No items found.
A new breed of chat bot is about to make Large Language Models a whole lot more useful in practice.
Just before Sam Altman’s shortlived departure from OpenAI, he delivered a keynote at the annual public event where he presented a stunning new feature called “Custom GPTs” to enthusiastic applause, it was reminiscent of Steve Jobs announcing the App Store.
If you pick up a general model like Chat GPT and ask it to carry out a task, it will apply generalised knowledge and provide an output that is similar across users. A Custom GPT is fine-tuned for a very specific task and works in a bespoke and domain specific setting to be very good at a precise thing, It might be accessible to the public, exist within a single business or created for individual use. Ultimately OpenAI’s vision is for Custom GPTs to proliferate into all areas of digital life. We can expect to find Custom GPTs everywhere, including within our architecture practices.
I regularly host AI Workshops with practices and each session brings new ideas for new use-cases that could work with Custom GPTs; for example to compare and contrast a spreadsheet of tender returns, or hosting company policies and onboarding documentation or my personal favourite; training a GPT to create novel RfP draft responses from stock answers and project case studies. None of these ideas replace the work of architects, but they massively improve the experience of “leg work”; providing instant insight and getting you to a 60% ready draft that you can develop, rather than having to start from a blank white page.
The “T” in GPT stands for Transformer and this feature is most important because what Custom GPTs do really beautifully is to synthesize information from one form, into another.
Rather than looking something up manually, you are able to converse directly with dense collections of documents with your own natural language — you speak to the body of douments and they speak back.
Video demo of the BSA Bot:
Upon first hearing about Custom GPTs my mind jumped to several use cases in practice that I’m currently finding challenging. I have worked on projects that sit in the ‘grey zone’ of “High-Risk Building” designation. There are sometimes more nuanced cases where simply looking up a building regulation becomes very hard to navigate, particularly when I need to understand the latest on fire safety and how this affects CDM 2015 at the same time.
I have attended two CPDs on the BSA and I still feel technically adrift. Where I struggle is the interplay between the old regime and the new regime and the dates that make the new rules apply to some projects and not others.
So as a solution, I created The Building Saftey Act Bot which is a Custom GPT specifially trained on publicly accessible government guidance on the new regs: all 229 pages. In addition, it has been fed Part B Fire Safety Regulations (wording only) and supplementary HSE guidance about how CDM regs are impacted by the BSA. The bot will provide a BSA-informed contextual response to any project query about how the regs apply to a project. It’s not a final and definite answer, you will still need to check with the source material and take responsibility for your decisions, but it’s a very helpful navigational aid if you are feeling adrift.
The Building Saftey Act Bot is an experiment and will stay in “Beta” for now, it’s important to note it has not been stress-tested across many real-world projects, but I hope you will try it out and let me know how it goes for you, it will need continuous improvement!
Here is an example response from the bot:
If you’re feeling annoyed by the number of times I’ve written “GPT” so far, I don’t blame you. It all sounds like a technical headache, but in this new age of rapid change you simply have to set aside feelings of befuddlement and lean-in, I promise it’s much more simple than it sounds.
You may be wondering what advanced Computer Science skills do you need to create a Custom GPT? The answer is nearly zero — you instruct and configure the bot with plain English, or French or whatever your natural language of choice is. You explain the specification and requirements, feed the app your documents and GPT builder goes to work in the background, writing code to produce a bespoke model that behaves in the way you have instructed.
You no longer need Javascript or Python to code, plain English is the new coding language.
On a recent episode of the Proptech 3.0 Podcast, host Hattie Walker-Arnott flagged some of the acute risks associated with Custom GPTs. Specifically, they bring a newfound ability to spoof people and to imitate strangers with great ease. A custom model can be trained to write like you, to know your backstory and with enough audio material; even to sound very like you on the phone.
Organisations also need to be very careful about the information they feed Custom GPTs. Unless you are working with a Private LLM that has been ‘airlocked’, then the material you upload is likely being harvested by OpenAI. It is possible to enable privacy on your account by request and any business using ChatGPT should have this enabled. When you feed GPT files you should consider; “is the file copywritten?”, “does it contain GDPR sensitive information?”. Conversely, if you have already banned GPT, can you guarantee that your employees aren’t using it anyway? There is mounting evidence that there is widespread covert use of these tools across the industry already by people on personal devices.
More broadly, some quick and simple regulation is needed to help corral some of these emergent risks. A person’s face, their likeness, their voice and personality should be considered inalienable and sovereign to the individual. These characteristics should not be reused or sythesized by AI into novel settings without express permission. I don’t know many people who would object to this idea and it would help protect against deep fakes and fraud and force tech companies to make sure they are designing appropriate safeguards in their models.
Final note on Copyright and IP infringement, Open AI has made an unprecedented promise to provide legal protection to any user that suffers a lawsuit for copyright infringement following the use of their tools and this step is will provide significant cloud cover for creators using the platform and this clearly demonstrates how bullish the makers of this technology are feeling about its future.
Originally published in the Architects Journal
The term ‘ideation’ always sounds made up to me, but essentially this is what these image generation platforms aim to do.
