Key Emerging AI Trends

In partnership with

TWB’s SPONSOR OF THE WEEK

FUTURE

Key Emerging AI Trends

A year ago, OpenAI launched the o1 model, marking the beginning of the era of reasoning models and sparking a new wave of excitement and spending.

A year later, though, improvements have already been showing diminishing returns, with GPT-5 confirming that sentiment (although GPT-5 Pro’s “bad release” was deeply exaggerated, it was not the step function improvement we had hoped for).

We need a new thing. But what is it?

Today, we are looking at the five trends that are generating more interest, excitement, and funding. We’ll first focus on the next frontier in AI, memory, and we’ll then move to others, including things like ‘humility training’, ‘eval rethinking’, or ‘programmatic prompt engineering,’ that are shaping research and the next wave of progress and opportunities.

In a nutshell, this is for those who want to know what’s next.

Let’s dive in.

The Next Frontier: Memory

I’ll just go and say it: Memory is the next frontier in AI. That is, providing AI models with an ample, dynamic set of memory features that allow them to know you more and better.

Because in a world of marginal improvements in “intelligence”, context makes a huge difference.

AI Engineering is Context Engineering

With AI Labs in charge of improving model priors (what a model can or can’t do, aka skill building), our job as practitioners is to ground the model.

Yes, GPT-5 appears very smart out of the box, but it’s a general model; it needs context to work on your task. Thus, AI engineering is not only about choosing what model works best, but making sure the model actually works.

People tend to summarise this as ‘prompt engineering’, which can easily mislead others into assuming AI engineers simply craft the best possible prompt (choosing the best words, the structure, etc.), when in reality, the key to whether an AI engineer succeeds at the job is if they get context right.

Context engineering is, therefore, a much more appropriate way of understanding what implementing AI means (in the Premium subscriber section, we’ll cover one of the most interesting projects in a while that allows you to automate prompt engineering using AI).

And when memory is done well, the model’s responses just feel different. Superior.

Just like your mom’s advice always feels more accurate than your friend’s, because your mom knows your deepest secrets and can tailor the advice to your particular context, AI models become truly transformative experiences when they have the appropriate context to help you.

The issue is that this is easier said than done.

Do we actually know how to implement memory?

As I briefly mentioned on Thursday, there are two memory types: parametric and in-context.

Parametric memory refers to the type of memory the model has learned during training or fine-tuning. The model has been optimized to learn this information, making it an intrinsic part of its knowledge.

In-context memory refers to the information we provide…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.