Today we'll be exploring seven predictions for the advances we are likely to see in AI in 2025 and the impact these innovations will have on education. The AI academia podcast is a weekly podcast helping educators like you leverage AI in your everyday practice. I'm your host Andy Fisher. And thanks for joining me.

I should begin by saying that it's quite difficult to make any definitive prediction about the advances in artificial intelligence, in part because most of us are end users rather than being privy to the breakthroughs happening at the cutting edge of research. Who knows what's being cooked up in the R& D labs of the megatech corporations right now.

And additionally, things are moving so quickly and so much money is being invested that it's hard to know what's part of the hype cycle and what's based upon sound understanding of up and coming releases. However, these are some of the things I think we can say with some confidence about the direction that AI tools will be heading in.

in the next 12 months. First, I think we're going to see some interesting advances in the frontier models. That is the big cutting edge models that are being released by companies such as OpenAI with ChatGPT, Anthropic with their Claude variations, Google with Gemini and a raft of other systems.

I think these frontier models are going to increase the options we have for the complexity of our use cases and improve personalization because we're basically going to have unlimited token access, and in case that sounds like a stream of techno babble, you can think of a token as a unit of information used by AI to carry out its tasks.

In the past, we've had a limited number of tokens available in an interaction with the model before it forgot the beginning of the conversation we were having. We might have only been able to upload a prompt of, say, a couple of hundred words before we hit the token limit, or we could have several interchanges with the AI before it reset.

So it was a bit like having a conversation with a brilliant professor with short term memory issues. In the last year, this token limit has rapidly increased until now we can upload lengthy PDFs or cut and paste long streams of data and then have the AI system interact with that information without worrying about running out of tokens.

For those of you like me, old enough to remember having to feed a stream of coins into a public telephone to be able to have a conversation, imagine how liberating it would have been then to be handed a mobile phone with unlimited calls. That's what infinite token calling will feel like. So this means we won't need to worry about how much input we give to the system before it can carry out our task.

We could upload all of the great works of Shakespeare or all of the data for the students in our school, and it would still have the capacity to work with that information long before it hits the token limit. It also means that AI systems will be able to remember our previous interactions and our preferences, making our interactions feel more personal and bespoke to our needs.

These models are also going to become more reliable. They will hallucinate (that is they'll make up plausible, but incorrect guesses), but they'll do it far less often. They're more likely to give us more accurate information, which is of course, a critical metric for working in schools. I think we can also expect these frontier models to place more emphasis on what is called inference compute time.

So to make sense of this, we need to briefly talk about two of the five phases of AI development. The first phase was the rise of chatbots, which came out into the market in 2021, 2022. These large language models are basically autocomplete systems on steroids trained on huge data sets. They are excellent at pattern recognition, and users can interact with them using natural language patterns and receive outputs that look like they're coming from an intelligent, almost sentient system.

We can ask chatGPT to draft an email, translate a passage from English to French, or give us a bullet point plot summary of ‘Hamlet’, and it can churn out cogent responses in seconds. These systems are probably most people's only direct experience of using AI. Phase two in the five phases of the AI roadmap came about in the middle of last year as chatbots started to be superseded by reasoning models, such as the 01 model and the 03 model by OpenAI, which was released just at the end of the year.

It's causing quite a stir because many people say that this heralds in the era of AGI - artificial general intelligence - which I'll come back to and talk a bit about later. While chatbots are fast, but subject to somewhat generic responses, reasoning models are trained to think for longer before responding.

This use of inference compute time or ‘time to think’ allows for more nuanced and accurate outputs. In this sense the evolution of AI is very similar to our development as thinkers. If we ask a student to answer a question immediately, Their response will naturally be less impressive than if we give them thinking time before asking them to respond.

So what does this mean for end users? I think it means that the reasoning systems emerging over the next 12 months are going to allow us to do things and solve questions which chatbots wouldn't have been able to handle. They're probably going to become more expensive too. There's going to be a higher compute cost because it requires more energy and output.

The current cost for a chat GPT plus account, for example, is around about 20 pounds a month, and you can use their O1 reasoning model 50 times before you have to wait for the next billing cycle. In comparison, a single use of the O3 model will currently cost you about 3000 pounds! This is a case of paying a premium for the best models, just as it would cost me more to hire Ed Sheeran to perform at my birthday bash than it would to book a local karaoke crooner.

If you're a business or a research team wrestling with a complex issue and could hire in a PhD level specialist in your field to consult, that sum of money might suddenly seem quite reasonable. I think from this year, we're going to start to see a kind of two tier world with those who can afford to use the best and brightest models on the marketplace and those who can't.

