Claude is currently spearheading 26% of Anthropic's model development efforts; the company asserts that this signifies the model is capable of executing the majority of assigned tasks "end-to-end from high-level prompts" under human oversight.
Claude 現正主導 Anthropic 旗下 26% 的模型研發工作;公司宣稱,此舉意味著該模型得以在人類監督之下,「從高階提示端到端」遂行大多數指定任務。
Nevertheless, the model has yet to attain a state of fully autonomous operation.
惟該模型尚未臻於完全自主運作之境。
However, roughly ninety percent of the company's research and development work is carried out in a mode of "collaboration" with Claude;
然而,公司約莫九成的研發工作,皆是在與 Claude「協作」的模式下遂行;
Anthropic points out that this means the model is able to take on "a considerable amount of work" under close human supervision.
Anthropic 指出,此即意味著該模型得以在人類嚴密督導之下,承擔「為數可觀的工作」。
At the time this announcement was made, a number of leading figures in the field of AI — a chorus whose momentum was largely spearheaded by Anthropic CEO Dario Amodei — were, out of concern for safety, calling for a slowdown in the pace of the technology's development.
此公告問世之時,若干 AI 領域之領軍人物——其聲勢大抵由 Anthropic 執行長 Dario Amodei 所引領——正以安全為念,籲請緩其技術之進程。
As the company's leadership weighed the pace of AI development, the company stated in a blog post: "We should do everything in our power to narrow the gap between what frontier labs know and what the public knows."
在公司領導層權衡 AI 發展之步調之際,該公司於一篇部落格文章中表示:「我們應竭盡所能,縮小前沿實驗室所知與公眾所知之間的落差。」
"This means measuring AI development more effectively, reporting on it publicly, and giving society the opportunity to decide how this information is used."
「此即意味著要更妥善地衡量 AI 的發展、公開報告相關情況,並讓社會有機會決定如何使用這些資訊。」
The company stated that models accelerating their own development might render "these systems harder for humans to understand or control."
該公司表示,模型加速自身發展,或將致使「人類更難理解或控制這些系統」。
The company contends that disclosing such metrics would afford outsiders a clearer grasp of how close leading AI laboratories are to achieving recursive self-improvement — that is, a state in which a model is capable of autonomously producing its own successor versions.
該公司主張,披露此類指標可令外界更清晰地掌握領先 AI 實驗室距離達成遞迴自我改進尚有多近;所謂遞迴自我改進,係指模型得以自主締造其後繼版本。
Based on the information Anthropic has disclosed, it remains unclear how close the company believes it is to achieving recursive self-improvement; nevertheless, Claude's contribution to research and development is growing by the day, and the pace of that growth is truly not to be underestimated.
根據 Anthropic 所披露的資訊,該公司自認距離實現遞迴自我改進尚有多近,仍未明朗;然而,Claude 對研發之貢獻與日俱增,其增速著實不容小覷。
The proportion of work that Claude "led" — that is, work it largely completed while still subject to human oversight — stood at zero in February.
Claude「主導」——意即在仍受人類監督之前提下大量完成——之工作占比,於 2 月時為零。
Six months later, it had already taken the lead on a quarter of research and development work, reaching that benchmark in August.
六個月後,其已主導四分之一之研發工作,並於 8 月達到此一基準。
The company also disclosed specific details of its current agent oversight mechanisms, noting that as of August, approximately 30,000 agents were engaged in research and engineering-related work.
該公司亦披露其現行代理監督機制之具體細節,指出截至八月,約有30,000個代理正從事研究與工程相關工作。
The company stated that oversight measures are of paramount importance for ascertaining the frequency with which monitoring systems detect misconduct on the part of agents.
公司表示,監督措施對於掌握監測系統偵測代理不當行為之頻率,實屬至關重要。
Anthropic recently committed to installing external third-party evaluators, who will be embedded within the company to oversee safety-related work.
Anthropic近日承諾設置外部第三方評估人員,彼等將進駐公司內部,監督安全相關工作。
Last week, upon resigning, a researcher at Anthropic issued a stark warning about the threat that this technology poses to humanity, thereby igniting much of the recent debate surrounding AI safety.
上週,Anthropic 一名研究人員於辭職之際,就該項技術對人類所構成之威脅發出嚴厲警示,遂引發近期圍繞 AI 安全之諸多論辯。
Subsequently, tech leaders such as Amodei, OpenAI CEO Sam Altman, and Elon Musk all voiced support for the notion of slowing down development, whereas other tech leaders and US President Donald Trump took the opposing stance.
其後,Amodei、OpenAI 執行長 Sam Altman、Elon Musk 等科技領袖均表態支持放緩發展之構想,惟其他科技領袖與美國總統 Donald Trump 則持相反立場。