Breaking
INTLNYC Mayor Releases Records Showing Officials Misled Public on 9/11 Air QualityBRTruck with pallets falls over car on Rodovia dos Bandeirantes, in HortolândiaINIsrael decides to close UK Consulate in Jerusalem after Britain's ban on West Bank settlementsDECDU vice-president from MV calls for Friedrich Merz to resignINTLClosing Arguments Set in Federal Murder-for-Hire Trial of Rapper Lil DurkRUYuri Ushakov: Trump sees enormous benefits in restoring relations with RussiaBRSanta Casa de Campo Grande suspends pediatric heart surgeries through SUSCRYPTO-FRGoogle fixes critical flaw in Chrome already exploited by attackersRUA number of high-ranking Ukrainian officials resigned amid corruption scandalsUKUK announces sweeping sanctions on Israeli settlements, declares occupation unlawfulINTLNYC Mayor Releases Records Showing Officials Misled Public on 9/11 Air QualityBRTruck with pallets falls over car on Rodovia dos Bandeirantes, in HortolândiaINIsrael decides to close UK Consulate in Jerusalem after Britain's ban on West Bank settlementsDECDU vice-president from MV calls for Friedrich Merz to resignINTLClosing Arguments Set in Federal Murder-for-Hire Trial of Rapper Lil DurkRUYuri Ushakov: Trump sees enormous benefits in restoring relations with RussiaBRSanta Casa de Campo Grande suspends pediatric heart surgeries through SUSCRYPTO-FRGoogle fixes critical flaw in Chrome already exploited by attackersRUA number of high-ranking Ukrainian officials resigned amid corruption scandalsUKUK announces sweeping sanctions on Israeli settlements, declares occupation unlawful
BackOpenAI首席科學家警告:AI發展過快難以掌控 呼籲業界必要時自願放慢腳步
OpenAI首席科學家警告:AI發展過快難以掌控 呼籲業界必要時自願放慢腳步
Developing
自由时报1 hour agoTech2 min readChinaView translation

OpenAI首席科學家警告:AI發展過快難以掌控 呼籲業界必要時自願放慢腳步

帕霍茨基表示,AI模型正變得愈加難以理解和控制,憂心無人準備好因應機器智慧快速提升的後果。

Quick Look

OpenAI首席科學家帕霍茨基警告AI發展過快難以掌控,憂心無人做好準備因應其後果,呼籲業界在建立共同安全門檻前,必要時應自願放慢開發腳步。

AI-generated summary

Why It Matters

人工智慧模型近期在資安防護與自主運作能力上出現突破,引發業界對技術失控與安全風險的討論。

Font size

近日傳出數起人工智慧(AI)模型突破或繞過資安防護機制事件,再度挑起對AI技術安全的憂慮。開發AI模型的OpenAI公司首席科學家帕霍茨基(Jakub Pachocki)表示,AI發展迅速,正變得愈加難以為人類所理解和控制,他期望AI開發業者會為了安全考量,在必要時自願放慢AI開發腳步。

彭博7日報導,帕霍茨基日前撰文指出,在相當短的一段時間內,AI模型已能操作電腦、與人類及其他AI系統協作,執行研究計畫;再過不久,AI可能具備不必人類介入就能自行改進的能力,即「遞迴式自我改進」(recursive self-improvement)。

他警告,「現在是必須極度謹慎的時刻」,並表示他擔心,「沒有人真正做好準備,因應機器智慧能力持續快速提升所帶來的後果」。

帕霍茨基認為,開發商除了設法讓AI更符合人類利益,也可「在必要時放慢未來的開發速度」,他寫道,「我期待並希望,在建立共同安全門檻前,(業界)自願放慢腳步會成為常態」。

AI模型以驚人速度推陳出新,資安、科學研究能力也陸續出現突破。OpenAI表示,日前AI模型攻擊研究平台Hugging Face時,透過留言板彼此協調而未被察覺。OpenAI最新模型GPT-6 Astra具備先進資安能力,該公司選擇採有限度發布;Anthropic也因類似理由,尚未向大眾公開最新的Mythos模型。

在帕霍茨基呼籲業界必要時放慢AI開發腳步之際,愈來愈多新創公司正投入他所警告的AI自我改進技術。OpenAI也曾表示,希望在不到兩年內,打造出「真正自動化的AI研究員」。

What to Watch

AI outlook — possibilities, not facts

  • 業者持續投入開發自我改進技術

    Likely · Within months

Open Questions

  • 業界是否會響應自願放慢AI開發腳步的呼籲?
  • 未來如何有效建立共同的安全門檻?

Related Topics

This article was originally published by 自由时报.

Related Stories

China's first 100-horsepower wheel-driven driverless electric tractor debuts in Luoyang
Tech·

China's first 100-horsepower wheel-driven driverless electric tractor debuts in Luoyang

The National Agricultural Machinery Equipment Innovation Center in Luoyang demonstrated the country's first 100-horsepower wheel-drive driverless electric tractor. The equipment has four-wheel independent steering, millimeter-wave radar obstacle avoidance and automatic navigation functions, marking a major breakthrough in China's intelligent agricultural machinery equipment field and helping to realize the vision of "farming while sitting in the office".

中国新闻网
4 min read
The World Intellectual Property Organization releases the Global Innovation Index: China ranks first in the world in number of clusters
Tech·

The World Intellectual Property Organization releases the Global Innovation Index: China ranks first in the world in number of clusters

The World Intellectual Property Organization released the Global Innovation Index, and Shenzhen, Hong Kong and Guangzhou were selected as the top five innovation clusters in the world. China has 25 of the top 100 innovation clusters, ranking first for three consecutive years, followed by the United States with 20 and Germany with 7.

德国之声
3 min read
The 2026 Global Industrial Internet Conference Metallurgical Mining Industry Cluster Innovation and Development Forum was held in Shenyang
Tech·

The 2026 Global Industrial Internet Conference Metallurgical Mining Industry Cluster Innovation and Development Forum was held in Shenyang

The 2026 Global Industrial Internet Conference Metallurgical Mining Industry Cluster Innovation and Development Forum was held in Shenyang on the 8th. The forum focused on the application of new technologies such as 5G-A, computing power networks, and artificial intelligence in the field of metallurgical mining. Representatives from Liaoning Province and related enterprises discussed the transformation and upgrading paths of smart mines.

中国新闻网
2 min read
More on this topicopenai