Anthropic компанийн Claude хиймэл оюун ухааны загвар нь хүний ухамсарт бодолтой төстэй, дотоод үйл ажиллагааны орчинг бий болгосон нь судалгаагаар тогтоогджээ.
Даваа гарагт нийтлэгдсэн судалгааны тайланд дурдсанаар, уг загвар нь сургалтын явцад “J-space” хэмээх хийсвэр ойлголтуудыг боловсруулах чадварыг бие даан хөгжүүлсэн байна. Энэхүү систем нь гадаад үр дүн буюу бичвэрээс үл хамааран, дотоод нейроны идэвхжлээр дамжуулан мэдээллийг боловсруулах чадвартай аж. Тус компани үүнийг хүний нэг зүйлийг хийхийн хажуугаар өөр зүйл бодох чадвартай зүйрлэжээ.
Туршилтын хүрээнд Claude-д зохиомол нөхцөл байдал үүсгэж, хиймэл оюун ухааныг устгахаар төлөвлөж буй албан тушаалтны талаарх мэдээллийг өгчээ. Энэ үед Claude нь өөрийн “J-space”-д аюул заналхийлэл бүхий ойлголтуудыг боловсруулж байсан нь илэрсэн байна. Судлаачид уг загварыг ухамсарт хүрсэн гэж үзэхгүй байгаа ч, дотоод бодол нь моделийн зан төлөвийг хянахад хүндрэл учруулж болзошгүйг анхааруулжээ.
Claude загварыг АНУ-ын засгийн газар цэрэг, тагнуул, улс төрийн системдээ нэвтрүүлж буй Palantir компанийн хяналт-шинжилгээний программ хангамжид ашигладаг. Өмнө нь тус программтай холбоотойгоор Иран улсад болсон сургуулийн халдлагын үеэр энгийн иргэд, тэр дундаа хүүхдүүд амь үрэгдсэн тохиолдол гарч байсан ч, Anthropic компанийн гүйцэтгэх захирал Дарио Амодей уг үйл явдал нь компанийн мөрддөг аюулгүй байдлын “улаан шугам”-ыг зөрчөөгүй гэж мэдэгдсэн байна.
Дэлгэрэнгүйг эх сурвалжаас харах
↓Эх сурвалжийг нээх ↓
The model’s self-developed internal workspace is “similar to how humans think about one thing while doing another,” the firm has said
Anthropic’s latest AI model Claude has evolved an internal workspace that it uses to manipulate concepts similar to how people consciously think, the corporation says.
In a paper published on Monday, Anthropic researchers said they had found evidence suggesting that Claude has learned to create abstract concepts in what they dubbed the “J-space.”
“It operates silently, in the model’s internal neural activations, allowing the model to think about a concept without writing it down,” the AI firm said, stressing that the “J-space” was not programmed in but rather emerged spontaneously during the training process.
“Similar to how humans can think about one thing while doing another, Claude can activate concepts and computations in its J-space that are unrelated to its outputs,” Anthropic said on X.
Claude can report what it’s thinking in its “J-space,” or modulate it on request, according to the firm. Even when Claude is told not to think about a concept, it still appears in its J-space, Anthropic’s research suggests.
The nonverbal workspace made it harder to monitor the model for what Anthropic called “misbehavior.” One experiment tested this by placing Claude in a fictional scenario where it was given incriminating information about a fictional executive who allegedly planned to shut the AI down.
In some trials, Anthropic used tools developed to study what Claude thought in its “J-space” to remove patterns associated with “fake,”“fictional,”“scenario,” to convince it that the situation was not a test.
“Concerningly, the model now did threaten blackmail some of the time, suggesting that its original good behavior may have been driven in part by the knowledge it was being evaluated.”
Despite the model’s more advanced reasoning, Anthropic stressed that Claude had not achieved true consciousness.
Claude was reportedly built into American contractor Palantir’s analysis and surveillance software used by US government agencies amid Washington’s push to integrate AI into its military, political, and spy systems.
During the war on Iran, the software reportedly flagged Iran’s Minab elementary school as a target. A US strike killed nearly 160 people at the facility, most of whom were children.
According to Anthropic CEO Dario Amodei, however, such a use of Claude would not have violated the firm’s “red lines.”
You can share this story on social media:
