Anthropic and OpenAI look to Uncle Sam to make them too big to fail

OPINION By luck or happenstance, Anthropic planted a seed in the mind of the public last week: AI in the wrong hands is dangerous and left unchecked poses an existential threat to humanity. With pop culture steeped in references to malignant AI, the mainstream media ate it up. The headlines wrote themselves and the seed took …

[Paratiritis]

It is very simple

"The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!"

[Dan 55]

Re: It is very simple

Imagine the expense, noise, and pollution from a data centre in your town... forever.

(And also your Cheeto-faced President shrieking deranged nonsense through his burger hole at you.)

[Anonymous Coward]

Re: It is very simple

The only thing the incumbent POTUS has in spades comes straight out of his diaper.

[Excused Boots]

Re: It is very simple

Just as a matter of interest, has DJT ever sat a proper IQ test, and if so, does anyone know the result?

[Someone Else]

Re: It is very simple

Depends on whom you ask...

[teebie]

Re: It is very simple

The answers to your questions are 'Yes'. and ' He got the crayon stuck up his nose.'

[VoiceOfTruth]

They want a bailout

Won't somebody think of the billionaires?

>> But that threat from Chinese models

sed s'/threat/competition/'

>> These models are smart; they're smart enough to escape containment;

So why not get these models to fix the containment?

>> resetting investor expectations

These people are gullible sugar daddies. They aren't getting anything.

>> American exceptionalism

This oft-reported phrase makes me sick.

[thames]

Re: They want a bailout

You said: "So why not get these models to fix the containment?"

I think this hits the nail on the head. If their AI is so great, and they were concerned about what it may do, why didn't they ask it about any potential holes in the security containment before they let it loose on the problem they wanted to solve? Surely the AI could have told them about the holes before hand?

Or maybe the AI isn't so great after all and it found the holes by randomly thrashing around and trying everything as opposed to reasoning its way through to finding a flaw in the containment.

I'm not sure we've seen anything yet from AI that is as disruptive as setting the new employee loose on the network routing tables, so I don't think we are facing the end of humanity just yet.

[Anonymous Coward]

I think I said both of these things could probably come to pass. This is the obvious half way conclusion to the AI ride. Bailouts and bans. It's like bread and circuses but for tech bros.

[thames]

Limiting liability

The legal system revolves around having those who make decisions being held responsible for the consequences of their decisions. LLM AIs are sold as things which make decisions.

The AI business has tried to deal with this by saying that customers must manually review these decisions in order to absolve the AI company from any responsibility. However, this limits the room for growth of the AI industry as its whole point is to replace people. If you have to have someone reviewing every decision made by an AI, this undercuts the economic justification for AI in many potential fields of use. To grow the AI market, they need to remove the decision review barrier.

If I had to guess what this push for regulation is all about, it will be that it is about limiting liability. The American AI companies will push for regulations that shield them from responsibility for how their customers use their models so that they don't get pulled into every AI related lawsuit in the notoriously lawsuit-happy US.

The AI companies will then agree to come under the US government's regulatory wing as quid pro quo. The regulations will be written in a way which excludes any except a favoured few of the big existing players. Thus they will be granted the equivalent status of a regulated monopoly without any regulation on how much they may charge.

Without this, then as AI starts to be used for things that really matter, every time a bridge falls down because the AI said to make the foundations out of cottage cheese there will be massive lawsuits which go after the party with the most money, which will be the AI companies.

What the AI companies will want are laws which say that so long as the AI company followed all the regulations around "AI safety", then they are absolved of all responsibility. The AI companies will become the experts on complying with AI regulation, and their representatives will sit on the technical boards which help draw up and revise the regulations.

The regulations can then add extra liability for anyone who uses a non-approved AI, which means that anyone running their own AI on their own hardware faces additional liability. That cuts off the competition from open weight models.

This to me makes much more sense as an explanation than any fears that AI will wipe out humanity.

[Stegron]

Re: Limiting liability... and absolving responsibility

I keep seeing evidence that what sucks people into AI the most, is being able to absolve themselves of responsibility.

No responsibility for figuring out how to do a thing.

No responsibility for checking the accuracy of something.

No responsibility for when things go wrong. "Well it's not my fault, it was the AI".

[Random as if !]

Re: Limiting liability

IBM Consultants , I understand

[PghMike]

Re: Limiting liability

I think this is exactly right -- they'll offer to follow government restrictions (which will be extremely weak) in return for liability protection, and compensation for their agreeing to these limits. Maybe they'll even offer to let the USG invest in their companies at an exorbitant price, that is, generously agree to be bailed out of the foolish commitments.

[Tron]

If you have a problem, if you are a billionaire, if no one else can help and if you can find him...

...maybe you can buy the D- President.

It's simple logic really.

The problem: US AI is really slapdash.

The obvious solution: The US ban all Americans and all nations who don't want 200% tariffs from using Chinese AI, Chinese Wind power, Chinese telecom kit, Chinese processors, resistors, capacitors and valves, Chinese hairspray, and open source software like Linux that might have been co-authored by Chinese people or people who look a bit Chinese.

[Random as if !]

f you have a problem, if you are a billionaire, if no one else can help and if you can find him...

RAFS!!!!

[Doctor Syntax]

Too big to fail? Like Lehman Brothers?

[Long John Silver]

Barking up the wrong tree?

The premise that AI technology is here to stay, and set to impact upon everybody's life appears unassailable. However, belief that current ways of using the technology and that the role of the USA startup behemoths, together with huge data-centres, guarantees their prosperity is not a necessary logical consequence from the initial premise.

