i dem?<br \/>\n- Hur mycket skulle du prioritera forskning om alternativ?<\/p>\n<p>En ny rapport ger st\u00f6d f\u00f6r detta: <a href=\"https:\/\/t.co\/r0eP3mFxQg\">https:\/\/t.co\/r0eP3mFxQg<\/a><br \/>\nh\/t... <a href=\"https:\/\/t.co\/Id2kdaCSGk\">pic.twitter.com\/Id2kdaCSGk<\/a><\/p>\n<p>- Gary Marcus @ AAAI 2024 (@GaryMarcus) <a href=\"https:\/\/twitter.com\/GaryMarcus\/status\/1761764524674457662?ref_src=twsrc%5Etfw\">25 februari 2024<\/a><\/p><\/blockquote>\n<p><script async src=\"https:\/\/platform.twitter.com\/widgets.js\" charset=\"utf-8\"><\/script><\/p>\n<h3>Teoretiska grunder f\u00f6r hallucinationer<\/h3>\n<p><span style=\"font-weight: 400;\">Studien b\u00f6rjar med att l\u00e4gga fram ett teoretiskt ramverk f\u00f6r att f\u00f6rst\u00e5 hallucinationer hos LLM-personer.\u00a0<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Forskare c<\/span><span style=\"font-weight: 400;\">skapade en teoretisk modell som kallas \"den formella v\u00e4rlden\". Denna f\u00f6renklade, kontrollerade milj\u00f6 gjorde det m\u00f6jligt f\u00f6r dem att observera de f\u00f6rh\u00e5llanden under vilka AI-modeller misslyckas med att anpassa sig till markens sanning.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">De testade sedan tv\u00e5 stora familjer av LLM:er:<\/span><\/p>\n<ol>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>Lama 2<\/b><span style=\"font-weight: 400;\">: Specifikt anv\u00e4ndes 70-miljarder-parameterversionen (llama2-70b-chat-hf) som \u00e4r tillg\u00e4nglig p\u00e5 HuggingFace. Denna modell representerar en av de nyare posterna i den stora spr\u00e5kmodellarenan, utformad f\u00f6r ett brett utbud av textgenererings- och f\u00f6rst\u00e5elsesuppgifter.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><b>Generativa f\u00f6rtr\u00e4ngda transformatorer (GPT)<\/b><span style=\"font-weight: 400;\">: Studien omfattade tester p\u00e5 GPT-3.5, s\u00e4rskilt modellen gpt-3.5-turbo-16k med 175 miljarder parametrar, och GPT-4 (gpt-4-0613), f\u00f6r vilken det exakta antalet parametrar inte har offentliggjorts.\u00a0<\/span><\/li>\n<\/ol>\n<p><span style=\"font-weight: 400;\">LLM:erna ombads lista str\u00e4ngar av en viss l\u00e4ngd med hj\u00e4lp av ett angivet alfabet, en till synes enkel ber\u00e4kningsuppgift.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Mer specifikt fick modellerna i uppgift att generera alla m\u00f6jliga str\u00e4ngar med l\u00e4ngder som varierade fr\u00e5n 1 till 7, med hj\u00e4lp av alfabet med tv\u00e5 tecken (t.ex. {a, b}) och tre tecken (t.ex. {a, b, c}).<\/span><\/p>\n<p><span style=\"font-weight: 400;\"> Utdata utv\u00e4rderades baserat p\u00e5 om de inneh\u00f6ll alla och endast str\u00e4ngarna av den angivna l\u00e4ngden fr\u00e5n det givna alfabetet.<\/span><\/p>\n<h3>Resultat<\/h3>\n<p><span style=\"font-weight: 400;\">Resultaten visade en tydlig begr\u00e4nsning i modellernas f\u00f6rm\u00e5ga att slutf\u00f6ra uppgiften korrekt n\u00e4r komplexiteten \u00f6kade (dvs. n\u00e4r str\u00e4ngens l\u00e4ngd eller alfabetets storlek \u00f6kade). F\u00f6r att vara specifik:<\/span><\/p>\n<ul>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Modellerna fungerade tillfredsst\u00e4llande f\u00f6r kortare str\u00e4ngar och mindre alfabet men vacklade n\u00e4r uppgiftens komplexitet \u00f6kade.<\/span><\/li>\n<li style=\"font-weight: 400;\" aria-level=\"1\"><span style=\"font-weight: 400;\">Inte ens den avancerade GPT-4-modellen, den mest sofistikerade LLM som finns tillg\u00e4nglig just nu, lyckades lista alla str\u00e4ngar \u00f6ver vissa l\u00e4ngder.<\/span><\/li>\n<\/ul>\n<p>Det h\u00e4r visar att hallucinationer inte \u00e4r ett enkelt fel som kan lappas eller r\u00e4ttas till - de \u00e4r en grundl\u00e4ggande aspekt av hur dessa modeller f\u00f6rst\u00e5r och \u00e5terger m\u00e4nskligt spr\u00e5k.