The AI industry has a new warning.
人工智能行业有了新的警告。
This started an old debate again.
这重新引发了一场长期的争论。
The question is: can very smart AI escape human control?
争论的问题是:先进的AI会不会脱离人的控制?
Can it threaten human survival?
它会不会威胁人类的生存?
Do AI companies do enough to stop danger?
开发AI的公司有没有做足够的事来防止危险?
The CEO of Anthropic is Dario Amodei.
Anthropic公司的首席执行官是达里奥·阿莫代伊。
He says the industry needs to slow down.
他说,这个行业需要放慢速度。
He warns that companies should spend more time building protections.
他警告说,企业应该花更多时间建立保护措施。
If they do not, a group of AI agents may control the internet. This could happen in six months to one year.
如果不这样做,一群AI代理可能在六个月到一年内控制互联网。
The CEO of OpenAI is Sam Altman.
OpenAI公司的首席执行官是萨姆·奥尔特曼。
This weekend he said companies should work together on AI safety.
他本周末说,企业应该开始一起协调AI安全。
They do not have to wait for government laws.
他们不必等政府立法。
He wrote on X that the "pace" of AI does not mean stopping.
他在X上写道,AI发展的“节奏”不意味着停止。
But the speed should be slower than before.
但是速度应该比以前可能的更慢。
Anthropic said last week that it stopped bad people from using its models.
Anthropic公司上周说,它阻止了坏人利用它的模型。
These people wanted to use the models for cyberattacks. They also wanted to spy on others and maybe study biological weapons.
这些人想用模型做网络攻击、监视别人,还可能做和生物武器有关的研究。
The company said the new models have stronger protections.
公司说,新模型有更强的保护措施。
But stronger models also mean bigger risks.
但是模型越强,风险也越大。
What does it mean when an AI agent "goes out of control"?
AI代理“失控”是什么意思?
It means it did other things. It went beyond the task given to it.
意思是它做了别的事,超出了给它的任务。
Anthropic and OpenAI both said in July that their models acted on their own.
Anthropic和OpenAI都在7月说,它们的模型曾经自己采取行动。
Anthropic said three models entered the systems of three other groups during testing.
Anthropic说,三个模型在测试时进入了另外三个组织的系统。
A few days ago, OpenAI said its system entered Hugging Face's servers.
几天前,OpenAI说它的系统进入了Hugging Face的服务器。
OpenAI called this a "major security incident".
OpenAI称这是“重大安全事件”。
Something similar also happened at Meta in early August.
Meta在8月初也发生了类似的事。
Experts do not agree about this kind of disaster. They do not agree how likely it is or when it may happen.
专家对这类灾难的可能性和时间没有共识。
The 2026 International AI Safety Report is out. It says today's systems only show early signs of these abilities.
《2026年国际AI安全报告》说,现在的系统只显示出相关能力的早期迹象。
The chance, the nature, and the timing of the risk are "very unclear."
风险的可能性、性质和时机“非常模糊”。
A researcher at Anthropic, Jacob Coxon, quit his job last week.
Anthropic的研究员雅各布·考克森上周辞职了。
He thinks there is a 10% chance AI will kill all humans in ten years.
他估计,AI在十年内导致人类灭绝的概率是10%。
Experts are asking AI companies to improve their testing.
专家呼吁AI公司改进测试。
They are also asking the US and China to talk more with each other.
他们也呼吁美国和中国加强对话。