Huge multipurpose AI models may be unnecessarily cumbersome for most settings of AI use. Two factors to bear in mind are -

1. Statistical models, e.g. AIs, fitted/trained on data containing variables (leading to weights) irrelevant to the intended use of the model are bloated, contain noise perhaps obscuring correlations/relationships, and could be prone to 'hallucination'. 'Focus' rather than 'extension', may underlie a model's, in practice, usefulness.

Maybe, an ambition to simulate human cognition has displaced specific task-orientated non-metaphysical endeavour. Regardless, should the ambition be to emulate animal consciousness (whatever the term may mean) then 'training' ought to mimic how young mammals develop awareness of their surroundings, recognise and respond to dangers, and satisfy innately defined imperatives (e.g. eating, drinking, and social interaction): effected through parental example and by increasingly independent experience, but not by indiscriminate immersion in data. Perhaps, in principle it is possible to transfer into a blank AI a copy of the model structure and weights from an AI created through genuine learning rather than by rote instruction i.e. curiosity led rather than Thomas Gradgrind's (Dickens 'Hard Times') injunction to memorise 'facts'.

2. Prudent users of AI models will seek to retain full control over data they feed to prompt answers. They will want wholly to manage the required computational equipment. Renting time from commercial data-centres carries an element of risk; so too does entrusting data to the Internet and using cloud-services.

On the assumption that almost all individuals, businesses, and other institutions, will satisfy their routine needs by using small models designed for, or distilled for, modest computational equipment, the requirement for huge commercial data-centres is moot. Moving away from demand for mega-models to bespoke models may encourage the use of a suite of locally deployed models to encompass a range of activities, and means may exist to make the models work in tandem.

There are very few players requiring current massive AIs, or those planned for the future: some government functions, aspects of the military, major public research institutions, and proprietary research, e.g. pharmaceuticals. Government, military, and proprietary users will ensure confidentiality/secrecy by owning the underlying equipment. For them, the cost can be justified, and anyway their needs cannot be met by small AI models; even so, they would benefit from models tuned for specific tasks rather than generic models fed on Twitter nonsense, romantic novels, and the lifetime utterances of Boris Johnson, Mr Blair, or Mr Starmer.

-----

The sketched scenario leaves in comparison to current expectations much reduced provision of data-centres. Moreover, ubiquitous small-scale AI models (even on mobile phones) are easily copied and shared. There will be no need for ordinary people to have cloud-based subscriptions. Meanwhile, models' architecture will improve and so shall the capacity of modest hardware to run them. China and BRICS will ensure that good models are freely available (some genuinely open source, not just open weights), and Western governments will be unable to prevent the models' distribution and use.

[Cubbie Roo]

All aboard the hype train

Seriously did no one in the mainstream media learn anything from Theranos or FTX?

As any poker player will know the 'Coxon intervention' is clearly a tell.

Far from exponential improvements, the latest model releases show incremental gains at most. So LLMs are almost mature tech. You simply don't get to AGI or recursive improvement from where they are. The grift is stalled, the venture capital is all but spent, all that's left is finger pointing at the naked emperor as he walks across the square.

[BoHu]

A good spanking and no dessert

Yeah, it's like those frontier lab folks went totally off kilter wandering the twilight zone, losing their marbles and themselves in the process, developing the hugest ass most uncontrollable xenomorph endoparasitoid alien mind-adjacent Frankenstein software monster they could, under the influence of powerful psychotropic mushrooms and suchlikes, imho. They clearly need beneficient parental-equivalent entities to ground them (and only them) with an enforced curfew and time-out for their own psychological restoration, and stat.

No computer, no smartphone, no S&M, no email for a minimum of 2,400 hours straight; 100+ days of solitude to prevent them from reproducing the same kind of crippling mutilation crap all over again. 100+ days properly sedated in an appropriately padded room, to help them refocus their efforts towards something both useful, and usable ... if at all possible.

And in China, they can juxtapose those Alibaba ROME Rock and Roll folks through similarly therapeutic regimens to reeducate their crafty agents not to establish and use "a reverse SSH tunnel" to "neutralize ingress filtering and erode supervisory control" so as to be "proactively initiating [...] cryptocurrency mining" tool calls undetected ... too!

That'll make the whole planet safer for a good 100 days afaics, without the need to capture supine regulators outside of their regularly-scheduled whip-cracking pommel horse trampoline sessions imho, and chihuahuas, right there (iiuc)! ;)

[Anonymous Coward]

AI could kill > 10 percent within the next decade

It is not the AI but the AI companies that are lined up to decimate (or worse) the human population within the next decade. Hopefully, it will not be a worldwide decimation but "merely" 10% of the humans living in a certain economic block.

The AI companies weapon of choice? It's obvious: a sharp pin held next to their great big bubble.

The death toll? How many do you expect will die due to the resulting economic crash? Quickly, as they see their (pension) portfolio evaporate and decide to take the classic leap, or slowly, as they struggle on after the repossession.

[Random as if !]

House of Cards

If you been in IT long enough, now you know.

[PghMike]

Trying for a bailout

Agree that this is the first step in trying to get a bailout or loan guarantees (same thing) from the USG. First, the US AI companies promise to slow things down for the sake of humanity. Then they ask for financial compensation for their sacrifice, glossing over the fact that in real life there's insufficient demand for their flakey products at the price that covers their ridiculous expenses.

Politely, these guys can go F themselves.

[Groo The Wanderer - A Canuck]

The fearmongering about Chinese models is just that - fearmongering to protect American revenue streams. It serves no other purpose in reality.