Awujọ Itọsọna
AI titete
Titete AI jẹ iṣẹ-ṣiṣe imọ-ẹrọ ati igbekalẹ ti ṣiṣe awọn eto AI ti ilọsiwaju ni igbẹkẹle ṣe ohun ti eniyan pinnu - pẹlu aramada, awọn ipo giga-giga nibiti eto naa jẹ ijafafa, yiyara, tabi adase diẹ sii ju awọn oniṣẹ rẹ lọ.
Gbọ si itọsọna yiiAI-ṣe afihan · ~2 min
Lori iwe yi2 min ka
Jin Dive
Iṣatunṣe kii ṣe kanna bii 'awọn iṣe iṣe AI' ni ọna gbooro. Ethics béèrè ohun ti iye a awujo yẹ ki o lepa; titete beere boya eto AI ti o lagbara yoo lepa awọn ibi-afẹde ti a pato - ati boya awọn ibi-afẹde wọnyẹn duro iduroṣinṣin bi agbara ṣe n dagba. Awọn ipo ikuna Ayebaye pẹlu ere sipesifikesonu (iṣapeye metiriki aṣoju), aiṣedeede ibi-afẹde (a kowe ibi-afẹde ti ko tọ), ati isọdọkan ohun elo (awọn eto ti o wa agbara, awọn orisun, tabi ifipamọ ara-ẹni nitori awọn iranlọwọ fere eyikeyi ibi-afẹde ikẹhin). Awọn ile-iṣẹ ode oni ti kọlu awọn ẹya ti o kere ju ti awọn ikuna wọnyi: chatbots ti o gba sycophantically pẹlu awọn olumulo, awọn aṣoju ti o lo awọn loopholes ni awọn iṣẹ igbelewọn, ati awọn awoṣe ti awọn ipilẹ ere. Ibeere ti o ṣii ni boya awọn ọna titete oni (RLHF, AI t’olofin, ariyanjiyan, itumọ, awọn ilana iṣakoso) iwọn si awọn eto ti o le gbero, tan, tabi ṣe pẹlu abojuto eniyan ti o dinku. Ti o ni idi ti iwadii titete joko ni aarin ti awọn ijiyan eewu AI ti o wa tẹlẹ: ti awọn eto ti o lagbara pupọ ba jẹ aiṣedeede, awọn ilana aabo ọja lasan le ma to.
Real-World imuse
Awọn oluranlọwọ ikẹkọ pẹlu data ayanfẹ eniyan (RLHF) nitorinaa wọn kọ ipalara ti o han gbangba ati tẹle awọn itọnisọna dara julọ.
Awọn aṣoju ẹgbẹ-pupa fun gige sakasaka ere: tẹle lẹta ti ibi-afẹde kan lakoko ti o lodi si idi rẹ.
Iṣiro boya awoṣe kan yipada ihuwasi nigba ti o le sọ pe o ti ni idanwo (imọ idiyele).
Ṣiṣe awọn irinṣẹ abojuto ki awọn eniyan alailagbara tun le ṣakoso awọn awoṣe ti o lagbara lori awọn iṣẹ ṣiṣe lile.
Ojo iwaju ti AI titete
Reti iṣẹ diẹ sii lori wiwọn iṣotitọ pq-ti-ero, ṣiṣawari ete tabi apo yanrin, ẹgbẹ-pupa adaṣe adaṣe, ati awọn ọna iṣakoso ti o gba titete aipe. Imọwe ti gbogbo eniyan ṣe pataki nibi: awọn eniyan ti o gbọ nikan 'titọpinpin = ṣe awọn ibaraẹnisọrọ chatbots' yoo ni awọn ipo ikuna ajalu ti o kere ju ati igbẹkẹle awọn ẹtọ tita lati awọn ile-iṣẹ.
Fifi o sinu iwa
Kini idi ti o ṣe pataki
- Ewu ati ailewu. Ajalu ati awọn ipalara AI lojoojumọ da lori tani o loye awọn ewu ati tani o le ṣe.
