Anticipating the curve - tracking AI’s exponential growth and opportunities

With the rate at which AI develops, it’s easy to forget where it came from. Most obviously, generative AI has gone from comically bad to professionally viable, with video experiencing the largest quality jump in the last year. However, Large Language Models (LLMs) remain the most impactful across industries – and it’s why each release is so eagerly anticipated.

In this article, I explore the evolution of LLMs and what the next big innovation is set to be.

Our AI journey to date

In 2017, the breakthrough with Transformers in AI changed everything for LLMs. This introduced a model using self-attention mechanisms, enabling more efficient, scalable, and versatile processing of data – all the ingredients that make today’s AI models so good. 

In 2019, OpenAI's GPT-2 introduced preschool-level intelligence, with the model able to perform basic counting and write simple stories, including one about a four-horned unicorn. The release of GPT-3 elevated the model’s abilities to that of an elementary student, with the model able to generate simple text that could be applied in real-world use cases, such as for meta and product descriptions. 

GPT-4 made a significant leap, comparable to a high schooler in writing, math, and coding. From my point of view as a copywriter, its ability to write coherent – and sometimes decent – copy was a clear indication of the trajectory we’re moving towards. Since then, we’ve seen a wave of LLMs introduced and there’s no sign that development is slowing. 

The next chapter: AI with PhD-level intelligence or better?

Just a few months ago Kevin Scott, Microsoft CTO, outlined in an interview that some of the newer models he’s seen could pass qualifying exams as a PhD student. It turns out, Scott was likely talking about what he’d seen with Open AI’s latest o1 series models. o1 is a groundbreaking innovation, introducing advanced reasoning capabilities that systematically break down tasks to generate more accurate outputs. 

It’s a step closer to mimicking what a human brain does and achieving Artificial General Intelligence (AGI) – AI that can match or surpass human ability across diverse cognitive tasks.  

AI’s development curve is already there for us to see in the history books. From GPT-2 to today’s flagship GPT-4o model, there was a rapid increase in effective compute, algorithmic improvements, and practical enhancements like vision capabilities. And it’s not only OpenAI that’s innovating, other companies like Anthropic, Google, and Meta are also advancing quickly.  

While compute scaling is clearly a factor in AI’s rapid progress, it’s not the only one. AI models are getting much smaller, more efficient and far cheaper. The recently released GPT-4o mini is claimed to be more than 60% cheaper than GPT-3.5 Turbo through the API – while also offering text and vision capabilities. The cost and quality improvements indicate algorithms are getting better, meaning AI can still advance without considering long-term compute-scaling challenges.

We’re seeing this firsthand with Microsoft’s announcement of agentic capabilities coming to Copilot Studio and Dynamics 365. Copilot Studio allows organisations to reimagine their business processes by creating agents for Copilot that draw on the context of their data. Meanwhile, Dynamics 365 is introducing 10 autonomous agents to support sales, service, finance, and supply chain teams.  

OpenAI has recently released its experimental ‘Swarm’ framework – a more dynamic, multi-agent system. Instead of being confined to rigid workflows, Swarm agents can dynamically switch roles, share information and adjust their behaviour. This should – in theory – allow agents to handle more complex tasks by dividing responsibilities, improving efficiency, and allowing greater autonomy.  

But it’s when we get true, autonomous multi-modal AI agents out of the box – powered by the best LLMS – that we’ll see the greatest leap in productivity gains. We’re seeing an experimental version of this with Anthropic’s new Claude update, which allows its AI to autonomously control a user’s PC – just like a human.  

Eventually, these innovations will mature, and every marketer’s PC will have an AI agent like this on standby. However, even at this advanced stage we won’t be able to trust AI 100%, and it’s why human value will remain essential for authenticating and enhancing AI-generated work.

Don’t underestimate human value 

As Voltaire said, “With great power comes great responsibility”, and we should look through every AI development with this lens. With the trajectory we’re on, it’s inevitable that AI’s capabilities and potential use cases in a marketing context will soon exceed our wildest dreams of only a few years ago. But how we choose to use and integrate it into a modern workplace matters. 

For starters, empowering everyone in an organisation to confidently and creatively navigate AI tools is essential. This helps to produce the best results from AI, ensuring it follows instructions more diligently to produce more consistent results in both basic and advanced use cases. 

Thanks to the proactive AI approach being taken by Revere’s leadership team, everyone across the agency has completed a prompt engineering training session. This means we’ve all established a strong, best-practice foundation that’s allowing us to embrace and work with AI. So no matter who you are or what your role is, you’re able to complement your own unique skills and abilities, allowing us as a collective to deliver innovative, enhanced and faster results for our clients.