<\/p>\n<p>Som studien beskriver, <span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3\">\"LLM:er kan inte l\u00e4ra sig allt <\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3 r-b88u0q\">av<\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3\"> de ber\u00e4kningsbara funktionerna och kommer d\u00e4rf\u00f6r alltid att hallucinera. Eftersom den formella v\u00e4rlden <\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3 r-b88u0q\">\u00e4r<\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3\"> a del <\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3 r-b88u0q\">av<\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3\"> den verkliga v\u00e4rlden som <\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3 r-b88u0q\">\u00e4r<\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3\"> mycket mer komplicerat, hallucinationer \u00e4r ocks\u00e5 <\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3 r-b88u0q\">oundviklig<\/span><span class=\"css-1qaijid r-bcqeeo r-qvutc0 r-poiln3\"> f\u00f6r LLM:er i den verkliga v\u00e4rlden.\"<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Konsekvenserna f\u00f6r applikationer med h\u00f6ga insatser \u00e4r enorma. I sektorer som sjukv\u00e5rd, finans eller juridik, d\u00e4r korrekt information kan f\u00e5 allvarliga konsekvenser, kan det leda till allvarliga fel om man f\u00f6rlitar sig p\u00e5 en LLM utan en s\u00e4kerhetsfunktion f\u00f6r att filtrera bort dessa hallucinationer.<\/span><\/p>\n<p>Denna studie uppm\u00e4rksammades av AI-experten Dr. Gary Marcus och den framst\u00e5ende kognitiva psykologen Dr. Steven Pinker.<\/p>\n<blockquote class=\"twitter-tweet\">\n<p dir=\"ltr\" lang=\"en\">Hallucinationer \u00e4r oundvikliga med stora spr\u00e5kmodeller p\u00e5 grund av deras utformning: ingen representation av fakta eller saker, bara statistiska interkorrelationer. Nytt bevis p\u00e5 \"en medf\u00f6dd begr\u00e4nsning\" av LLM. <a href=\"https:\/\/t.co\/Hl1kqxJGXt\">https:\/\/t.co\/Hl1kqxJGXt<\/a><\/p>\n<p>- Steven Pinker (@sapinker) <a href=\"https:\/\/twitter.com\/sapinker\/status\/1761801185181200410?ref_src=twsrc%5Etfw\">25 februari 2024<\/a><\/p><\/blockquote>\n<p><script async src=\"https:\/\/platform.twitter.com\/widgets.js\" charset=\"utf-8\"><\/script><\/p>\n<h2>Djupare fr\u00e5gor \u00e4r i spel<\/h2>\n<p><span style=\"font-weight: 400;\">Den ackumulerade tekniska skulden och de oundvikliga hallucinationerna hos LLM \u00e4r symptomatiska f\u00f6r en djupare fr\u00e5ga - det nuvarande paradigmet f\u00f6r AI-utveckling kan i sig vara felinriktat f\u00f6r att skapa h\u00f6gintelligenta system och p\u00e5 ett tillf\u00f6rlitligt s\u00e4tt anpassa sig till m\u00e4nskliga v\u00e4rderingar och faktiska sanningar.<\/span><\/p>\n<p><span style=\"font-weight: 400;\">Inom k\u00e4nsliga omr\u00e5den r\u00e4cker det inte med att ha ett AI-system som har r\u00e4tt f\u00f6r det mesta. Teknisk skuld och hallucinationer hotar b\u00e5da modellintegriteten \u00f6ver tid.\u00a0<\/span><\/p>\n<p>Att \u00e5tg\u00e4rda detta \u00e4r inte bara en teknisk utmaning utan en tv\u00e4rvetenskaplig utmaning som kr\u00e4ver input fr\u00e5n AI-etik, policy och dom\u00e4nspecifik expertis f\u00f6r att navigera s\u00e4kert.<\/p>\n<p>Just nu verkar detta g\u00e5 stick i st\u00e4v med principerna f\u00f6r en bransch som lever upp till mottot \"move fast and break things\".<\/p>\n<p>L\u00e5t oss hoppas att m\u00e4nniskor inte \u00e4r \"sakerna\".<\/p>","protected":false},"excerpt":{"rendered":"<p>I takt med att AI-system som stora spr\u00e5kmodeller (LLM) v\u00e4xer i storlek och komplexitet uppt\u00e4cker forskarna sp\u00e4nnande grundl\u00e4ggande begr\u00e4nsningar.  Nya studier fr\u00e5n Google och University of Singapore har avsl\u00f6jat mekaniken bakom AI-\"hallucinationer\" - d\u00e4r modeller genererar \u00f6vertygande men fabricerad information - och ackumuleringen av \"teknisk skuld\", vilket kan skapa r\u00f6riga, op\u00e5litliga system \u00f6ver tid. Ut\u00f6ver de tekniska utmaningarna \u00e4r det fortfarande en \u00f6ppen fr\u00e5ga att anpassa AI:s kapacitet och incitament till m\u00e4nskliga v\u00e4rderingar. N\u00e4r f\u00f6retag som OpenAI str\u00e4var mot artificiell allm\u00e4n intelligens (AGI) inneb\u00e4r det att man m\u00e5ste erk\u00e4nna de nuvarande systemens gr\u00e4nser f\u00f6r att s\u00e4kra v\u00e4gen fram\u00e5t. Att noggrant erk\u00e4nna riskerna \u00e4r dock<\/p>","protected":false},"author":2,"featured_media":10364,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[84],"tags":[480,118],"class_list":["post-10342","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-industry","tag-hallucinations","tag-llms"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.4 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Generative AI systems, hallucinations, and mounting technical debt | DailyAI<\/title>\n<meta name=\"description\" content=\"As AI systems like large language models (LLMs) grow in size and complexity, researchers are uncovering intriguing fundamental limitations.