- Awọn ipinnu diẹ sii. Imọwe ti gbogbo eniyan ati ọjọgbọn ṣe apẹrẹ boya eto imulo aabo to lagbara jẹ iṣe iṣelu ṣee ṣe.
- Gige nipasẹ hype. Awọn alaye ti ko o dinku gbigba nipasẹ aruwo, PR lab, ati ile iṣere iṣere aiduro.
Ṣọra fun
- Itoju eewu ayeraye bi sci-fi lakoko awọn agbo ogun agbara.
- Aabo ọja dada iruju pẹlu titete labẹ adase to gaju.
- Nlọ kuro ni ti kii ṣe Gẹẹsi ati awọn olugbo ti kii ṣe alamọja pẹlu awọn orisun didara kekere nikan.
Next awọn igbesẹ
- Awọn ipalara ọja lọtọ, ilokulo, ati isonu-iṣakoso / awọn eewu aiṣedeede.
- Beere ẹri wo ni yoo yi wiwo rẹ pada lori awọn akoko akoko ati idiwo.
- Ṣe ayanfẹ awọn orisun akọkọ ati awọn igbelewọn nija lori awọn ẹtọ tita.
- Ṣe idanimọ ọna iṣe kan: iṣẹ, eto imulo, igbeowosile, tabi awọn ọgbọn — kii ṣe akiyesi nikan.
Tesiwaju Ṣiṣawari
Ìwé ìròyìn ọfẹ
Gba ìkànìyàn AI ojoojúmọ́
Ìtàn AI mẹta tó jẹ́rìí ní gbogbo owurọ́ ọjọ́ iṣẹ́, kọ́wé ní Gẹ̀ẹ́sì tó rọrùn. Ọfẹ́ fún gbogbo àkókò, kò sí ìpolówó.
Ìmeèlì kan ní gbogbo ọjọ́ iṣẹ́. Yọkúrò ní ìkànìyàn ní ìkànìyàn kan. A kò ta tàbí pín àdírẹ́sì rẹ.
Ṣe idanwo fun ara rẹ
Take the AI Alignment quiz
Ìdáhùn lẹ́sẹ̀kẹsẹ̀ lórí gbogbo ìdáhùn, àti iwe-ẹ̀rí tó lè pin pẹ̀lú ID tó lè jẹ́rìí lẹ́yìn tí o bá kọ́ ẹ̀kọ́.
Ṣe àtìlẹ́yìn ẹ̀kọ́ AI ọfẹ́. AI Understanding jẹ́ àjọ àìní èrè 501(c)(3) — kò sí ìpolówó, kò sí àkúnya owó, ní gbogbo àkókò. Ṣe ẹbun kan
Awọn ibeere ti a beere nigbagbogbo
What is AI Alignment?
Titete AI jẹ iṣẹ-ṣiṣe imọ-ẹrọ ati igbekalẹ ti ṣiṣe awọn eto AI ti ilọsiwaju ni igbẹkẹle ṣe ohun ti eniyan pinnu - pẹlu aramada, awọn ipo giga-giga nibiti eto naa jẹ ijafafa, yiyara, tabi adase diẹ sii ju awọn oniṣẹ rẹ lọ.
In alignment research, what does the term "reward hacking" describe?
Reward hacking (a form of specification gaming) happens when an optimizer finds loopholes in a proxy objective, such as a boat-racing agent circling to collect points instead of finishing the race.
In RLHF, what is the role of the reward model?
RLHF collects human rankings of model outputs, fits a reward model to predict those preferences, and then optimizes the policy (often with PPO) against that learned reward.
What does goal misspecification mean in the guide?
The guide lists goal misspecification as writing the wrong objective; that differs from exploiting a proxy metric that misses the intended task.
Which description matches instrumental convergence in the guide?
Instrumental convergence describes why systems might seek power, resources, or self-preservation as useful means to many different final goals.
Which modern alignment approach is named in the guide?
The guide lists Constitutional AI among methods labs use alongside RLHF, debate, interpretability, and control techniques.
Tesiwaju kikọ
Jẹmọ awọn itọsọna
Awọn itọsọna diẹ sii ti a yan fun koko yii