I like to commence an article with an update on current affairs. To ensure it is always relevant and contextual, it usually centers around some key event for the world, and has broader implications for the construction industry at large. Here is the latest event:

Earlier last week, OpenClaw - recently acquired by OpenAI - was announced at NVIDIA’s GTC conference as the “future computer” making reference to it as an operating system. Additionally, a robotic version of Olaf trained on a world model featured alongside CEO Jensen Huang to usher in a new age of industrial robotics.

Now, if you are like any of the professionals I work with, a small percentage of readers know exactly what these events and the bold words mean (perhaps someone in your BIM Team) and an even smaller percentage know what to do with this information (perhaps someone in your IT department).

ABOVE: Demonstration of a neural network with hidden layers classifying a dog Source: https://duynguyenngoc.com/posts/cats-vs-dogs-classification-using-cnn-keras/

The breakthrough in Artificial Intelligence which we all currently experience is because of a subset of machine learning called deep learning.

Deep learning, although around for decades, erupted in 2012 and 2016 respectively when it enabled computers to ‘perceive’ images and beat the international world champion in the game of Go.

Deep learning is responsible for the success of computers ‘understanding’ our language through large language models like ChatGPT. They are also responsible for the success of the reverse of recognising images - generating them - through applications like Google’s Nanobanana.

However, both platforms are only interfaces to engage with the technology which works beneath their surfaces. The magic of deep learning lies in its mini-programming scripts which all work together in a way inspired by the human brain to ‘learn’ - or change themselves based on new inputs. These mini-scripts are called neural networks.

Perhaps an oversimplification of what is really going on, but it will do for now, neural networks are at their heart automated statistical procedures. Neural networks identify patterns in large sets of data which would take immense time and effort to decipher. By doing this, they can help us make more sense of the data we supply them.

This is what we do at the institute - we use neural networks and their applications to identify patterns in data to help professionals make better decisions. Using a stepped approach to assess how to tackle the problem with AI, professionals build the solutions to problems the experience in their businesses through the programme.

Like Tylo Barends (now working at Dhk Architects) who created Zonelytics - an application which converts zoning information from the City of Cape Town into volumetric parameters in a Revit model.

And if that did not sound useful enough - Tylo did this all without having any knowledge of software programming within 3 days.

The outcomes clearly showed that AI was not a one dimensional technology: it operated at different levels of problem solving. And through this, the greatest outcome was that students demonstrated that there are various levels of AI integration.

ABOVE: Tylo’s Zonelytics overview of project workings: Read more

ABOVE: AI Levels of integration, AI Strategy for an Architectural Business (version 2). Published by An Architect + Institute of Building Design.

The greatest benefit of humans being intelligent is that we can foresee many possible futures. We generally call this form of insight the ability to think.

Artificial intelligence as part of its definition is aimed at mimicking human thinking. The value of this may be in making machines appear to think - but what is of greater value is in providing us with thoughts we might never have had.

By simulating various possible outcomes, humans assess the probability of failure or success for an objective when we think. This specific task - simulating outcomes based on previous data - is what AI is exceptional at doing.

ABOVE LEFT: An architectural structure located in Cape Town, blending elements of local South African architecture, Parametricism and Computational Design. Generated by Akheel Khan, Midjourney. 2023. ABOVE RIGHT: An architectural structure located in Cape Town, blending elements of local South African architecture, Parametricism and Computational Design. Poster generated for SAIBD CPD Workshops. Generated by Tahiyya Naicker, Midjourney. 2024.

In 2024, I wrote a piece titled “Think Again” in the newly launched ArchSA edited by Dr Sechaba Maape on what I believed AI is actually for: helping one improve the quality of thoughts.

So counter intuitively, the most effective use of AI is not to use it for automation - but to add value to human thought for better judgement.

Which leads to the first step in the levels of integrating AI into a process: identifying a problem. The problem must be meaningful, painful and frequent enough to qualify.

Examples I see everyday which professionals bring to me: finding a new candidate to work with, using software for rendering and assistance with professional practice examinations and compliance. All of these problems have various levels of AI interventions which can be of benefit.

The power is in deciding when not to use it.

Iteration. Variation. Generation.

All these words are familiar to anyone in the built environment - especially on the design side.

By definition, these words all suggest multiples which have a similar essence. Like bricks - all brick buildings are made of roughly similar materials, but what kind are they? Facebrick, clay, hollow core? All variations of the same idea - but essentially different depending on region, structural integrity and aesthetic purpose.

Now the interesting part is that these same words - iteration, variation, generation - also exist in the field of computer science and machine learning.

So the next part of this process is then to use a large language model - which is exceptionally good at creating variations of output through statistical prediction - which simply means that it can help you frame out your problem more clearly and identify already existing solutions.

Like Bronwen Renaud and Jaco Landman from BHC School of Interior Design, who used Google’s Gemini to reason out their problem as academics and figuring out how to lessen the cognitive load on students whilst maintaining curriculum alignment.