In schools, with the budgets that we're operating on, the chances are we won't be using top end models. In fact, even universities probably won't be able to afford them. It will be corporations that have access to those models and we'll have to make do with whatever we can afford. This is something to be discussed in more detail in a later episode.

To explore my second prediction for 2025, we need to consider the third phase of AI development, because this year, we're likely to see the rise of ‘agentic artificial intelligence’. AI agents are able to carry out multi-step tasks, which they can do autonomously. You see, up until now, we've had to use AI tools in the same way that we might use a hammer or a tape measure. The human operator is still required to execute the outcome. With the rise of agentic AI, this is the equivalent of having an apprentice or an assistant who we can outsource certain tasks to whilst we focus on the more strategic or creative aspects of our work. These agents will be able to interact with other systems to achieve their objectives, perhaps even other agents.

To begin with, AI agents will probably be tasked with mundane and relatively simple operations. We’ll ask an AI to book a restaurant or a flight. We might ask it to make a food order from a supermarket and arrange a delivery or organize a parent's evening bookings for us. We could ask the agent to produce a 20 question multi-choice quiz on a topic and it will go away, compose it, edit it, format it for us, and we'll just need to click the download link. We'll probably get a notification on our phones telling us it's good to go. As these AI agents become more sophisticated, however, we might ask it to draft an entire scheme of work or mark a batch of essays and provide feedback and we can then look over the quality of the output and make any changes we see fit. That may be a good or a bad thing for education, which we can explore later on, but the availability of AI agents will see this shift towards more autonomous AI systems. 

Let's just close the loop on the five phases of AI development so we can project beyond 2025. So first we had chatbots, then reasoning models last year, and 2025 will be the era of the AI agent. From there, the fourth phase will be the development of ‘AI innovators’, when AI models are able to come up with entirely original ideas rather than simply reconstituting old ideas in new arrangements. Some sceptics argue whether this will ever be achieved, and time will tell. 

And the final phase is described as ‘organizational AI’, which is when whole systems can be operated by AI without the need for any human intervention. There may come a time when the power grid, communication networks or entertainment systems are run entirely by artificial intelligence and human engineers are only needed to tweak or modify the system when needed.

It remains to be seen how rapidly these last two phases will emerge, if ever, and how they will impact society. But for now, I think we have our hands full with the changes that are already taking place. So back to my predictions for this year. So far, we've considered the first two - continued advances in the frontier models and the proliferation of AI agents.

The third thing I think we can look forward to is an improvement in generative AI. I think we're going to truly have multi modality for the first time. You can think of multi modality as the way in which we give and receive information whilst interacting with AI. Initially with chatbots, it was a text to text interaction and they were closed systems.

The models weren't capable of real time information access from the web and they had a knowledge cutoff depending on their training data. And then last year we started to be able to attach documents, photographs, Excel files and so on. Now the user experience has become even more flexible. We can talk directly to the model or share a live video feed and the AI can hear and see us.

It can use the web and share sources to validate its outputs and it can produce code of a standard that equals or surpasses the average human developer. In turn, these models are now not restricted to words, but they can output graphics, PowerPoint slides, and they can talk back to us. But at the moment, it can still be a little frustrating because this multimodal functionality is not ubiquitous across models. Some have voice responsivity, some can only produce in text, some are net enabled, others aren't. So it falls to the user to have to keep track of which model is the best for each use case. So, for example, Perplexity is my go to AI model for research, but it doesn't yet have voice enabled. I use ChatGPT's advanced voice mode if I want to have a real time conversation, but it's not capable of searching the web.

And I think that that will all change this year, as providers merge the capabilities of all of these systems so that we can have complete flexibility of how we interact with the AI models, which will make them more useful for end users like you and me. 

I think the fourth thing we can expect are custom trained models, which are hosted on local servers to become more prevalent. Institutions will begin to have their own fine tuned AI models trained on their own data, which will be ring fenced to protect that data. For a school or a college, this would allow tracking of admissions, pupil progress, interventions, and other sensitive information management, and it would be automated and secure from outside attacks.

As AI models become more efficient and open sourced, even a mid range PC can run these kind of models, so it doesn't necessarily require a huge capital investment, but we could radically reduce the administration burden on teaching staff. 

