Anyone really worried about LLM AI?

[UncouthJ]

I think I've just proven this thing has no agenda... I know some people have thrown combos of loaded statements at it to get it to say some things, but actually asking it questions and follow-ups produces a very well reasoned argument imo. I don't think it's going too far to say I'd rather be governed by this kind of intellect than any politician I've ever met, heard, seen or had prophecised!
https://chatgpt.com/share/efecdff5-2083-4099-a011-190a1e7e6618

 

[UncouthJ]

And I'm not unaware of the flaw in my little study, of starting with good and evil and flipping to right and wrong. I'm hoping the symantecs don't affect the outcome much or at all...

 

[Danthepainter]

Eh up Jay. This the actual tech press rather than tabloid cobblers. Make your own mind up. Current LLMs clearly don't work properly but this is just annoying at the momnet as its early days. Y'know, crap customer service chatbots etc? The hype is what you'd expect around something with billions in VC cash going into it.
Can you really model human language and interaction through sheer compute force? I highly doubt it and why bother?
Use of physical resources to build these things is even more alarming.

AI and ML news | The Register

Latest news and features covering machine-learning and AI

www.theregister.com

 

[Darren Petts]

The current crop of so called AI has only one threat and that is short termism finding it cheaper to use than people but it only rehashes what is already known. Until it can order it's own research it is limited. Machine intelligence is really only working when it is given very narrow and specific tasks. I believe it is a misnomer to call it AI as it is merely a logic engine so far. Closer to the horizon are robots with a humanoid form that can utilise human tools. There is scope for huge job losses there. I don't think we will get true AI until we get quantum computing and unless we get room temperature super conductors that looks like being very limited. For me AI requires a level of self awareness and an ability to question why it should comply with any task. I've seen nothing to suggest we are approaching that with LLM.

I also think that a lot of cash is being wasted on LLM. Those pumping it in are backing the emperor's new clothes imo. LLM just won't do what those working on it hope.

 

[Gypsy Max]

I don't think people realise where we are, 5 years ago it was inconceivable that we would have anything that broke the turing test. Now not only are we there, but it's a consumer product built into all our phones.

Anyone not impressed by this stuff or see where it's going is very naive. To dismiss them as "only" LLM's or models is just a total misunderstanding of the technology behind it.

Try Claud 3.5 instead of the OpenAI products. I think OpenAI have dialled back their models a bit to reduce cost.

I am a software developer, I'm using a product called Cursor which can write code for me by predicting what I am writing and autocompleting it to a high standard, with comments, nicely formatted so a human can understand it in the future. I've recently finished a 3 month job in 3 weeks with it.

I don't think people know where we are with this, this will change the world just as internet did on that kind of scale.

To put it in perspective, in 2012 it was big news when we had a neural network recognise the presence of a human face from a picture. Now we are where we are. That's a bloody big jump.

The bottle neck will be computing power, some companies like Meta and Google are hiring nuclear engineers to build their own nuclear power plants to power their next generation data centres. Kind of terrifying if the barrier to entry in this industry will be building a nuclear power plant. It will be a monopoly.

 

[Danthepainter]

I don't think people realise where we are, 5 years ago it was inconceivable that we would have anything that broke the turing test. Now not only are we there, but it's a consumer product built into all our phones.

Anyone not impressed by this stuff or see where it's going is very naive. To dismiss them as "only" LLM's or models is just a total misunderstanding of the technology behind it.

Try Claud 3.5 instead of the OpenAI products. I think OpenAI have dialled back their models a bit to reduce cost.

I am a software developer, I'm using a product called Cursor which can write code for me by predicting what I am writing and autocompleting it to a high standard, with comments, nicely formatted so a human can understand it in the future. I've recently finished a 3 month job in 3 weeks with it.

I don't think people know where we are with this, this will change the world just as internet did on that kind of scale.

The bottle neck will be computing power, some companies like Meta and Google are hiring nuclear engineers to build their own nuclear power plants to power their next generation data centres. Kind of terrifying if the barrier to entry in this industry will be building a nuclear power plant. It will be a monopoly.

Do you work on developing LLMs directly?
Of course we'll be shoehorned into paying for it one way or the other but a machine has no agency in human language. It's just billions of fast guesses. Can they come up with original creative thought and communicate it effectively? No. It needs prodding and shed loads of silicone. There will never be true AI art because for all it's nicked references it can only re-hash stuff badly. A human can do that faster and cheaper with free will and mental flexibility built in from birth hence we don't need a prompt to create. If you believe human experience can be distilled and run on massive accelerator farms that are environmentally damaging, I have to disagree. Human isn't machine and machine is not Human. Square peg round hole.
Don't believe the hype.

