AI in 2024 and Predictions for 2025. Part 2

December 9, 2024  |  Lee Simon Bristow

The Future of AI: Autonomous Agents, Digital Twins, and Personalized Innovation

Continuation of Highlights of AI Progress, Challenges, and Governance in 2024 with Lee Bristow

The EU AI Act and Prohibited Practices

Maybe we can circle back to the AI Act. As I mentioned, in 2025, the prohibited AI practices will come into effect. I’ll go into a bit of detail about what that entails. General-purpose AI obligations will also start coming into effect in the latter part of 2025, alongside the first level of penalties.

Looking ahead to 2026, the rest of the AI Act will begin to take effect. One key aspect to note about the EU AI Act is that it’s a risk-based piece of legislation. It categorizes AI applications into prohibited AI, high-risk AI, limited-risk AI, and minimal-risk AI. For the most part, the legislation aims to ensure that most AI applications fall into the minimal-risk category.

However, there are nuances. Even if an application is classified as minimal risk, it might still play a role in critical infrastructure within the EU. This can elevate the risk profile of the underlying asset. That’s a crucial point to understand.

In terms of prohibited AI, starting in 2025, social credit scoring systems will be outlawed. These are systems similar to those in China, like WeChat. Emotional recognition in workplaces and education settings will also be prohibited.

When it comes to AI classifications, there are four types: reactive AI, limited memory, theory of mind, and self-aware AI. Theory of mind refers to AI becoming aware of others. While we haven’t reached that point yet, AI can infer emotional states in humans.

Concerns About AI Cloning Personalities and Predictive Policing

Recently, a Stanford study demonstrated AI’s ability to clone human personalities. Researchers conducted tests on humans to map their personalities, then superimposed these personalities onto AI agents in a simulated environment. The AI agents behaved similarly to their human counterparts. Two weeks later, when tested again, the AI agents mirrored 85% of the human personalities. This ability to clone personalities is concerning and will be prohibited under the EU AI Act.

 It’s a forward-thinking move by the EU, as this technology could be exploited for behavioral manipulation. We’ve already seen the impact of such practices with the Facebook leaks and Cambridge Analytica, if you now kind of take those kinds of technologies, those kinds of behaviors in terms of big tech, we wouldn’t want AI in its current form of technology to start changing or influencing behaviors. It just wouldn’t work out really well. And then in terms of facial recognition, biometric categorization, and predictive policing, I think is a really big challenge as well.

This touches on the Minority Report scenario, where behaviors and actions could be predicted based on personalities. Law enforcement using real-time biometric identification in public spaces will also be prohibited.

And we’ve seen that coming out in some of those research and then law enforcement using real time biometric identification in public spaces. So those will be the prohibited AI or the application of prohibited AI. And the reason why I’m kind of peppering that list with a series of researchers that have recently come out is that I’m actually really glad that this is starting to come out, because if those researchers are coming out, there would be a lot of tool sets and technology that would start to exploit vulnerability, start doing behavioral monitoring, potentially even start doing predictive policing using those personality clones and then creating certain scenarios for them to play out.

AI Governance, Ethical OS, and Organizational Resilience

Looking ahead to 2025, areas like AI trust, transparency, and governance will take center stage. These shifts will be driven by EU legislation and emerging regulations in the US.

But what was really interesting is that it’s also aligned to the quality management systems like ISO and NIST are also utilizing or going along with a risk based approach. And I also think there is one other call out that I’d like to mention that is also kind of going along the lines of that would be the ethical OS.

Everyone realizes that you can’t just run out and implement AI. You actually need to figure out what it is that you want to do with the AI tool sets. How are you going to use them? Where the data is going to come from? How the algorithms are going to be built? What are the ethical considerations? What are the governance considerations? What are the people considerations within your organization, both in terms of currently employed and potentially future employed? How are you going to educate? How are you going to retool the people within your organization to be able to do that? These elements fall under what I call the “governance umbrella,” ensuring organizational resilience in adopting AI technologies.

The Rise of Autonomous Agents

On the technology front, autonomous agents are expected to become more powerful. Tested in pilot programs in 2024, these agent-based workflows will gain traction through 2025 and beyond. Studies by KPMG or PwC show that over 90% of executives view AI as a core strategic pillar. However, many fail to realize the urgency of preparing for it. Autonomous agents in sales, marketing, supply chain, and logistics can provide significant competitive advantages. Without readiness, organizations risk falling behind.

