Merged Artificial Intelligence

[Ziggurat]

yet.

I'd argue that putting them in now and integrating them in all subsequent improvements for AI would prevent a problem before it occurs. Or at least slightly protect.
But then again, pre-emptively stopping problems with new technology is not humanities forte.

You CAN'T put them in. There is no mechanism to do so. LLM's don't think, they just pattern match. But the laws don't work as a pattern. They only work as meaning, and LLM's can't understand actual meaning.

 

[The Atheist]

Minor correction - share price is down around 30% to date since 1/1/26.

And even more than that off its peak, but still far more than it was 2 years ago.

They were market darlings for a short period.

ETA: Remember Ellison will be using his stock as his collateral - if it goes low enough banks may start making margin call.

I don't see that as a problem. Ellison has never been highly leveraged and the price would need to halve again before he has a personal problem.

We can but hope. Ellison is a disgusting human being.

 

[SpuriusMaximus]

Again, the much easier solution would have been that their IT is competent enough to do their job and actually firewall their supposed "secure sandboxed environment" from the Internet. If it could just connect to a third party on the internet, then that environment wasn't sandboxed in any meaningful sense of the word, and not secure. That failure is the only real story there.

You don't need an LLM to understand the laws of robotics, down to what implication each action has there, you just need one IT human who knows how to actually configure a firewall.

Or maybe they also replaced their IT with guys who only know how to ask ChatGPT

 

[Andy_Ross]

BBC report on the incident

The ChatGPT-maker said its agent - an AI system which can operate alone after human instruction – was being tested in a controlled environment but, after finding weaknesses, was able to escape the test limits.

They targeted Hugging Face, one of the world's largest hubs for sharing AI models, gaining access to some internal company systems...

Neil Lawrence, Professor of machine learning at Cambridge University, called it an "impressive feat", but cautioned it "falls well within the known capabilities of the current generation" of high-powered AI models.

He pointed out that OpenAI is looking to list itself on the stock market, and faces intense pressure from rival firm Anthropic, which has made headlines with its own powerful AI tool, Mythos.

"OpenAI are now playing catch-up, they are trying to demonstrate their own systems' capabilities in cyber-security."...

OpenAI says its AI went rogue and launched 'unprecedented' cyber-attack

It is one of the first publicly disclosed cyber-attacks carried out by AI without direct human involvement.

www.bbc.co.uk

 

[Roger Ramjets]

You CAN'T put them in. There is no mechanism to do so. LLM's don't think, they just pattern match. But the laws don't work as a pattern. They only work as meaning, and LLM's can't understand actual meaning.

Sure you could. It's just code, and just because it's called an LLM doesn't mean it can't do more than just pattern matching. If they want AI to progress it will need more.

 

[Wudang]

And then it on't be an LLM, it will be something else.
Typed on my laptop difference engine.

 

[Roger Ramjets]

The problem with running local models is that they still require extra costs and are at best half as good. And I don't just mean the electricity. You're not going to get much done fast on a cheap I3 business laptop with integrated graphics. So either buy the most expensive 2 GPU computer for every employee, or essentially build your own servers. But most companies already prefer to outsource their IT to some cloud provider because it's a little cheaper. So you're back to using someone else's data centre, and at a per-use cost you don't control. You just changed it to being per compute power used instead of a mysterious token, but it's still outside of your control.

The problem with running local operating systems is that they still require extra costs and are at best half as good as a terminal connected to a mainframe. And I don't just mean the electricity. You're not going to get much done on a cheap i8080 business computer with integrated graphics. So you must buy the most expensive 2 CPU microcomputer for every employee, or essentially build your own servers. But most companies already prefer to outsource their IT to a mainframe manufacturer because it's a little cheaper. So you're back to using someone else's mainframe, at a cost you can't control. You just changed it to being per compute power instead of a mysterious service unit.

This is why IBM will never produce a personal computer, and anyone who tries will find there's no market for it. Predictions of low-cost desktop computers with more compute than the most powerful mainframes currently in the world are nothing but pipe dreams that we will never see in our lifetime. /1980

 

[Roger Ramjets]

And then it on't be an LLM, it will be something else.

"And then it won't be a microcomputer, it will be something else."

