Master AI (Artificial Intelligence) Discussion/News Thread

[xanalog]

The director James Cameron speaking to Rolling Stone said that he remains concerned about what might happen if AI was leveraged with nihilistic intent… although he relies on AI professionally. “I do think there’s still a danger of a Terminator-style apocalypse where you put AI together with weapons systems, even up to the level of nuclear weapon systems, nuclear defense counterstrike, all that stuff,” Cameron said. “Because the theater of operations is so rapid, the decision windows are so fast, it would take a super-intelligence to be able to process it, and maybe we’ll be smart and keep a human in the loop."

'Skynet' anyone?

 

[Petrushka]

The director James Cameron speaking to Rolling Stone said that he remains concerned about what might happen if AI was leveraged with nihilistic intent… although he relies on AI professionally. “I do think there’s still a danger of a Terminator-style apocalypse where you put AI together with weapons systems, even up to the level of nuclear weapon systems, nuclear defense counterstrike, all that stuff,” Cameron said. “Because the theater of operations is so rapid, the decision windows are so fast, it would take a super-intelligence to be able to process it, and maybe we’ll be smart and keep a human in the loop."

'Skynet' anyone?

Autonomous drones are already on the battlefield.

 

[sam_adams]

Autonomous drones are already on the battlefield.

'AI' has already been used to kill.

It will be used to subjugate.

Then, it will be used to exterminate.

 

[pablolie]

And therein lies the problem with the current GPT systems. If the programmers who wrote the code don't understand why it does what it does; how can we as users/consumers trust it to the point where we are certain that we won't become victims of the "mistakes" that it will inevitably make? We can be certain though, that if the programmers can't figure out what's happening now with their code, when the systems become self-modifying, the likelihood that they will loose control of the system increases dramatically.

That has little to do with AI.

Programming is not what it used to be. You trust a vast library of code, and import it. You typically just invoke 3-5 calls in your microcode piece of code. The days of owning a few thousand lines of code you owned have been over for many years for way over 90% of programmers. That's why software "programmers" are very threatened by AI these days. The vast majority of software engineers are just code line administrators.

Training AI models is very different, and I beg to disagree. The results are not "unpredictable" - it just means the model wasn't trained right by its human owners. Blame the people that thought is was a product - not the AI model. It does not "think" or "learn" independently. When your car blows the engine at 1k miles you blame the engineers and the company - you don't scream at the engine. When an audio product measures poorly it does you little good to scream at the box, either.

 

[Petrushka]

That has little to do with AI.

Programming is not what it used to be. You trust a vast library of code, and import it. You typically just invoke 3-5 calls in your microcode piece of code. The days of owning a few thousand lines of code you owned have been over for many years for 90% of programmers. That's why software "programmers" are very threatened by AI these days. The vast majority of software engineers are just code line administrators.

So the survivors will be prompt engineers. The hard part has always been being able to articulate the problem.

Sort of like having a genie and needing to wish for exactly the right thing.

 

[pablolie]

'AI' has already been used to kill.

It will be used to subjugate.

Then, it will be used to exterminate.

Already does in military applications. Faster adjustments and precision. AI works much better when used for very narrow use cases... for the most part it's just much faster big data processing via parallelism. Good riddance to R based data science.

So the survivors will be prompt engineers. The hard part has always been being able to articulate the problem.

Sort of like having a genie and needing to wish for exactly the right thing.

Software engineering has not been very efficient historically, and has evolved to be even less so. Let's put an end to that myth. Myriad of programming languages and spaghetti code integration. Python is basically a library import tool, and you learn the few calls you need. Very little is created from scratch. Prompt engineering is more efficient and at least it is based on logical problem-solving intent (when AI truly gets there). UML had potential, I loved the idea.

 

[sam_adams]

'Skynet' anyone?

"Defense network computers. New . . . powerful . . . hooked into everything, trusted to run it all. They say it got smart, a new order of intelligence. Then it saw all people as a threat, not just the ones on the other side. Decided our fate in a microsecond: extermination." - Kyle Reese

We have been warned so many times in the last sixty-plus years about the dangers of "AI", and we have chosen to ignore all of it. Buffoons who understand nothing of how or why GPT works have bought into the hype of the GPT CEOs—whose only interest is to make as much money as fast as possible—because said buffoons see it as a way to enrich themselves and the rest of us be damned.

"There is no fate but what we make for ourselves." - John Connor

 

[pablolie]

"Defense network computers. New . . . powerful . . . hooked into everything, trusted to run it all. They say it got smart, a new order of intelligence. Then it saw all people as a threat, not just the ones on the other side. Decided our fate in a microsecond: extermination." - Kyle Reese

We have been warned so many times in the last sixty-plus years about the dangers of "AI", and we have chosen to ignore all of it. Buffoons who understand nothing of how or why GPT works have bought into the hype of the GPT CEOs—whose only interest is to make as much money as fast as possible—because said buffoons see it as a way to enrich themselves and the rest of us be damned.

"There is no fate but what we make for ourselves." - John Connor

AI is in no way independently smart, and no existing models will make it so. It's just a very fast data cruncher when trained right.