A recent MIT study highlights this. Conducted in November 2024 by researcher Aidan Toner Rogers, it examined how AI-assisted research in a large US R&D lab improved material discovery by 44%, increased patent filings by 39%, and boosted innovation output by 17%. This demonstrates how AI can accelerate novel idea generation rather than just iterative improvements. For top scientists leveraging AI, the results were orders of magnitude better, though lower-tier scientists saw less benefit.

And so all it’s ever going to do is regurgitate the same stuff. AI was able to create 57% of the ideation process or idea generation. And what was really interesting in the research paper was that the AI actually looked at the data differently, which meant that it was actually able to create novel ideas as opposed to iterative improvements.

However, one downside noted in the study was a reduction in job satisfaction for 82% of scientists involved. They felt disconnected from the core creativity process, which diminished the joy of overcoming challenges. Humans thrive on tasks that feel achievable yet challenging, providing micro-hero moments that drive satisfaction.

Digital Twins and Personalized Customer Interactions

In hardware developments, we may see innovations moving away from transformer models, which are opaque “black boxes,” to more interpretable models. Additionally, new chip technologies could reduce the resource intensity of AI operations.

Digital twins, widely used in manufacturing, are extending to other areas, such as healthcare. A recent case in the UK involved a woman from Portugal who was almost blind without glasses. Doctors created a digital twin of her eye and used AI to simulate surgeries. This approach enabled her to achieve better-than-20/20 vision post-operation.

This was essentially using the digital twin of her eye. If we extrapolate that further, if we can already create personalities or essentially a digital twin of a human based on their personality test, and we can do an eye, that means we can apply this concept to a customer. We could potentially extrapolate that all the way up to understanding who that individual is. Now you can understand why the EU legislation exists around policing.

So, bringing it all back to how to create a customer persona: In traditional marketing and sales, we would create a persona. Inside that persona, we would know about Lee, where he lives (he’s South African, living in Dublin, 48 years old).

We would build up this view of who he is, then consider factors like his disposable income, where he would spend it, and so on. You can start to really build up an idea of who that individual is. In a corporate sales environment, you could target the way in which you write emails to that individual in such a way that they would understand it.

You could use the right tone, adjust the number of people in meeting rooms, and personalize the communication. In a non-corporate retail environment, it could get pretty scary.

For instance, you could change the lighting as Lee walks down the aisle in the supermarket to encourage him to buy the next cereal. This concept of hyper-personalization takes things to a whole new level. While you might not want to use it in all environments, it can be applied effectively in a contact center. The digital twin customer is a compelling, emerging concept—it’s still foundational at the moment.

I think spatial computing and humanoid working robots are going to become much better understood through 2025 and probably won’t be implemented until 2026 or 2027. If we can bring everything we’ve learned about large language models and apply it beyond just typing a prompt, this will mark a shift.

It’s almost like the Internet in 1993, when you were basically interacting through a prompt. You could still get the information, but it wasn’t as open or accessible.

The next interface, however, could involve robots. These robots will allow us to start interacting with humanoid types in new ways. We’re still at an infant stage. Some of these robots have already started to break out of the research and development labs. Look at Tesla with their robots and Boston Dynamics with theirs—those are the top examples.

We’re moving beyond just written data formats. The AI will need to process and understand data in real time. This year, with GPT-4 and others, we’ve seen significant progress in inference.

Test time compute, where large language models improve their responses almost in real time, is advancing. We’re moving from test time compute to test time training, where the large language model uses a retrieval-augmented generation (RAG) architecture to improve and contextualize its output.

This is a frontier-level development, where the AI takes in more information, stores it in memory, and then updates its response accordingly. This process updates the weights for that response in real time but doesn’t permanently alter the model’s underlying weights.

What this does is speed up the process, reducing the need for as much compute power. This method of going into memory and updating the model’s weights for the response in real time will become mainstream, impacting agents in future models. These agents will be the primary way we engage with AI outside of a prompt environment.

    Access full story Leave your corporate email to get a file.
    Yes

      Subscribe A bank transforms the way they work and reach
      Yes