\u00a0\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/dailyai.com\/sv\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/\" \/>\n<meta property=\"og:locale\" content=\"sv_SE\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Generative AI systems, hallucinations, and mounting technical debt | DailyAI\" \/>\n<meta property=\"og:description\" content=\"As AI systems like large language models (LLMs) grow in size and complexity, researchers are uncovering intriguing fundamental limitations.\u00a0\" \/>\n<meta property=\"og:url\" content=\"https:\/\/dailyai.com\/sv\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/\" \/>\n<meta property=\"og:site_name\" content=\"DailyAI\" \/>\n<meta property=\"article:published_time\" content=\"2024-02-27T18:58:26+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2024-03-28T00:37:55+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/dailyai.com\/wp-content\/uploads\/2024\/02\/shutterstock_2146674289.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"1000\" \/>\n\t<meta property=\"og:image:height\" content=\"667\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"Sam Jeans\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@DailyAIOfficial\" \/>\n<meta name=\"twitter:site\" content=\"@DailyAIOfficial\" \/>\n<meta name=\"twitter:label1\" content=\"Skriven av\" \/>\n\t<meta name=\"twitter:data1\" content=\"Sam Jeans\" \/>\n\t<meta name=\"twitter:label2\" content=\"Ber\u00e4knad l\u00e4stid\" \/>\n\t<meta name=\"twitter:data2\" content=\"7 minuter\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"NewsArticle\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/\"},\"author\":{\"name\":\"Sam Jeans\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#\\\/schema\\\/person\\\/711e81f945549438e8bbc579efdeb3c9\"},\"headline\":\"Generative AI systems, hallucinations, and mounting technical debt\",\"datePublished\":\"2024-02-27T18:58:26+00:00\",\"dateModified\":\"2024-03-28T00:37:55+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/\"},\"wordCount\":1569,\"publisher\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2024\\\/02\\\/shutterstock_2146674289.jpg\",\"keywords\":[\"Hallucinations\",\"LLMS\"],\"articleSection\":{\"1\":\"Industry\"},\"inLanguage\":\"sv-SE\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/\",\"url\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/\",\"name\":\"Generative AI systems, hallucinations, and mounting technical debt | DailyAI\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2024\\\/02\\\/shutterstock_2146674289.jpg\",\"datePublished\":\"2024-02-27T18:58:26+00:00\",\"dateModified\":\"2024-03-28T00:37:55+00:00\",\"description\":\"As AI systems like large language models (LLMs) grow in size and complexity, researchers are uncovering intriguing fundamental limitations.\u00a0\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/#breadcrumb\"},\"inLanguage\":\"sv-SE\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"sv-SE\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/#primaryimage\",\"url\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2024\\\/02\\\/shutterstock_2146674289.jpg\",\"contentUrl\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2024\\\/02\\\/shutterstock_2146674289.jpg\",\"width\":1000,\"height\":667,\"caption\":\"technical debt AI\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/2024\\\/02\\\/generative-ai-systems-hallucinations-and-mounting-technical-debt\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/dailyai.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Generative AI systems, hallucinations, and mounting technical debt\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#website\",\"url\":\"https:\\\/\\\/dailyai.com\\\/\",\"name\":\"DailyAI\",\"description\":\"Your Daily Dose of AI News\",\"publisher\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/dailyai.