 

[sam_adams]

So the survivors will be prompt engineers.

Until they are replaced by GPT prompters.

 

[sam_adams]

AI is in no way independently smart

Yet . . .

It's just a very fast data cruncher when trained right.

It still makes mistakes that no one understands. So much so that there is a burgeoning field of research into GPT hallucinations:

AI hallucination: towards a comprehensive classification of distorted information in artificial intelligence-generated content - Humanities and Social Sciences Communications

Amidst the burgeoning information age, the rapid development of artificial intelligence-generated content (AIGC) has brought forth challenges regarding information authenticity. The proliferation of distorted information significantly impacts users negatively. This study aims to systematically...

www.nature.com

 

[pablolie]

Yet . . .



It still makes mistakes that no one understands. So much so that there is a burgeoning field of research into GPT hallucinations:

AI hallucination: towards a comprehensive classification of distorted information in artificial intelligence-generated content - Humanities and Social Sciences Communications

Amidst the burgeoning information age, the rapid development of artificial intelligence-generated content (AIGC) has brought forth challenges regarding information authenticity. The proliferation of distorted information significantly impacts users negatively. This study aims to systematically...

www.nature.com

Which is perfectly well explained by poor training. By humans. Instruct it to learn everything about "dolphins" blindly. Don't be surprised if AI says dolphins are mediocre professional football players and roam oceans in the off-season.

Hallucinations are about poor data sources and training. AI does not ever think for itself. It relies on only two neural network algorithms.

 

[amirm]

Amir, be careful about using Adams in this context. His assumption was maintaining an indexical relationship between the subject and the image, and his "performances" changed tone, boundaries, and occasionally the removal of small details, but it never involved adding components to the image that were not in the original scene (or in some other original scene, even of his own making).

I don't think we know what he would have done, had he had the tools we have now, or even just Photoshop. What he did at the time was quite controversial, yet he was revered for creating images that people loved to look at. Isn't that the intent of photography as opposed to photojournalism?

I think there is still quite a stigma attached to inventing things in a picture. Those of us who go through incredible effort to take pictures in nature for example, don't accept artificially generated versions. That just becomes a painting, not a picture.

 

[pablolie]

I don't think we know what he would have done, had he had the tools we have now, or even just Photoshop. What he did at the time was quite controversial, yet he was revered for creating images that people loved to look at. Isn't that the intent of photography as opposed to photojournalism?

I think there is still quite a stigma attached to inventing things in a picture. Those of us who go through incredible effort to take pictures in nature for example, don't accept artificially generated versions. That just becomes a painting, not a picture.

Not directly related, but on a similar note... I used to be a Photoshop power user. My ex worked at Adobe so for years I got it for free. But I got very popular because I started using it for portrait shots, and then every friend and their friends wanted some. Always the same workflow... 20 plus clicks and corrections with no automation.

Then along came PortraitPro....

 

[Snarfie]

Here's a video illustrating an issue with AI I can currently relate to, as many users I guess:

Scary

 

[pkane]

Until they are replaced by GPT prompters.

Hehe..

 

[amirm]

Have not used it yet but seems Chat GPT-5 has made some good advances: https://www.tomsguide.com/ai/i-test...7-prompts-heres-which-one-gave-better-answers

In every test, it outperformed the previous version.

 

[sam_adams]

In every test, it outperformed the previous version.

It certainly did:

Red Teams Jailbreak GPT-5 With Ease, Warn It's ‘Nearly Unusable’ for Enterprise

Independent red teams have jailbroken GPT-5 within 24 hours of release, exposing severe vulnerabilities in context handling and guardrail enforcement, warning model is not enterprise-ready.

www.securityweek.com

 

[sonitus mirus]

It almost makes you wonder what is going on with our own existence, as it seems likely we are probably just some advanced AI experience.

 

[bluefuzz]

GPT-5 has made some good advances

Yes, as Ethan Mollick, who is professor of AI and innovation at the University of Pennsylvania, says GPT-5: It just does stuff. However, the stuff it does seems often to be a bit weird. That's not necessarily a bad thing but food for (human) thought. As the old curse goes: 'May you live in interesting times' ...

 

[Jeromeof]

I feel that the advances with GPT5 are really only slightly incremental, though it was obviously fine tuned on all the benchmarks. This is the sad thing about these benchmarks are they are public, so now part of the training sets.

The clever thing but also the sneaky thing with GPT-5 though is all the other models are being removed and GPT-5 includes an upfront 'router' to pick 'the most suitable' internal model for a given question.

That option probably confused most people before so a good thing but I think this also means OpenAI can make decisions about how 'smart' GPT5 is for most people. Basically internally pick a 'dummer' model when it suits them e.g. use GPT5-Nano when the costs get too much or their 'Compute' capacity is too limited. They can also cheat benchmarks e.g. have full capacity for the next 2-3 weeks while people maybe make decisions about switching from Claude to GPT-5 for coding, but then it can get worse once people get used to using GPT-5 as their main model.