They then moved on to the next step - using an existing multi-faceted platform - Google’s AI studio to build an evaluator of briefs which assess against their criteria. Their total hours spent on building it: less than 72 hours. Note that the version you are seeing here is only the final iteration made during the short time they had to complete the project - there were several iterations which did not make the cut.

What Bronwen and Jaco achieved was time gained by reducing effort spent on figuring out where there was an alignment mismatch. They now joke that their next focus is to resolve marking - a workload that they are currently procrastinating on.

ABOVE: A screenshot of the application developed by Bronwen and Jaco which maps architectural curriculum to its alignment with Bloom’s Taxonomy, an educational learning methodology

Olaf is a fictional character in Disney’s computer generated feature film Frozen. He was brought to life at GTC (a tech conference) through a robotic body with software powered by NVIDIA GPUs (a global GPU manufacturer) which were used to trained neural networks on large datasets - including videos.

Datasets which are primarily trained on video imagery have been particularly useful at doing something incredible: generating ‘worlds’. Akin to video game digital environments, they call these world models. As one could imagine, world models are particularly useful in the construction industry - for everything from producing digital twins to controlled simulations of light on objects (better known as rendering). There are many models like this - models for various things like image, text and sound generation.

Alex Van Rensburg from Team Architects has been using a similar method - training neural networks on set patterns from his companies archive to produce novel products. Alex downloaded various models and has since created complex geometry sets after engaging with neural networks on a popular database for machine learning called HuggingFace (what a name!).

Alex’s work demonstrates that you do not need to be an expert to do effective work; you need a well defined problem and care enough to work through the details to solve it.

ABOVE: A screenshot of the application developed by Alex which takes inputs of patterns [left] and produces outputs of variation [right]

The closest thing to automation that I have come across is the ability for a machine to write its own software and control its own operating system via simple commands from its user. Andrej Karpathy, a founding member of ChatGPT’s creator OpenAI, says it is the closest thing to science fiction in reality he has seen.

The computer science community calls this “agents”. It is the dream of many businesses - virtual aids working together independently and of their own accord doing tasks that a human would do.

This dream was brought to reality by Peter Steinberger a few weeks ago with his operating system called OpenClaw (its latest name - the product actually had a few). OpenClaw is available for anyone to publicly download and use.

In this domain, when professionals say one should “use AI”, they actually mean which large language model shall I use to perform a task? This method of thinking frames AI as an action taker.

For some fields like computer science, this is helpful. If your job is to write software, then this actually does execute your work for you. But if your work is to make good decisions - especially ones which have a physical effect like we see in the built environment - then this is not the most valuable use of your efforts.

What becomes the most valuable usage is having options to make a good decision on.

As of 2026, I have not seen a successful deployment of an AI agent meet the lofty expectations of businesses. And whilst attempts have been made, the core failure I have observed comes down to interpretation.

ABOVE: Visualisation by Moebio of how LLMs work as probability prediction engines. Interactive source: https://moebio.com/mind/utm_campaign=Thought%20Leadership&utm_content=286963216&utm_medium=social&utm_source=linkedin&hss_channel=lis-CKsguRgtEw

Large language models which follow the transformer neural network architecture (the “T” in ChatGPT”) all have an internal function called “temperature” which makes the conversations sound more human.

It effectively acts as a randomiser of co-efficients that helps make the next word coming through in a sequence of words more interesting.

And that means that no two outputs - even if given the same prompt - should ever be the same. In order to produce exactly the same output temperature must be constant, GPUs must be the same and models must also not have changed.

All of the above are almost impossible to control - especially when companies like OpenAI do not share internal data about their systems architectures.

And this is where agents have a structural flaw - if no two outputs are ever meant to be the same, we cannot guarantee a determined, predictable outcome. Outputs from the large language model, whether a document or a piece of software programming scripts - may have the same intent but will never be replicated.

And to solidify the concept further, analysing human sentiment has the same issue. A machine is statistically predicting the likelihood of what you said with what its view of the world suggests you meant. That leaves us with one final issue - one we all experience - my AI doesn’t understand what I’m saying!

So given the context I provided in this article, and hopefully the newfound knowledge on deep learning which is which what powers current AI applications, world models which are trained on video datasets, temperature which gives large language models a ‘human’ feel and agents which are claimed to be the future - one final question you will need to ask yourself make before engaging with this bewildering technology to make the most effective use of it is:

To AI or not to AI?

ABOVE: Me (center, Akheel Khan) presenting how we approach AI at the Institute of Building Design during the Big 5 Conference in 2025. Photograph by Videsh Boodu

Akheel Khan is the co-founder of the Institute of Building Design (IBD) and founder of impact organisation, An Architect which works with businesses, professional bodies and organisations to make architecture accessible. His previous notable roles include Systems Researcher in AI at SAOTA, where he led research and development work at the AI unit; Lead Examiner at SACAP’s Professional Practice Examination, where he led the integration of financial literacy

and emerging technology concepts into the national papers, and finally as a Project Leader at Y: A SAIBD Initiative where he is responsible for building the VA of the Future. You can find him on LinkedIn where he posts frequently on AI + Architecture, or if you’d like to meet him in person he will be at the Inc conference at Century City, Cape Town, later this week.