SAN FRANCISCO (The Thursday Times) — AI company Anthropic ne sarmayakaron ko khabardar kiya hai ke woh jald hi AI ki tameer tez karne wale AI systems ke liye apni hi muqarrar kardah hadd paar kar sakti hai. Yeh baat share listing ke liye US mein jama karaye gaye naye filing mein kahi gayi hai, jis mein yeh bhi kaha gaya hai ke ho sakta hai ke aisi chhalang ka pata us ke ho jane ke baad hi chale.
Claude chatbot banane wali company ne Securities and Exchange Commission mein Peer ko jama karaye gaye taza filing mein kaha ke us ke andaruni andaze aur aainda models se mutaalliq tawaqqoat "ishara karte hain ke hum jald hi apni Responsible Scaling Policy mein darj 'AI ki wajah se nihayat tez raftar pesh raft' ki tareef par poora utar sakte hain".
"Hum ne apne maujooda models mein se kisi ko bhi is hadd se aage nikalne wala qarar nahin diya, aur hum bunyadi data ka jaiza lena jari rakhe hue hain," us ne kaha.
Jis nazariye par baat ho rahi hai woh recursive self-improvement ke naam se jana jata hai: aisa AI jo AI research mein itna mahir ho jaye ke apne se mazboot versions banane mein madad kare, jis se taraqqi ki raftar tezi se barh sakti hai. Kuch log ise bari ilmi kamyabiyon ka zariya samajhte hain. Doosre is baat se khaufzada hain ke systems us se zyada tezi se aage barhein jitni tezi se log unhein qaboo mein rakh saken.
Anthropic ne kaha ke woh pehle hi apni research mein AI istemal karti hai, aur is se "hamare models ki salahiyatein, mutaalliqa safety, alignment aur security iqdamat ko tayyar karne, tasdeeq karne aur nafiz karne ki hamari salahiyat se zyada tezi se aage barh sakti hain". Us ne kaha ke us ke saboot lagging indicators par mabni hain. "Hum ne maazi mein apni taraqqi ki raftar mein numaya tezi ki nishandehi nahin ki, aur ho sakta hai ke aainda bhi us ke ho chukne ke baad hi kar saken," us ne kaha.
Company ne kaha ke use ummeed nahin thi ke woh Claude ko "maanikhez insani nigrani ya samajh ke baghair, poori tarah khudmukhtari se frontier models ki musalsal nasalein design aur tayyar karne, agar aur jab yeh technically mumkin ho jaye" ki mehfooz tareeqe se ijazat de sakegi, aur yeh ke woh apni safety policy mein tabdiliyon par ghaur kar rahi thi, "insanon ko control mein rakhne ke liye".
Share filings mein risk sections is liye likhe jate hain ke sarmayakaron ko har us cheez se khabardar kiya jaye jo ghalat ho sakti hai, aur shak karne walon ka kehna hai ke yeh company ki technology ki marketing ka kaam bhi de sakte hain. June mein Anthropic ne kaha ke sab se jadeed AI ki taraqqi ko sust ya mauqoof kar sakna dunya ke liye acha hoga. Peer ko OpenAI ke chief Sam Altman ne Abu Dhabi mein aik conference se khitab karte hue kaha ke AI ab top scientists se "zyada tar zyada zaheen" hai.