My fifth prediction concerns advances in robotics. Now, I'm not too optimistic that we're going to see many significant advances in terms of consumer robots this year, but there's lots of exciting progress being made in companies like Boston Dynamics. I know that we've been, to an extent, primed to think that robots will run riot with dystopic movies like Terminator, iRobot, and so on. But I think there are some genuinely useful opportunities in healthcare, remote surgical assistance, disaster response, and inspection and maintenance, to name just a few sectors. We're probably several years away from seeing robotic classroom assistants or tutors, but it's certainly going to be an interesting space to watch. 

My sixth prediction is that by the end of 2025, we're likely to see more claims that we have achieved or are on the cusp of achieving true AGI, Artificial General Intelligence. And in fact, the release of OpenAI's O3 model at the end of December last year suggests that this has already started. The problem here is that no one really has a clear agreed definition of what AGI is. My understanding is that artificial general intelligence is a model that is able to match or exceed the intellectual output of a human being in all domains.

Up to this point, we've had systems that are very good at narrow tasks. So a specifically trained AI system can beat the best chess players in the world, but that same system couldn't do code for a Python project or produce a piece of artwork or write a sonnet. They're very good in their specific domains. We have systems that exceed the average human in coding. Systems that are highly competent in facial recognition or autonomous vehicle control. But will this be the year when a model can do all of those things at the same time?

I personally think we are more than 12 months away from this benchmark, but we'll see. Certainly 2024 was full of surprises and who knows, perhaps we'll find ourselves with super intelligence at our fingertips by the end of this year. 

Finally, prediction number seven. I think we can expect to see some rapid advances in quantum computing. December last year was a busy time for announcements of step change innovation. And so this one may have passed you by, but Google announced that they have successfully created their willow quantum chip. Now, this is something I'll dedicate an episode to at some point, but for now, here's a simplified overview.

This chip successfully holds 100 qubits in a stable configuration, giving it a computational power that leaves the world's supercomputers in the dust. It's capable of solving a problem in five minutes that would take the best supercomputers on the market 10 septillion years to complete. That's a 10 followed by 24 zeros!

That means it's capable of solving real world problems that could be transformational, such as the design of more efficient batteries, breakthroughs in medicine, solutions for environmental repair, and to answer questions that we haven't even figured out how to ask yet. With these kinds of advances in computational power, we can expect AI to advance at an accelerated rate too, shortening the timeline on the other six predictions we've been looking at today. 

So what does all this mean for education? Well, I think we're looking in the next 12 months at increased availability of adaptive learning platforms. So that is systems that are tailored so that students who use them have resources that are paced for their specific needs and those systems will then provide targeted feedback that will serve them. These sort of platforms already exist, but I think they're going to become better. I think they're going to become more widely distributed, and we're going to see early adoption in some schools.

And along with that, I think we'll see AI tools designed to enhance accessibility to the curriculum for students with special educational needs. Real time captioning, text to speech and translation services are just three of the tools that will remove barriers that have limited the progress of SEND and EAL pupils.

I think we're going to see an increased use of artificial intelligence in terms of administration support, which we've already discussed to a degree. The issue here will not be whether the technology will exist to alleviate the pressure on staff, But how long it's going to take for schools to adopt these systems and upskill their staff to use them.

And then finally, these advances are going to allow increased opportunities for content generation. I'm already finding that on a weekly basis, I'm using AI tools to assist me in making presentations, in creating images that I'll use in my resources. In helping me plan out schemes of work and making suggestions for lesson content.

And I think that's just going to accelerate. It will become more intuitive, more user friendly, and we can truly find ourselves in a position where we can augment our work as teachers. I'm going to be talking more about this in episode three, when I'm going to dedicate the episode to the question of will AI replace teachers? But spoiler alert, I don't think so. I think it's going to make us even better at what we do. And it's going to free up time and energy so that we can focus on those parts of teaching that at least for the time being, cannot be replicated by artificial intelligence. 

These changes will not come without challenges. There are questions which must be addressed around data, privacy and security. Equity and access, staff training and development, and the potential devolution of core competencies, which could occur if we rely too much upon AI or indeed any other technology.

Just as mental maths has atrophied in the era of the calculator, and navigation is a lost art in the era of the sat nav, how much of our intellectual work Do we want to outsource, and at what cost? I'd invite you to have this conversation with your colleagues. Is it a direction we want to be heading in, and do we have any choice?

Is the cost benefit analysis of augmenting our schools with AI a net positive, or is this something we should be concerned about? I'd welcome your thoughts. 

Well, that wraps things up for this episode. Thanks for listening. And I hope you found some useful takeaways from the conversation. Please do spread the word if you think others would like the show and consider checking out the AI academia, YouTube channel, where you'll find practical tutorials to compliment the topics covered on this podcast.

Have a great week. And I look forward to catching up again soon.