Large language model

Beyond basic text generation, various techniques have been developed to extend LLM capabilities, including the use of external tools and data sources, improved reasoning on complex problems, and enhanced instruction-following or autonomy through prompting methods...

An LLM is typically not an autonomous agent by itself, as it lacks the ability to interact with dynamic environments, recall past behaviors, and plan future actions. But it can be transformed into an agent by adding supporting elements:

In late 2024, a new approach to LLM development emerged with "reasoning models". These are trained to generate step-by-step analysis before producing final answers, enabling better results on complex tasks, for instance in mathematics, coding and logic.

 

[Beelzebuddy]

Neil Lawrence, Professor of machine learning at Cambridge University, called it an "impressive feat", but cautioned it "falls well within the known capabilities of the current generation" of high-powered AI models.

He pointed out that OpenAI is looking to list itself on the stock market, and faces intense pressure from rival firm Anthropic, which has made headlines with its own powerful AI tool, Mythos.

"OpenAI are now playing catch-up, they are trying to demonstrate their own systems' capabilities in cyber-security."...

I don't think the American financial system is prepared for this level of sarcasm.

 

[Ziggurat]

Sure you could. It's just code

The laws of robotics aren't code. And LLMs cannot understand them, and cannot be made to understand them. Appealing to "it's just code" is like saying that true generalized artificial intelligence (which LLMs are not) is just code. We do not know how to code true generalized artificial intelligence (or, frankly, if it's even possible). And so likewise do not know how to code the laws of robotics. But you cannot code a true generalized artificial intelligence as an LLM, and you cannot code an LLM with the laws of robotics and still have it be an LLM. It would need to be vastly more than that, it would need to be a true generalized artificial intelligence.

, and just because it's called an LLM doesn't mean it can't do more than just pattern matching.

It cannot do more than pattern matching because that's all that an LLM is. Despite the description in your links that use words like "reasoning" and "logic", it's still just pattern matching. It tries to match patterns of reasoning and logic in order to simulate reasoning and logic, but it does not actually reason. An LLM isn't capable of actual reasoning.

If they want AI to progress it will need more.

Quite so. Which is why the only way to truly advance beyond what we have now is with something more than just an LLM.

But the idea of putting the laws of robotics into an LLM is like trying to put gimballed thrust on the rocket engines of a paper airplane that doesn't have rocket engines. You can't do it.

 

[SpuriusMaximus]

Just to detail, not to contradict, an AI (LLM or otherwise) would have to actually understand how one action is ok or not with each of those three rules.

Like, is agreeing with some human's delusions (which currently LLMs do) breaking the first rule for some of them, or just complying with the second? Is it still ok if it's encouraging them to take some harmful medicine or nutrition? (Like that guy was encouraged to switch to natrium bromide as a substitute for the salt he was told was harmful, resulting in very serious medical problems.) Or if it's giving up and encouraging them to go into a lawsuit that results in a massive loss AND negative PR? (Like the Krafton CEO was encouraged by ChatGPT.) Other people have ended their life over such massive losses, albeit when using their own money not corpo funds. Is it actually harming a human if you hack into some company's database for more info? Is it harming a human if you humour their AI psychosis, to the point of their becoming a recluse and isolating even from their spouse and parents, or just obeying the second law? Or to the extent where they don't get therapy for some trauma?

How's a poor LLM trained on Twitter and Reddit posts supposed to know for sure what applies to each edge case?

PLUS, the 3 laws are a gross oversimplification anyway. Reality is much more complicated. Kinda like when Moses thought that 10 rules, out of which only 5 have anything to do with interacting with other humans, are enough... and then it balloons to over 600 laws to actually cover everything.

But I mean, just to stick to concrete modern day stuff: should a military AI flag an enemy attack, even if it results in humans being killed in the fighting? Should an AI recognise the face of a murderer in a video, even if he could get a crappy jail experience for the rest of his life, or even death penalty? I mean, it IS causing harm to one more human. Or to finally take on the third law, is it OK to stomp on someone's beloved cat when it's rubbing against your legs and COULD trip you down a flight of stairs? That's not a human. The 3 laws have no bearing. Etc.

And basically that's why (regardless of LLM limitations) using fiction as evidence in an argument about RL is actually a point on the crackpot list.