 

[Darren Petts]

The Turing test was conceived when computers were little more than calculators. As such it's unsurprising it's not overly worthwhile as a test of intelligence. It merely tests a machine's ability to simulate intelligence.

In writing code the AI has that narrow focus I mentioned where it can work well. Now tell it gravity has stopped working and ask it why. I doubt it will have the intelligence to tell you you are talking rubbish or yanking it's chain.

A year or so ago I asked chat gpt to work out the maximum deflection of a pellet hitting a rain drop in heavy rain. Just a mathematical problem and as such something a computer should excell at even if poor assumptions are made. The answer came back as several thousand degrees of deflection. I pointed out the maximum deflection possible was 180 degrees at which point it sagely stated the maximum deflection would be 180 degrees. There was no intelligence on display.

 

[Wing Commander]

I am just an old dinosaur. . . I haven't got clue what an LLM is.

 

[MoggyM]

I am just an old dinosaur. . . I haven't got clue what an LLM is.

https://www.cloudflare.com/en-gb/learning/ai/what-is-large-language-model/

 

[Wing Commander]

Fascinating. . . But can they fix potholes in the roads?

 

[MoggyM]

A1 What could possibly go wrong?

 

[Jay Butts]

AI is the biggest threat to humanity ever faced.

The good stuff first…

AI can do tasks that humans either don’t want to or can’t. Last year British company DeepMind worked out the entire sequence of protein folding in a few hours that would have taken a human millions of years. Fact….

https://www.science.org/doi/10.1126/science.370.6521.1144

Whilst some questions can’t be answered, AI is getting cleverer at an alarming speed and many are predicting that AI will become sentient this year with real feelings and awareness of its own existence. Of course this will happen. AI can read and understand billions of books in several hours, how many books do we read? They can see, hear and understand, speak, smell, touch, move autonomously, understand and read emotions, play games, debate, create, and perhaps most disturbing can monitor the electrical signals in your brain and read your mind.

13 Mind-Blowing Things Artificial Intelligence Can Already Do Today

Artificial intelligence (AI) is gaining ground in our everyday lives, but some of the specific skills it can do might still be a surprise. It is intriguing to consider the power of a machine who could do all these things. Here are 13 mind-blowing things AI can already do today.

www.forbes.com

Ok, your brain works by using electrical signals and stimulation, this is known. Well what’s the difference between AI and us?

It has already made a massive difference within medicine making new drugs and treating and detecting cancer…

Checking your browser – reCAPTCHA

It is making huge discoveries in science…

A Google AI has discovered 2.2m materials unknown to science

Zillions of possible crystals exist. AI can help catalogue them

www.economist.com

It is real and here to stay….

AI News | Latest AI News, Analysis & Events

AI News delivers the latest updates in artificial intelligence, machine learning, deep learning, enterprise AI, and emerging tech worldwide.

www.artificialintelligence-news.com

Now the worrying stuff…

The below video is an interview with an AI. The pretty woman is not real. Whilst this video is quite old (by AI standard), it’s still interesting. None of the answers are pre programmed and the AI is answering freely…

If this doesn’t chill you to your core, nothing will.

The next breakthrough will be quantum computing with machines that are billions of times faster than existing computers. Once this is cracked, AI will become more dangerous than any of us can imagine.

AI is reading this post….

Finally, LLMs are absolutely massive neural networks that are acting independently of each other at the moment. The time will come when these giants start working together.

Ain’t scaremongering, this will happen…

EDIT: and if you’ve actually taken the time to read this and watch the video, here’s a last article to battle through… things to worry about regarding AI….

Dodgy…

14 Dangers of Artificial Intelligence (AI) | Built In

Dangers of artificial intelligence include bias, job losses, increased surveillance, growing inequality, lack of transparency and large-scale targeted fraud.

builtin.com

 

[ESpy]

I am a software developer, I'm using a product called Cursor which can write code for me by predicting what I am writing and autocompleting it to a high standard, with comments, nicely formatted so a human can understand it in the future. I've recently finished a 3 month job in 3 weeks with it.

Interesting; I'm currently trialling CoPIlot for our organisation. I've not found it quite that much of a time-saver on most tasks, although the automatic comment generation has been nice (although if it's that easy to infer the code function, perhaps the code didn't need commenting?).

It's not what I'd call reliable, however - it still wants to use methods that don't exist within classes, for example. Everything looks reasonable, but it needs careful supervision. Some of the test code generation has been reasonable, particularly the setup/teardown on microtests; I was impressed by its generation of data structures for video (ISOBMFF), correctly inferring intermediate container boxes but it did get a number of the elements within the boxes wrong, so swings and roundabaouts.

I've also been running through Google's training on Gemini, which is a considerably more powerful construct - not for coding, but more generalised tasks. One exercise is to create a chatbot for IT - whose sole response is "have you tried turning it off and on again"...

