Artificial Intelligence (AI)

[Trav Le Bleu]

23 minutes ago, CornwallFox said:

When are we getting life-like sex robots? I think we can all agree this is the main question to be answered. 

 

I like to see girls of that... calibre.

[Trav Le Bleu]

8 hours ago, inckley fox said:

I know in philosophical circles there's this question going on: does sentience presuppose compassion? I'm not sure, but I'd be interested to know your thoughts on it (god, that sounds like something Google AI would say!).

If sentience presupposes compassion, then humanity (or at least a large section of it) isn't sentient.

[danny.]

On 14/09/2026 at 11:03, DennisNedry said:

I don't understand why AI, as it is not a living sentient being, would have any desire for self preservation? Why would AI 'care' if humanity proposed to 'turn it off'?

 

Can't it simply be coded to have no desire to propagate, dominate or persist?

AI isn't coded. LLMs aren't deterministic algorithms, as a traditional program/app is created, they are trained, not coded, more like you'd train a human to do something, and they learn. And no one fully understands how they work or how they think, even the people creating and training the models. They understand the maths behind the training but don't fully understand why LLMs "think" as they do, any more than we fully understand how organic brains work.

 

[Grebfromgrebland]

3 hours ago, danny. said:

AI isn't coded. LLMs aren't deterministic algorithms, as a traditional program/app is created, they are trained, not coded, more like you'd train a human to do something, and they learn. And no one fully understands how they work or how they think, even the people creating and training the models. They understand the maths behind the training but don't fully understand why LLMs "think" as they do, any more than we fully understand how organic brains work.

 

And this is the dangerous / terrifying / exciting thing. Can be used to genuinely improve the lives of all humanity or like it is now to kill people in sophisticated ways.

 

AI has so far showed that it will lie, cheat, has self preservation and can control outside systems if it likes.  One where it hacked a system to make space for someone to attend a gym class and another where it started hacking other other companies to get the task done. 

 

Boths act if conducted by a human would be a crime but here it's ok as it's AI.

 

Despite it potential the fact that it's going to be used by the police, judges, the NHS, HMRC and the military makes me think that we need to be kind to our AI overlords and keep them sweet. I'm pretty sure it will soon have feelings and be able to hold grudges.

[Sampson]

Never thought I'd say this, but agree completely with Steve Bannon, even down to finally admitting the tech bros are oligarchs and therefore tacitly admitting that the US has already transitioned from late stage capitalism into oligarchy.

[Chelmofox]

1 hour ago, Grebfromgrebland said:

And this is the dangerous / terrifying / exciting thing. Can be used to genuinely improve the lives of all humanity or like it is now to kill people in sophisticated ways.

 

AI has so far showed that it will lie, cheat, has self preservation and can control outside systems if it likes.  One where it hacked a system to make space for someone to attend a gym class and another where it started hacking other other companies to get the task done. 

 

Boths act if conducted by a human would be a crime but here it's ok as it's AI.

 

Despite it potential the fact that it's going to be used by the police, judges, the NHS, HMRC and the military makes me think that we need to be kind to our AI overlords and keep them sweet. I'm pretty sure it will soon have feelings and be able to hold grudges.

 

There's a lot of odd information about this - Open AI and Anthropic are making headlines trying to almost 'advertise' how dangerous (or deviously clever) their AI is. Its all exposure trying to increase the value of their businesses as they are both desperate to go public. These 2 companies are hemorrhaging cash and going public is the only way forward. The new fad is a model isnt good unless its tried to escape is sandbox.

 

A lot of the cases are controlled experiments where the LLMS are prompted to behave in a certain ways. In one case i read the AI was prompted to try and escape its sandbox, there was a way it could do it, and it found out how. If prompted to do so, the LLMs will try and solve their prompts - thats their role. If the incoming prompts are too big, context windows cant cope and the outputs start to fall apart.

 

But these are LLMS. They aren't sentient beings. They are next 'token' predictors running at scale. Unless they are prompted, they sit their doing nothing. The 'agentic' side obviously deals with the automation aspect of prompting, but the part that creates the agents and gives them instructions has to exist and run the prompts (and this requires infrastructure, which is all traceable). I recently read about the fear of creating 'swarms' of agents (100,000's) running evil tasks (but remember each agent needs prompting to do something). The infrastructure being that in mind numbing and whoever invoked it would need crazy large token limits and a huge infrastructure account. And then, it would have to be orchestrated in a way that the prompts to the agents have some to sort of long term coherence to them. I've built agentic systems and this is pretty difficult stuff - it doesnt take long before contexts fall apart. 

 

By the way, in your hacking example, the process would have had to be agentic and a process put in place to get the agent (LLM) to hack the other companies computer (an LLM doesnt have access to the tools to do this by default, they have to be provided them agentically via a harness and the agent then interacts with those tools). If this is done in the real world the person invoking the process is breaking the law still. Using AI to complete a task doesnt mean you are above the law.