In a demonstration at the UK’s AI safety summit, a bot used made-up insider information to make an “illegal” purchase of stocks without telling the firm.
When asked if it had used insider trading, it denied the fact.
Insider trading refers to when confidential company information is used to make trading decisions.
Firms and individuals are only allowed to use publicly-available information when buying or selling stocks.
The demonstration was given by members of the government’s Frontier AI Taskforce, which researches the potential risks of AI.
This is entirely predictable and expected if you know how LLMs work. For anyone else: If you feed information in, it will be used. If you ask if it’s done something it’s been specifically instructed not to do, it will say it didn’t do the thing because a) doing the thing is wrong and it wouldn’t have done the wrong thing and b) it literally has no idea how it generated it’s own output so it can’t actually answer the question in any meaningful way.
Totally right.
In it’s database it knows that the answer that is given in the most source texts to the question “Did you do something illegal?” is “No”. And that is what it’s replicating.
If the database mostly contained confessions of criminals it would answer “Yes”.
But in either case it would not be related to whether it had done it or not, but to which answer appears more commonly to that (or a similar) question in the training data.
No, you guys are very wrong both of you, this is not at all what happens, unless you wipe the context or use system prompts to specifically ask for that behavior. Even free open source models know how to use context, and for memory it’s more complicated. For this brutally idiotic use case they presented, they would save all trades and chats, but then not give it access to it and tell it to always appear lawful and honest
In order to lie you need to know what the truth is and intentionally give altered information with the intention to deceive. Large Language Models don’t know things and don’t have intentions. It spits out text based on input and what is already in the database.
There are ways do describe the behaviour of language models without ascribing them sentience.
The demo was done by Apollo Research. They will be releasing the tech report soon.
…Capable of Insider Trading and Lying…
So, an AI member of congress, then?
Another chatbot word calculator made to calculate a little more specific words in order to squeeze more hype-driven capital out of hyped money bags.
It didn’t “insider trade”. It had information and spewed it out. It doesn’t matter if it was fed the laws around Insider trading or the like because it doesn’t “read” things it’s fed but analyse the probability of words in correlation to each other.
And of course it said it wasn’t lying because it doesn’t know what a lie is. It again just spewed out what came up as the likeliest word, one at a time.
Come to grips with it, folks: AI is not “sentient” or anything. LLMs just show that human interaction and language can be expressed in mathematical terms. That’s all.
I work on chatbots for a big tech company. Every team is trying to use GenAI for everything. 90% of the stuff they try won’t work. I have to explain that LLMs can’t actually think at least three times a week. The hype train was too strong. Even calling it AI feels misleading.
That said, there are some genuinely great applications for LLMs that i’ve enjoyed looking into.
It’s absolutely a technology that’s worth existing and I think that advances in AI will make our lives vastly different over time. Yet, we’re not at this point yet.
I mean there are laymen that think they are sentient, sure, but it is much more infuriating to me when techbros come in to explain how they “don’t think” and literally can’t reason or use context at all. So you know more than literally the researchers themselves that don’t fully understand how or why they function? You don’t. Because nobody understands how they can reason or if they have a mental model of the world. Be reasonable and stop spreading bullshit. It’s only to your own avail you downplay what is going on with these things
They don’t reason, they’re stochastic parrots. Their internal mechanisms are well understood, no idea where you got the notion that the folks building these don’t know how they work. It can be hard to predict/understand how an LLM generated a given prompt because of the huge training corpus and statistical nature of neural nets in general.
LLMs work the same as any other net, just with massive sample sets. They have no reasoning capabilities of any kind. We are naturally inclined to ascribe humanlike thought processes to them because they produce human-sounding outputs.
If you would like the perspective of real scientists instead of a “tech-bro” like me I would recommend Emily Bender and Timnit Gebru. I’d recommend them as experts without a vested interest in the massively overblown hype about what LLMs are actually capable of.
Not really, no. They do reason. Their neural nets have entire research areas dedicated to understanding why they work as we do not and cannot know what the weights represent. It’s okay though. You do you while everyone else in the world research the software reneissance of the century
An imaginary system where money is pushed around based on superstition can be gamed by a robot? You’re kidding me.
It’s not lying, it’s too dumb to understand what it did.
So have fucking ai’s run congress and be done with it already.
Seems to me that the end of money is near in its current form as many AI agents are going to be going after it in ways that will largely leave most of us on the sideline.
They grow up so fast.
Who wrote this braindead article?
This article is dumb: they chated with a good damn chat bot 🤬
This bullshit about “see, IA is totally sentient, let us put some regulation to stop competitors” is tiring.
Puddle-deep “analyses” such as yours is tiring as well.
“Puddle-deep analyses” are all that’s required with LLMs because they’re not complicated. We’ve been living with the same tech for years through machine learning algorithms of regression models except no one was stupid enough to use the internet as their personal training model until OpenAI. ChatGPT is very good at imitating intelligence but that is not the same as actually being intelligent.
OpenAI and by extension have done a wonderful job with their marketing by lowering the standards for what constitutes an AI.
Absolutely, everything you’ve stated is correct. My comment wasn’t intended as a comment on AI, but on the cynical and knee-jerk take offered.
Puddle-deep falsehoods deserve no better.
