AI

[hauritz]

An attack or a malfunction? In this video an AI powered robot appears to be throwing hands with people in the crowd. The implications of an AI basically taking matters into its own hands are things we should be concerned about.

Did it perceive a threat?

Did it stumble and correct itself before it fell?

You decide.

 

[hauritz]

Taingiwilta, is Australia's latest supercomputer or more accurately a computer system. It has been under development since 2022 and achieved Final Operational Capability in December 2024. It will work as a hub connecting multiple defence platforms.

It will be a powerful tool used to help manage and co-ordinate military forces and operations. Rather than a potentially cumbersome hierarchical command system which can take hours, days and even weeks to navigate, AI can do the same job almost instantly. It could analyse, priotitise any information it gathers and ensure that information is sent to the right people. It could even present the commander with a list of options.

It even has its own limited decision making capabilities. For example it might control ISR assets and make decisions as to where and how these assets will be deployed.

ADF supercomputer brings AI to decision-making

'Critical’ defence system hits major milestone.

ia.acs.org.au

 

[Sandhi Yudha]

An attack or a malfunction? In this video an AI powered robot appears to be throwing hands with people in the crowd. The implications of an AI basically taking matters into its own hands are things we should be concerned about.

Did it perceive a threat?

Did it stumble and correct itself before it fell?

You decide.

It looks quite threatening, but it still can be a sudden corrective action. In this video, we can see a second case of unexpected robot behaviour. Here we can clearly see, the robot, while still attached to his crane/stand, behave aggressively, which is clearly not to correct an imbalance situation.

In this video we have a closer look to that martial arts robot, quite frightening.

Chinese Kung Fu robot shows impressive martial arts skills

This is the impressive moment a Kung Fu robot executed a stunning 720-degree roundhouse kick to disarm its opponent.

youtube.com

Ten years ago, the Honda Asimo was already impressive, the world's most advanced humanoid. At that time people had the dream to create robots for simple tasks in places like hotels, restaurants and hospitals, or becoming an assistant or pet for elderly people living alone. But the extreme fast development of robotics and specially AI, will undoubtedly cause that humanity will loose control of its own creation, all in the name of progress and prestige.

In this last video we can see what we can expect in the near future.

 

[hauritz]

It is getting increasingly hard to separate fact from fiction. What seems like sci-fi perhaps a few years ago is now pretty much reality. Even now I can sit down with an AI chatbot and hold relatively normal conversations. Something I couldn't imagine 10 years ago. The growth in capability of these systems is almost exponential to the point of where we may risk losing control.

AI alignment is a process of ensuring that AI systems behave in ways that align with human values, goals, and intentions. Basically it is what stops AI from rising up and taking over the planet. This isn't regulated by the way. Everybody thinks it is a good idea but it would be extraordinarily difficult to actually enforce.

One thing that should concern us is the speed at which this is all happening and whether those in control of these programs actually know what they are doing.

The following video presents two scenarios. One in which AI aligns with with human values and the other where it doesn't.

This is the original scenario.

https://ai-2027.com/

 

[John Fedup]

This article describes how AI is now enhancing cyber attacks. I am sure AI vendors will respond with AI cyber defence. IMHO we may be approaching a point where the best cyber defence is a massive disconnect.
AI can now power every stage of a cyberattack - Defense One

 

[Vivendi]

This article describes how AI is now enhancing cyber attacks. I am sure AI vendors will respond with AI cyber defence. IMHO we may be approaching a point where the best cyber defence is a massive disconnect.
AI can now power every stage of a cyberattack - Defense One

From this article published in July: "The shift does not yet amount to fully autonomous hacking..."

The Hugging Face incident has demonstrated that this is no longer the case. The AI risk is growing rapidly. The latest update on Hugging Face, and also some information on other incidents that have previously not been published: OpenAI just disclosed more 'concerning' AI behavior following the Hugging Face incident. Here's everything you need to know


Anthropic has published this, showing that in August an AI was leading 26% of AI development, up from 0% just a few months ago: Measurements for understanding the pace of AI development inside frontier labs \ Anthropic

The level of autonomy is already quite disturbing, and also how fast things are moving. Where will this technology be in 12 or 24 months? We need better oversight.

 

[kato]

The level of autonomy is already quite disturbing, and also how fast things are moving.

Completely offtopic, but for the back-and-forth:

There's a website for advertising real estate over here that just introduced (in beta release) an AI function for its users to automatically send an "i'm interested in ..." message and data package to sellers/landlords if something new fitting their preferences is posted.
For the sellers/landlords it already previously had AI-based filtering and auto-reply to block off "i'm interested in ..." messages not fitting their preferences...

 

[StingrayOZ]

The AI hack of openai of Australian records is interesting..
Yet the americans won't let us use the current model to protect ourselves.

They are almost creating a situation where we either have to start locally hosting chinese models to do it, or/and develop our own model.

The chinese have opened sourced all the good technology, you just need to be ruthless about data harvesting like US/China is. No body wants to know how the sausage is made. With enough processing you can retrain existing open source models, if you have enough money, compute and resources and time.

 

[Vivendi]

The AI hack of openai of Australian records is interesting..
Yet the americans won't let us use the current model to protect ourselves.

They are almost creating a situation where we either have to start locally hosting chinese models to do it, or/and develop our own model.

The chinese have opened sourced all the good technology, you just need to be ruthless about data harvesting like US/China is. No body wants to know how the sausage is made. With enough processing you can retrain existing open source models, if you have enough money, compute and resources and time.

There are European open source models available, perhaps more attractive for Australia to build on those?

One area Mistral is particularly proud of is Le Chonk's performance in cybersecurity workloads. The company claims the model is one of the strongest performers in cybersecurity benchmarks like Artificial Analysis' new Cyber index.

European AI flag bearer Mistral's new open weights model is 'Le Chonk'

 

[StingrayOZ]

There are European open source models available, perhaps more attractive for Australia to build on those?
European AI flag bearer Mistral's new open weights model is 'Le Chonk'

I've used mistrals stuff before, its different, its quite good in the maths, and it has a particular flair to its writing style and with languages. There were hopes for Le Chonk, it doesn't seem to be bad, but its not quite frontier level. But its differences and the fact its not Chinese will make it attractive. IMO it would be a better deal for schools and universities to locally host (with a cloud provider, but in Australia) something based off Mistral.

I also think Mistral would make a good basis for a sovereign AI, retrained or layered with Australian language, attention, history, details etc. At least the public source versions, and maybe copyright holders could agree to give allowance to train on their material for a Australian AI, as used by the Australian government. Something that could be made available to every government department and hosted locally, controlled by the government. We should have a large pilot in a school system or something to leverage negotiations with all providers.

Its the best non-chinese open source model. But the Chinese are significantly ahead. The French are 6-12 months behind the Chinese. Which isn't huge, but is a thing. Chinese are also pushing the architecture more. However, people really don't need the top model for everything, a smaller, faster but still capable enough is good enough for 99% of work people do, 99% of the time. For businesses, and gov departments they can build workflows around it with out it randomly changing or breaking or doing weird things.

But we are under pressure to sign deals with US AI companies. Lock everything, the entire government/country into Anthropic or OpenAI.

We are at war with OpenAI currently due to hacking, and general shitbuggery. Also OpenAI has a huge funding problem currently. Anthropic is heavily courting the government. But entities like Universities are signing deals with OpenAI after some of the biggest educational intuitions signed deals with them.

Mistral has been recruiting locally, but being very French, I think they will struggle commercially in Australia. They need better connected people on the ground, able to chase government and private.