AI is progressing fast and slow

The AI revolution continues to evolve and has been a key driver of equity markets over the last three years.

Significant sums of money are being spent on semiconductors, data centres, energy, models and applications, so there is a lot hinging on the profitable deployment of AI throughout the economy. It is therefore important to understand what is happening with the underlying technology.

ChatGPT is the most famous application and is powered by their GPT-5 AI model. Models have three key stages: pre-training, post-training and inference. Pre-training is like schooling, with a huge amount of computing power required to analyse massive amounts of data. The post-training stage involves humans supervising fine-tuning of the model for specific tasks and to operate responsibility. Inference is the process of the AI “thinking” through user requests to generate a response.

Scaling laws

A lot of the early focus was on the pre-training phase. There has been a debate around scaling laws. The basic question being whether more computing power and more data means better AI models. OpenAI (owner of ChatGPT) CEO Sam Altman has been the most vocal proponent of scaling laws. This is why OpenAI has committed to enormous levels of spending to secure access to computing power.

Anthropic’s CEO Dario Amodei also believes in scaling laws, noting that “every time we think we see a barrier, the river just kind of flows around it”. Anthropic, however, has been more disciplined on spending and emphasised the importance of using computing power as efficiently as possible.

Google recently launched Gemini 3. This model was widely praised and seen as important confirmation that scaling laws remain intact. If models stop improving even with more computing power and more data, then this is a risk for the equity market.

Facebook owner META has had a mixed experience with AI. They have been an early beneficiary of actually deploying AI to improve content targeting. However, their own AI model has lagged peers. META’s former chief AI scientist, Yann LeCun has recently left and set up his own venture. LeCun was outspoken on the topic of scaling, noting that simply adding more computing power does not guarantee better AI models.

The race for AGI

The large model providers are racing to achieve Artificial General Intelligence (AGI). There is considerable debate around this and it has no actual definition, but is generally defined as human-like cognitive abilities. Altman is most optimistic, noting “we are now confident we know how to build AGI as we have traditionally understood it”. Conversely, AI sceptics like Professor Angus Fletcher argue that the human brain is not a computer but a “story engine”. It’s a fascinating debate, but veers well into philosophy.

Beyond scaling

One challenge with scaling has been the lack of new data. Leading models have already read the entire internet. There has been a move to use synthetic data, but this comes with risk as unreliable data can mean unreliable models.

We have therefore seen a shift in emphasis to post-training and inference. In post-training, there is a focus on Reinforcement Learning. This involves models learning through trial and error and given a goal rather than specific instructions of what to do.

The training stage is very costly but happens once, whereas inference happens billions of times as the model “thinks” through answers. The view on how this stage works has evolved significantly.

Thinking fast and slow

In the influential 2011 book ‘Thinking, Fast and Slow’, Daniel Kahneman differentiated between System 1 thought and System 2 thought. The former being fast and instinctive thought based on intuition and the latter being slower, deeper thought based on logic. This is a helpful lens for AI.

Previously, AI models provided quick answers akin to System 1 thinking. The industry is now using additional computing power to “think through” questions before answering, more like System 2. This “thinking” requires more memory. Instead of making the model (brain) bigger by scaling, spending more time thinking through answers helps models improve.

Semiconductors and energy are both constraints on AI, but memory hardware is a new bottleneck. Companies selling memory hardware have seen significant stock price gains. The ability of AI to “remember” improves functionality. For example, it can tailor answers to specific individuals or companies based on their learning habits or preferences. With worries about a market bubble in AI, these advances in the underlying technology remain very important.

Efficiency is also important and technology has consistently seen huge efficiency gains over decades. Efficiency gains are a risk to the infrastructure powering AI, but falling costs of computing power can be offset by more demand. Equity valuations are important, but monitoring progress in the underlying technology is also time well spent.

Nicholas Rilley, CFA, Investment Manager and Strategy Analyst

Disclaimer: The views expressed are the opinions of the writer and whilst believed reliable may differ from the views of Butterfield Bank (Cayman) Limited. The Bank accepts no liability for errors or actions taken on the basis of this information.