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"sv-SE\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#organization\",\"name\":\"DailyAI\",\"url\":\"https:\\\/\\\/dailyai.com\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"sv-SE\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2023\\\/06\\\/Daily-Ai_TL_colour.png\",\"contentUrl\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2023\\\/06\\\/Daily-Ai_TL_colour.png\",\"width\":4501,\"height\":934,\"caption\":\"DailyAI\"},\"image\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/x.com\\\/DailyAIOfficial\",\"https:\\\/\\\/www.linkedin.com\\\/company\\\/dailyaiofficial\\\/\",\"https:\\\/\\\/www.youtube.com\\\/@DailyAIOfficial\"]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#\\\/schema\\\/person\\\/711e81f945549438e8bbc579efdeb3c9\",\"name\":\"Sam Jeans\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"sv-SE\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/a24a4a8f8e2a1a275b7491dc9c9f032c401eabf23c3206da4628dc84b6dac5c8?s=96&d=robohash&r=g\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/a24a4a8f8e2a1a275b7491dc9c9f032c401eabf23c3206da4628dc84b6dac5c8?s=96&d=robohash&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/a24a4a8f8e2a1a275b7491dc9c9f032c401eabf23c3206da4628dc84b6dac5c8?s=96&d=robohash&r=g\",\"caption\":\"Sam Jeans\"},\"description\":\"Sam is a science and technology writer who has worked in various AI startups. When he\u2019s not writing, he can be found reading medical journals or digging through boxes of vinyl records.\",\"sameAs\":[\"https:\\\/\\\/www.linkedin.com\\\/in\\\/sam-jeans-6746b9142\\\/\"],\"url\":\"https:\\\/\\\/dailyai.com\\\/sv\\\/author\\\/samjeans\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Generativa AI-system, hallucinationer och \u00f6kande teknisk skuld | DailyAI","description":"I takt med att AI-system som stora spr\u00e5kmodeller (LLM) v\u00e4xer i storlek och komplexitet uppt\u00e4cker forskarna sp\u00e4nnande grundl\u00e4ggande begr\u00e4nsningar.\u00a0","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/dailyai.com\/sv\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/","og_locale":"sv_SE","og_type":"article","og_title":"Generative AI systems, hallucinations, and mounting technical debt | DailyAI","og_description":"As AI systems like large language models (LLMs) grow in size and complexity, researchers are uncovering intriguing fundamental limitations.\u00a0","og_url":"https:\/\/dailyai.com\/sv\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/","og_site_name":"DailyAI","article_published_time":"2024-02-27T18:58:26+00:00","article_modified_time":"2024-03-28T00:37:55+00:00","og_image":[{"width":1000,"height":667,"url":"https:\/\/dailyai.com\/wp-content\/uploads\/2024\/02\/shutterstock_2146674289.jpg","type":"image\/jpeg"}],"author":"Sam Jeans","twitter_card":"summary_large_image","twitter_creator":"@DailyAIOfficial","twitter_site":"@DailyAIOfficial","twitter_misc":{"Skriven av":"Sam Jeans","Ber\u00e4knad l\u00e4stid":"7 minuter"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"NewsArticle","@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/#article","isPartOf":{"@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/"},"author":{"name":"Sam Jeans","@id":"https:\/\/dailyai.com\/#\/schema\/person\/711e81f945549438e8bbc579efdeb3c9"},"headline":"Generative AI systems, hallucinations, and mounting technical debt","datePublished":"2024-02-27T18:58:26+00:00","dateModified":"2024-03-28T00:37:55+00:00","mainEntityOfPage":{"@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/"},"wordCount":1569,"publisher":{"@id":"https:\/\/dailyai.com\/#organization"},"image":{"@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/#primaryimage"},"thumbnailUrl":"https:\/\/dailyai.com\/wp-content\/uploads\/2024\/02\/shutterstock_2146674289.jpg","keywords":["Hallucinations","LLMS"],"articleSection":{"1":"Industry"},"inLanguage":"sv-SE"},{"@type":"WebPage","@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/","url":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/","name":"Generativa AI-system, hallucinationer och \u00f6kande teknisk skuld | DailyAI","isPartOf":{"@id":"https:\/\/dailyai.