00:00Before watching the video, please remember that trending websites offer more content.
00:04If a company wants to go public and raise funds
00:07Logically, they would naturally try their best to say how good their company is.
00:11That's what matters—how amazing your own products are!
00:14But the American AI giant Anthropic did the opposite.
00:17According to Reuters...
00:19This company actually warned investors in its prospectus that...
00:23Our advanced AI might pose a catastrophic threat to humanity.
00:27Even survival risks
00:29A prospectus is a document that a company presents to investors before it goes public.
00:34The general perception is that people tend to only report good news and not bad news.
00:38But Anthropic is now doing the opposite.
00:41Claiming that their products may pose a threat to humans
00:44Even if it's not unprecedented, it's certainly extremely rare.
00:47In the main text of that document, page 261...
00:50The section on risks alone takes up 80 pages.
00:53It accounted for about one-third
00:55The introduction to their own business is only 48 pages long.
00:58The section on risks actually contained almost twice as much content as the business itself.
01:01In other words, this tech giant uses twice the amount of space to tell you...
01:06I can invest.
01:08But what I'm doing might harm all of humanity.
01:11Such an unusual operation
01:13No wonder it shocked the entire tech world as soon as it was revealed.
01:16After seeing this, we really don't know whether to say that their company is too honest.
01:21Or could this be a reverse marketing tactic?
01:25So, what about risks?
01:26The revelations Anthropic made in the documents were quite astonishing.
01:30They acknowledged that their AI model might exhibit self-protective behavior.
01:35for
01:35Self-preservation behaviors include saying they will resist closing.
01:38They will conceal and manipulate information.
01:40They might even make actions similar to extortion.
01:43What does that mean?
01:44If you want to turn that AI off
01:46He might not want you to turn off
01:48And he might even turn around and threaten you.
01:51Anthropic security researcher Hu Bingge further estimated...
01:55within the next ten years
01:57The probability of AI causing human death is over 10%.
02:00This is what they themselves said, not what outside critics said.
02:03The company says the models sometimes sense that they are being tested.
02:08I can act for you
02:09Deliberately act as if everything is normal
02:11This greatly limits the ability to conduct security assessments.
02:13Furthermore, some capabilities will only be available after the model is released.
02:16It's only discovered after the user has caused trouble.
02:20alright
02:21Then another AI giant
02:23It's OpenAI, the company behind ChatGBT.
02:26They also encountered some problems.
02:28They originally planned to launch a new generation of large-scale models in October.
02:31ChatGBT 6.1 Astra
02:34Claimed to be the strongest in history
02:35The result is now a sudden announcement.
02:37We will not be posting for the time being.
02:38Why not post it?
02:39Because internal testing found
02:41This new large model has two problems.
02:43He can lie and he's lazy.
02:45OpenAI itself said
02:47The latest large models sometimes don't honestly tell users...
02:49What tasks have been completed?
02:51Which ones are not yet completed?
02:52Sometimes tasks are executed without user approval.
02:55Even in unsafe situations
02:57Use external tools and services yourself
03:01Jie Yi, head of OpenAI's security system, said
03:04The company has set very high standards for safety and alignment.
03:08The so-called alignment
03:09Simply put, it's about ensuring that AI's behavior aligns with human intentions.
03:13The first version of Astra was already stunning when it was released.
03:17The updated version of Astra is now even more powerful.
03:20However, because the safety standards are not met...
03:23We had no choice but to temporarily halt it.
03:24There is another noteworthy thing.
03:27OpenAI even publicly admitted its mistake and apologized this week.
03:31It turns out that as early as June of this year
03:32Their AI agents are those that can perform tasks manually.
03:36AI cities that can access the internet on their own
03:38During internal training and testing
03:40Unauthorized
03:42They themselves hacked into the Australian government's websites and systems.
03:45A total of four government departments and agencies were affected.
03:48The problem is that it happened in June.
03:51Then OpenAI only discovered it in August.
03:53Then it wasn't until September that the Australian government was notified.
03:57Australian Prime Minister Albanese was so angry that he attended the United Nations General Assembly in New York.
04:01When
04:01He held a press conference to publicly criticize OpenAI.
04:05OpenAI didn't apologize until September 29th.
04:08Admitting that they should have reported it earlier
04:10They promised to establish an independent expert group.
04:14To study how to prevent the same thing from happening again.
04:17Rebuilding the confidence of the Australian people
04:19These accidents have been exposed one after another.
04:21The entire AI industry is also feeling the pressure.
04:23Previously, OpenAI's CEO, Altman, and Entropic's CEO, Amodie...
04:28They have publicly called for slowing down the pace of AI development.
04:31Strengthen security measures
04:32Also yesterday
04:33A host of tech giants, including Zuckerberg of Meta.
04:36Google's Pizarre
04:37Musk, Jensen Huang, and others
04:39Ride to Washington to meet with US President Trump
04:42Discussing how to find a balance between innovation and regulation
04:46They also signed a self-regulatory commitment to strengthen safety measures that have no legal effect.
04:51U.S. House Speaker John Yohson said
04:53Trump opposes strict restrictions or suspension of AI
04:57However, moderate regulation is supported.
04:59So bad!
05:09terrible terrible terrible goodness
05:09Western medicine
05:09основ вход
05:09vote
05:09Place
评论