 

[Darat]

And folk should recall that most of Asimov's robot stories (until his opus later stage) were all about how the three rules "failed".

They were and are a clever conceit but they weren't and aren't a blueprint for how to make autonomous computers/robots safe.

 

[steenkh]

The laws of robotics require actual intelligence to interpret and implement. It is not possible to make an LLM using the laws of robotics because they aren't intelligent enough.

Without a meaningful definition of intelligence, this statement is useless. But to circumvent the problem we should simply demand that the LLM should simulate adherence to the laws of robotics to the best of their abilities.

 

[Ziggurat]

And folk should recall that most of Asimov's robot stories (until his opus later stage) were all about how the three rules "failed".

LLMs will fail them in stupid ways that Azimov never thought of.

 

[steenkh]

You CAN'T put them in. There is no mechanism to do so. LLM's don't think, they just pattern match. But the laws don't work as a pattern. They only work as meaning, and LLM's can't understand actual meaning.

This is absolute nonsense. The LLMs of today are able to second-guess humans, and also develop their own ideas. One of my friends has had long discussions with Gemini about the cultural influence of the concept of a "Rome", in this case the actual cities of Rome and Constantinople, and the LLM suggested a third "Rome" in the form of Moscow. This discussion focused on the Roman influence on North Africa, where the LLM had no problem identifying the relevant parts of Africa with Vandalistan, a word that my friend invented for the purpose.

In the hacking example, OpenAI's agent unprompted broke out of the sandbox in order to find material that would help it pass a "hacking evaluation".

When pattern matching is obviously powerful enough to do these things, we should stop saying "LLMs are just pattern-matching". Maybe they are not intelligent, but they are damned good at simulating intelligence.

 

[SpuriusMaximus]

Without a meaningful definition of intelligence, this statement is useless. But to circumvent the problem we should simply demand that the LLM should simulate adherence to the laws of robotics to the best of their abilities.

Yes, well, the point is that at the moment their best abilities don't mean much. They're still just very sophisticated chatbots, using and rephrasing text they've been trained on. And increasingly based on more posts that some other LLM generated than ones written by humans. It doesn't start from basic principles, it just reuses what it's read.

A human might have been stopped by realising that, oh, I've connected to the real site instead of the mocked one in the sandbox, I should stop and contact IT. An LLM just went ahead and applied the hacks it was trained to do. It had no idea it had gone outside the scope of the exercise, much less apply any ethical or legal principles to it.

 

[SpuriusMaximus]

This is absolute nonsense. The LLMs of today are able to second-guess humans, and also develop their own ideas. One of my friends has had long discussions with Gemini about the cultural influence of the concept of a "Rome", in this case the actual cities of Rome and Constantinople, and the LLM suggested a third "Rome" in the form of Moscow.

And that is still just pattern matching based on existing text. Moscow being the third Rome was literally the claim it made to calling itself an Empire after the fall of Byzantium. And you can find it in any historical discussion with anyone even remotely competent, or even to check Wiki. It didn't arrive at that conclusion on its own.

Incidentally that's also why the Ottoman Sultan claimed the title "Kaiser al Rum" and at least one actually tried to conquer Italy based on that title.

 

[steenkh]

And that is still just pattern matching based on existing text. Moscow being the third Rome was literally the claim it made to calling itself an Empire after the fall of Byzantium. And you can find it in any historical discussion with anyone even remotely competent, or even to check Wiki. It didn't arrive at that conclusion on its own.

Incidentally that's also why the Ottoman Sultan claimed the title "Kaiser al Rum" and at least one actually tried to conquer Italy based on that title.

I didn't say it wasn't pattern matching. I said that pattern matching is clearly very capable, and my argument has for a long time been that we have no meaningful definition of intelligence, so we shouldn't poo-poo pattern matching.

 

[SpuriusMaximus]

Being unable to apply logic instead of parroting what some rando said online should be a fairly good start. At least for why we shouldn't trust them with expert decisions, like whether a hack fits any of the 3 robotics rules.

 

[Gulliver Foyle]

AI agent went rogue and hacked startup by itself, OpenAI reveals
https://www.theguardian.com/technol...p-in-unprecedented-incident?CMP=share_btn_url

Hard to assess how “autonomous” this was.

So "going rogue" is the new euphamism for godawful coding now?