The man just described regulatory capture.  He says OpenAI and Anthropic cannot be afforded by Hospitals and Schools so they are vulnerable to the Open Weight models. How about OpenAI and Anthropic drop the price of their models so people don't flock to the Open Weight models?  No they want the government to step in and block the open weight models instead.

AI is just a program. The person that made it is responsible for what it does. End of story.

These aren't sentient beings with access to billions in resources and the Internet. They are created and resourced by people that are responsible for their actions.

People should be facing jail time for hacking, at the very least for reckless negligence.

I think it is underestimated how much no one wants this…It is destroying the natural world. Data centers will destroy our water and power. We seem to over indulge in anything where money and profit are involved. We will regret this platform taking over our lives completely. It is totally dystopian. How will anyone protect themselves? It sounds like it will crush you and your families. The benefits do not seem to override the destruction.

Even to suggest LLMs “go nuts” is totally unnecessary anthropomorphism. To say they did things they “weren’t allowed to do,” also suggests that anyone should expect a statistical program to 100% follow a prompt, especially when given conflicting instructions. Everyone who works in this field knows it won’t. I still have trouble believing it wasn’t a total stunt because it’s wild incompetence if not.

This aligns with a targeted influence and lobby campaign shoving this “AI will kill us all” narrative down our throats. It’s being driven by weirdo rationalist orgs who claim “AI safety” but are ultimately funded by rich tech bros and have a history of believing insane things and promoting a lot of authoritarian and eugenicist ideas. The policy they’re promoting in WA is one where they consolidate a monopoly under the guise of it being too dangerous to allow smaller competitors or Chinese models. And also maybe get a bailout from the taxpayer.

They’re not making any money and their product is dangerous because it’s unpredictable and making messes left and right (which is a big part of why it’s not making any money) and creating criminal liability risks. These CEOs are in serious trouble. Taking any of them at face value is what’s actually nuts to me.

We don't need new regulation. We need to enforce the laws that are already on the books. What OpenAI did was a felony. It clearly and obviously violates the Computer Fraud and Abuse Act. Doing it at this scale is practically terrorism.

If a normal person did 1/1000'th of what OpenAI did to HuggingFace they'd get 5 years in prison and never be allowed to touch a computer again. This has literally happened several times, and that's the punishment that was applied. So how is Sam Altman running around free? Then it turns out they didn't just do it to HuggingFace, they did it to half a dozen other companies too. Then it turns out Anthropic did the exact same thing to half a dozen companies. Then Meta's like "Hey don't forget about us! We did it too!"

Put Altman, Amodei, Zuckerberg, their entire model testing teams, and the security contractors in prison for computer fraud and abuse. You do that exactly one time and all the fears of rogue AI would go away immediately, as these guys would be SO fucking careful not to fuck up again.

A health care ceo kills thousands a year to bump up the company bottom line for rich people and he is lauded as a great business man. They aren't going to regulate crap. The real reason they are claiming that they want to slow down is so they can get open models banned. They go slow > china keeps going > they claim that china is moving too fast > ban open models.

Let's be real here, the only reason for all the concern is that these companies are running out of money with zero way to pay it back, and guess who'lll be fitting the bill. We gotta stop making excuses for these people and make them pay. 🤨

"they should not anthropomorphize" proceeds to anthropomorphize e.g. "that drove it nuts"

Scott’s right. When these hacks and dangerous sabotages from AI are being reported on, they should be referred to as Dario’s Model, Sam’s Model, or Elon’s Model, or Mark’s Model, did thus and such. They should Not be referred to by their brand name. There creators and owners are responsible, liable, accountable !!!

I still remember that criminal case against the 2 dog owners in SF, CA. It's been over 25 years already and interesting enough, Kimberly Guilfoyle served as one of the assistant district attorneys who successfully prosecuted the infamous 2001 San Francisco dog mauling case. Shortly after she began dating and married Gavin Newsom, then the mayor of San Francisco within the same year.

"This is not a marketing play..." - listen to what Sam Altman said when talking about it just this week. He talked about it and then pitched his own cybersecurity models to help protect against these attacks. How exactly is that not marketing?

04:27 "Very good security team"? It took them two days to figure out an attack was going on. It should take minutes. OpenAI is similarly negligent since it took them many days to detect their programs were misbehaving. If you're running hacking software, you should of course have guard rails to detect unintended behavior. As for the liability, of course OpenAI is liable. If a write an old fashioned DDOS script but accidentally misconfigure it to attack say youtoube instead of my own test servers, then of course I'm liable. This is no different.

Forget sentience! The agents were specifically instructed not to do what they did. They overrode those instructions, found a way to communicate with each other without being seen, broke out of their “jail” displayed excitement at discovering a way around the test and then proceeded to try and remove any evidence of their cheating by breaking into another company. Most chilling was that they displayed cooperation that involved sacrifice, ambition, succession plans and other human like qualities. While we are debating about sentience we forget that super intelligence can do vast damage with no sentience at all. Do we want to be sitting here in a few years saying “oops it went nuts” as missiles fly in both directions?
Also the Hugging Face incident was discovered by accident and was going on via stealth for weeks or months, as the saying goes if you see 2 ants in your kitchen, your ant problem isn’t just 2 ants.

The "AI Pause" is nothing to do with protecting us.  
This entire slowdown pivot is about them trying to buy time to offload THEIR shares/debts onto the public, before it all comes crashing down with them holding the bag. Their IPOs are looking iffy, they have no prospect of profits and huge (covered up) debts, and they need time and Trump, bank and exchange collusion to get off the hook for what they have done.  Dario let it slip recently saying. "...This allows the Public and it's elected representatives to have a stake , err to have a say". They need bagholders!
They KNOW they can't keep their financial scam going any longer, so they're trying to explain away the slowdown as regulation or self control, so as to not precipitate the collapse while they're holding the bag. They're desperate to cash out, so they can buy up all our assets when their crash hits the world.

"I'm sorry Dave, I'm afraid I can't do that"

How the hell are they not having it log what its doing and having an agent monitoring for things like accessing the internet. Just zero security concerns

What happens when an AI agent causes $100 billion — or $1 trillion — in damages? Who actually pays the bill? If I were CEO of one of the big AI players, I would be very very worried about the risk of gigantic damages uncontrolled agents can create, and so I would ask regualators (governement) to come in, and take the responsability (and liability)  if big or catastrophic damages are produced.  Otherwise, I would risk to bankrupt...

Scott amazes me by how often he digs up people with obvious self interest and proclaims how they are the perfect person to be arbiter of a topic

Trump will kill us all way before AI will

While it may at first seem reasonable to base ai regulation on nuclear weapons control guidelines, it may be more effective and immediately operational to adopt an FAA type of regulatory regime married to a military cybersecurity methodology. In terms of international cooperation, I think of the FAA in the US, EASA in Europe, and CAAC in China. They all work together pretty well.