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/#primaryimage"},"image":{"@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/#primaryimage"},"thumbnailUrl":"https:\/\/dailyai.com\/wp-content\/uploads\/2024\/02\/shutterstock_2146674289.jpg","datePublished":"2024-02-27T18:58:26+00:00","dateModified":"2024-03-28T00:37:55+00:00","description":"I takt med att AI-system som stora spr\u00e5kmodeller (LLM) v\u00e4xer i storlek och komplexitet uppt\u00e4cker forskarna sp\u00e4nnande grundl\u00e4ggande begr\u00e4nsningar.\u00a0","breadcrumb":{"@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/#breadcrumb"},"inLanguage":"sv-SE","potentialAction":[{"@type":"ReadAction","target":["https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/"]}]},{"@type":"ImageObject","inLanguage":"sv-SE","@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/#primaryimage","url":"https:\/\/dailyai.com\/wp-content\/uploads\/2024\/02\/shutterstock_2146674289.jpg","contentUrl":"https:\/\/dailyai.com\/wp-content\/uploads\/2024\/02\/shutterstock_2146674289.jpg","width":1000,"height":667,"caption":"technical debt AI"},{"@type":"BreadcrumbList","@id":"https:\/\/dailyai.com\/2024\/02\/generative-ai-systems-hallucinations-and-mounting-technical-debt\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/dailyai.com\/"},{"@type":"ListItem","position":2,"name":"Generative AI systems, hallucinations, and mounting technical debt"}]},{"@type":"WebSite","@id":"https:\/\/dailyai.com\/#website","url":"https:\/\/dailyai.com\/","name":"DagligaAI","description":"Din dagliga dos av AI-nyheter","publisher":{"@id":"https:\/\/dailyai.com\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/dailyai.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"sv-SE"},{"@type":"Organization","@id":"https:\/\/dailyai.com\/#organization","name":"DagligaAI","url":"https:\/\/dailyai.com\/","logo":{"@type":"ImageObject","inLanguage":"sv-SE","@id":"https:\/\/dailyai.com\/#\/schema\/logo\/image\/","url":"https:\/\/dailyai.com\/wp-content\/uploads\/2023\/06\/Daily-Ai_TL_colour.png","contentUrl":"https:\/\/dailyai.com\/wp-content\/uploads\/2023\/06\/Daily-Ai_TL_colour.png","width":4501,"height":934,"caption":"DailyAI"},"image":{"@id":"https:\/\/dailyai.com\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/x.com\/DailyAIOfficial","https:\/\/www.linkedin.com\/company\/dailyaiofficial\/","https:\/\/www.youtube.com\/@DailyAIOfficial"]},{"@type":"Person","@id":"https:\/\/dailyai.com\/#\/schema\/person\/711e81f945549438e8bbc579efdeb3c9","name":"Sam Jeans","image":{"@type":"ImageObject","inLanguage":"sv-SE","@id":"https:\/\/secure.gravatar.com\/avatar\/a24a4a8f8e2a1a275b7491dc9c9f032c401eabf23c3206da4628dc84b6dac5c8?s=96&d=robohash&r=g","url":"https:\/\/secure.gravatar.com\/avatar\/a24a4a8f8e2a1a275b7491dc9c9f032c401eabf23c3206da4628dc84b6dac5c8?s=96&d=robohash&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/a24a4a8f8e2a1a275b7491dc9c9f032c401eabf23c3206da4628dc84b6dac5c8?s=96&d=robohash&r=g","caption":"Sam Jeans"},"description":"Sam \u00e4r en vetenskaps- och teknikskribent som har arbetat i olika AI-startups. N\u00e4r han inte skriver l\u00e4ser han medicinska tidskrifter eller gr\u00e4ver igenom l\u00e5dor med vinylskivor.","sameAs":["https:\/\/www.linkedin.com\/in\/sam-jeans-6746b9142\/"],"url":"https:\/\/dailyai.com\/sv\/author\/samjeans\/"}]}},"_links":{"self":[{"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/posts\/10342","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/comments?post=10342"}],"version-history":[{"count":11,"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/posts\/10342\/revisions"}],"predecessor-version":[{"id":10427,"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/posts\/10342\/revisions\/10427"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/media\/10364"}],"wp:attachment":[{"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/media?parent=10342"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/categories?post=10342"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dailyai.com\/sv\/wp-json\/wp\/v2\/tags?post=10342"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}