Added on: 2025-03-14

Tehisarul põhinevad suured keelemudelid

Large language models based on artificial intelligence

Tartu Ülikooli professor Meelis Kull räägib tehisaru vigadest ja sellest, kuidas mudelid saaksid paremini oma ebakindlust tunnistada.

Professor Meelis Kull from the University of Tartu talks about AI errors and how models could better recognize their uncertainty.

Näiteks vajavad nii masin kui ka inimene keerulisema ülesande lahendamiseks aega. Mõlemad suudavad teha üldistusi ka siis, kui nad pole kindla küsimuse vastust enne õppinud.

AI makes both human-like and less human-like mistakes. People also make mistakes in different ways. For example, both machines and humans need time to solve more complex tasks. Both can make generalizations even if they have not learned the answer to a specific question before.

Suur erinevus on see, et inimene teeb paremini vahet, kas väide tuleb mälust või on see välja mõeldud.

A key difference is that humans are better at distinguishing whether a statement comes from memory or is made up. AI sometimes makes mistakes that humans rarely make.

Meelis Kull rääkis ka tehisarusüsteemide läbipaistvusest. Kui tehisaru teeb otsuseid, tahaksime teada, miks just selline otsus tehti.

Meelis Kull also spoke about the transparency of AI systems. When AI makes decisions, we would like to know why a particular decision was made.

See tähendab, et mudel annab enesekindlalt vastuse, mis on tegelikult vale. Näiteks küsimusel "Kes on Friedebert Luts?" vastas mudel, et ta oli inimene aastatel 1886–1956. Tegelikult sellist inimest pole kunagi eksisteerinud.

Think before you speak. One major problem is hallucinations. This means the model confidently gives an answer that is actually wrong. For example, when asked "Who is Friedebert Luts?", the model answered that he was a person from 1886–1956. In reality, such a person never existed.

Iga osa kohta saab vaadata, kui kindel mudel oli. Kuid see ei tee mudeli vastust täielikult usaldusväärseks.

Models construct answers piece by piece. For each part, you can see how confident the model was. However, this does not make the answer fully reliable.

Kui aega on vähe, tekivad vead tõenäolisemalt. Nii inimestel kui ka mudelitel võiks enne mõelda, siis öelda.

People and models are similar. When time is limited, errors become more likely. Both humans and models should think before speaking.

Näiteks võiks kirjutada, et "vasta pikalt ja samm-sammuliselt". See aitab mudelil vähem vigu teha.

Meelis Kull recommends breaking tasks into smaller parts when using a model. For example, you could specify "answer in detail and step by step". This helps the model make fewer mistakes.

Meelis Kull juhib Eesti tehisintellekti tippkeskust. Näiteks liivikeelsete sõnade puhul võib mudel eksida, kuna ta pole neid varem näinud.

Meelis Kull heads Estonia's AI excellence center. The center studies how AI could work better with small languages. For example, with Livonian words, the model may make mistakes since it has not encountered them before.

Ideaalne oleks, kui mudel ütleks, et "Seda sõna ma tegelikult ei teadnud". See aitaks usaldust luua. Inimesed tunnistavad, kui nad ei tea vastust, ja see loob usaldust.

The ideal would be for the model to say, "I did not actually know this word." This would help build trust. People admit when they don't know an answer, and this builds trust.

Praegu on mudelite teadmine piiratud, aga inimkonna teadmine on suurem. Masinad teevad mõnikord vigu, mida inimesed ei tee.

Meelis Kull hopes that in the future models will provide better explanations. Currently, the knowledge of models is limited, but humanity's knowledge is greater. Machines sometimes make mistakes that humans do not.