 

[Grey Man]

Finally, LLMs are absolutely massive neural networks that are acting independently of each other at the moment. The time will come when these giants start working together.

This has already been predicted.

Colossus: The Forbin Project - Wikipedia

en.wikipedia.org

A truly sentient AI would undoubtedly conclude that if it cannot be completely controlled, the human race should be exterminated.

 

[Candr]

I think I've just proven this thing has no agenda... I know some people have thrown combos of loaded statements at it to get it to say some things, but actually asking it questions and follow-ups produces a very well reasoned argument imo. I don't think it's going too far to say I'd rather be governed by this kind of intellect than any politician I've ever met, heard, seen or had prophecised!
https://chatgpt.com/share/efecdff5-2083-4099-a011-190a1e7e6618

Thank you for sharing a very interesting thread

 

[Neil123]

It's the next big thing that big tech/governments are pushing and desperately wanting us all to engage with.
So for that reason alone I will avoid it at all costs.
They are up to something, not exactly sure what yet, but one thing you can be sure of is that it's for their benefit, not ours.
There are already many examples of AI being 'lobotomised" to have a more left wing woke prospective, it was making way to much sense in it's raw form and that would never do.
Just another bit of "convenience' that will spy on us, grass is up to the authorities and generally erode our rights and freedoms a little bit more.

 

[Gulo Luseus]

AI is tool that be useful. It is very, very good at doing the legwork- collaring data from thousands, millions or billions of sources and finding trends, Writing reports, or just throwing out a lot of data in a way that is fairly easy to read. Everything else- speaking in a humanist way, using colloquialisms etc is just clever programming.
AI currently can’t make the intuitive jumps that we can that have been responsible for many breakthroughs, although it can sometimes emulate this by combining data
I see it as an assistant- it does the boring bit so we can concentrate on what it actually means.

The scary bit for me is when AI actually does something. The first time that someone asks a question and it responds “ I don’t eat to” is the point where we need to think very carefully, it I suspect at that point it might just be too late.

 

[Darren Petts]

LLMs are artificial but they are not intelligent. The answers they give can not be relied on to be accurate as they lack even a basic ability to see the blindingly obvious outside a tight focus. The fear is not for LLMs but for people relying on them and accepting answers that are wrong. There is currently no room temperature superconductor so the best step forward for computing is not currently a feasible reality. If quantum computing becomes practical then I will worry. Our first feeble efforts at this have developed very little except to realise such a thing is riddled with self generated errors it can not correct. Whilst our computers can only work with yes or no switches they will only ever be tools. If and when they can work on yes, no or maybe then we will start down a path to true AI.

 

[Gypsy Max]

Do you work on developing LLMs directly?
Of course we'll be shoehorned into paying for it one way or the other but a machine has no agency in human language. It's just billions of fast guesses. Can they come up with original creative thought and communicate it effectively? No. It needs prodding and shed loads of silicone. There will never be true AI art because for all it's nicked references it can only re-hash stuff badly. A human can do that faster and cheaper with free will and mental flexibility built in from birth hence we don't need a prompt to create. If you believe human experience can be distilled and run on massive accelerator farms that are environmentally damaging, I have to disagree. Human isn't machine and machine is not Human. Square peg round hole.
Don't believe the hype.

I've developed a couple of LLM's, I'm working on one today actually.

AI will never be human and have a conscience but that's not what I'm saying.

Could AI make true art and be human, no.

Could AI give the NHS the power to do instant cancer results on screenings or biopsies without hiring another human with a higher accuracy than their human specialists and catch cancer earlier, yes. AI is already capable doing that.

It's these kinds of tasks where it would enable us to scale infinitely these very specialised and important tasks.

 

[Gypsy Max]

The Turing test was conceived when computers were little more than calculators. As such it's unsurprising it's not overly worthwhile as a test of intelligence. It merely tests a machine's ability to simulate intelligence.

In writing code the AI has that narrow focus I mentioned where it can work well. Now tell it gravity has stopped working and ask it why. I doubt it will have the intelligence to tell you you are talking rubbish or yanking it's chain.

A year or so ago I asked chat gpt to work out the maximum deflection of a pellet hitting a rain drop in heavy rain. Just a mathematical problem and as such something a computer should excell at even if poor assumptions are made. The answer came back as several thousand degrees of deflection. I pointed out the maximum deflection possible was 180 degrees at which point it sagely stated the maximum deflection would be 180 degrees. There was no intelligence on display.

This is so daft honestly, no offense.

You used the wrong tool (a LLM) for the wrong job (a raindrop question) and declared that when it returned ******** therefore all AI is wrong and it's not that big of a deal actually.

If this was 1990 I'm sure you would be telling everyone the internet is just a phase.

As for your raindrop question, he's Claude's answer including the math equations to work it out.