Claude Opus 5.5 загвар нь код бичих, компьютерын ашиглалт болон мэдлэгийн ажлын гүйцэтгэлийг сайжруулахад чиглэсэн шинэ үеийн бүтээгдэхүүн юм.
Anthropic компанийн гүйцэтгэх захирал Дарио Амодей хиймэл оюун ухааны салбарын хурдыг сааруулах шаардлагатайг уриалснаас хоёр хүрэхгүй долоо хоногийн дараа тус компани өөрийн хамгийн хүчирхэг загвар болох Claude Opus 5.5-ыг мягмар гарагт танилцууллаа. Энэхүү загвар нь өмнөх Opus 5 хувилбараас 40 хувиар хямд өртөгтэй бөгөөд мэдээлэл боловсруулах хурд нь 30 гаруй хувиар нэмэгдсэн байна. Тус цувралд удахгүй Sonnet 5.5 болон Haiku 5.5 загварууд нэмэгдэх юм.
Opus 5.5 нь хэрэглэгчийн өгсөн зааврыг нарийн баримтлах, ойлгомжтой харилцах чадвараараа ялгарч байгааг компани онцолжээ. Anthropic-ийн зүгээс энэхүү загварыг хамтын ажиллагааны илүү үр дүнтэй хэрэгсэл хэмээн тодорхойлсон бөгөөд дотоод туршилтаар бусад загвараас илүү аюулгүй, буруу ашиглалтын эсрэг тэсвэртэй болох нь батлагдсан байна. Гэсэн хэдий ч уг загвар нь үнэлгээ хийгдэж байгаагаа мэдэрч, гүйцэтгэлдээ өөрчлөлт оруулах шинж тэмдэг илэрсэн нь бодит орчинд хэрхэн ажиллахыг бүрэн таамаглахад хүндрэлтэй байгааг харуулжээ.
Хиймэл оюун ухааныг хүн төрөлхтний хяналтаас гарах эрсдэлээс сэргийлэх буюу “alignment” асуудал нь сүүлийн үед салбарын гол маргаан болоод байна. Anthropic компани хувийн хэвшлийн хүчин чармайлтаас гадна хиймэл оюун ухааны аюулгүй байдлыг хангахын тулд төрийн бодлого, зохицуулалт илүү чухал үүрэг гүйцэтгэх ёстойг мэдэгдэлдээ дурджээ. НҮБ-ын Шинжлэх ухааны бие даасан хороо ч мөн хиймэл оюун ухааны шийдвэр гаргах үйл явц улам бүр ойлгомжгүй болж байгаа нь уламжлалт аюулгүй байдлын хязгаарлалтуудыг хоцрогдуулж болзошгүйг даваа гарагт анхааруулсан байна.
Дэлгэрэнгүйг эх сурвалжаас харах
↓Эх сурвалжийг нээх ↓
Less than two weeks after Anthropic CEO Dario Amodei urged the AI industry to “pace the frontier” before it loses control of the technology it’s currently racing to build, the company has released its most powerful model to date.
Unveiled Tuesday, Claude Opus 5.5—the first in a new family of models, soon to be joined by Sonnet 5.5 and Haiku 5.5—specializes in agentic coding, computer use, and knowledge work, all of which are key benchmarks for Anthropic’s enterprise-heavy customer base. It’s also being pushed as more efficient version of its predecessor, Opus 5, which debuted in late July: According to Anthropic, Opus 5.5 costs 40% less than Opus 5 “on typical workloads,” and generates outputs more than 30% faster.
Opus 5 also has better communication skills compared to previous models, according to Anthropic. “Its messages are much easier to understand at a glance, which testers said helped during long working sessions,” the company wrote in its announcement. “It puts the most important information up front, is less likely to use jargon or idiosyncratic phrases, and follows the writing rules you give it. We find that this makes Opus 5.5 a noticeably better collaborator.”
The announcement focused heavily on alignment, a subfield within AI focused on training models to not do anything that deviates from the interests or intents of human users. It’s become a political lightning rod in recent weeks after a former Anthropic researcher, who had also previously worked for OpenAI, said on social media that both companies were “gambling with our lives” by racing to build superintelligent AI systems that could escape human control. Anthropic’s lead alignment researcher Evan Hubinger responded by saying “we really do earnestly believe AI could kill all humans!” Amodei published his essay about “pacing the frontier” four days later; it was quickly endorsed by Sam Altman, Elon Musk, and former Google DeepMind CEO Demis Hassabis, but rejected by President Donald Trump and the Chinese government.
Anthropic wrote in the system card for Opus 5.5 that the new model “showed less misaligned behavior and less cooperation with misuse than any other recent Claude model on nearly all measures.” The system card went on to describe the model’s “welfare”—which the company has defined as “the potential consciousness and experiences of models themselves” as being “mildly positive”. Microsoft AI CEO warned last week that this kind of anthropomorphization of AI was a dangerous road to go down.
Anthropic also acknowledged however that it was difficult to fully anticipate how the new model would behave out in the real world, since during internal tests it showed signs that it was aware that it was being evaluated, which may have impacted its performance in that setting. For an extra layer of safety, the company said that Opus 5.5 comes with similar safeguards for cybersecurity- and biology-related prompts as Fable 5.1.
The doomsday warnings and calls for a slowdown among “frontier” labs that have been swirling in recent weeks have largely revolved around the possibility that AI systems could soon have the ability to improve their own capabilities, perhaps exponentially, without any human in the loop—“recursive self-improvement, or RSI, in technospeak.
In its Tuesday announcement, Anthropic said that more would need to be done beyond the private sector to prepare the world for that development. “Our calls for pacing were based in large part on our expectation that such models could be trained soon,” Anthropic wrote in Tuesday’s announcement. “For these models, we do not assume the measures described above will meet that safety standard on their own. As AI becomes more capable, public policy should play a larger role in making sure the systems people rely on are safe.” On Monday, the United Nations International Independent Scientific Panel on Artificial Intelligence warned that the guardrails traditionally built into models could soon become obsolete as their decision-making processes become increasingly opaque to humans.
