OpenAI this week released these new “GPT apps” that has everyone in a tizzy, BUT over the past six months I can identify THREE founders whose business models are badly affected by their introduction.
So in this video, I’m going to do something that I don’t normally do – In this video then the thing I don’t normally do is a the kind of deep dive and analysis that I would do as a Fractional CTO.
One of the founders I’ve worked with over the past six months doesn’t have a business at all now, because these “GPT apps” are actually what their business was trying to produce. They’ll never compete with OpenAI (and everyone else who tried to copy them), so they need to pivot.
The other two of them need to use this technology to deliver their value, but what’s been announced today isn’t enough for them to start doing that, so strategically they have timing, funding, spending, and competition problems.
Effectively both those businesses today need to stop what they are doing because the existence of the new stuff is enough to indicate that something structural is happening to their market and they need to find out what that is. But I’m going to go through that at the end.
First though, a quick sidebar. OpenAI’s branding is AWFUL. ChatGPT is a terrible name. They have gotten away with it solely because the people who know that what they are doing is GENUINELY SPECIAL that they pump oxygen into keeping the fire going. These things they’ve announced called “GPTs” – that’s also a terrible name. (See before when I called them “GPT things”.) So in this video I’m going to call them “GPT apps”. I want to highlight just how bad OpenAI is in “talking to their customers where they are” in terms of branding. They need to hire a branding specialist.
But anyway, with that said, what are these GPT apps?
ChatGPT, and its sibling LLMs (large language models) in things like Bard and Bing Chat, do two things EERILY well. And I say “eerily” because what you get butts up against this maxim of “any sufficiently advanced technology looks indistinguishable from magic”.
The two things it does eerily well is a) understanding the user’s intent from plain English, and b) summarising data. It doesn’t do anything else particularly well, and I think for the most part if you steer away from specific use cases with ChatGPT, the outputs are honestly pretty lame and disappointing.
If you think about how frustrating an Alexa is to use, and think about how smooth ChatGPT is to use, that’s how eerily good the “determining intent” part is. When Alexa gets something wrong, you think it’s the one that’s dumb for not understanding an obvious question. When ChatGPT gets something wrong, you think you’re the one who’s dumb for asking the wrong question.
Like how ChatGPT is able to answer queries like, “tell you what, give me that information again, but this time broken out into a table I can paste into Excel”, and it just does it – that is eerie/spooky.
The summarising information part is the bulk of what ChatGPT does. It is ASTONISHINGLY good at taking unstructured data and summarising its contents. Like I can paste this script in and say, “in two sentences, explain what this video is about”, and it’ll get it SPOT ON.
Despite how magical it appears, all ChatGPT does it look at the set of data that it can see and work out what the “typical” response to the question it’s been asked would be. A way to think about “typical” is that it’s actually giving you the “average” response.
However, there’s a effect in play here which is that the larger a set of similar things becomes, the more AVERAGE that set as a whole becomes. It’s this that leads to ChatGPT giving lame, unimaginative, and rote responses.
Imagine Microsoft in the late 1970s when it had ten employees. The “average level of being a computer genius” in that organisation was very, very high. Microsoft now employs 221,000 people – the average level of “being a computer genius” is now far closer to the general population. (A good part of managing that business is working to keep that bar “above average”.) The bigger something gets, the more average it becomes.
ChatGPT basically is a massive pile of “words”, but the system’s purpose is to find the AVERAGE response from that massive pile. This is why when you ask it a question like, “write me a cold outreach email to a new prospect”, you get back something that is so BORING. It’s not “boring” per se, it’s average because what it’s doing is answering what MOST PEOPLE WOULD LIKELY SAY, but as humans we see average as boring. We need verve and elan to move the needle for us.
To get ChatGPT to do anything that’s not “average” in relation to the general population, we have to give it its own data to work on.
This is the first thing that the “GPT apps” are able to do. They are able to work with their own sets of data. This is sort of genius because, fundamentally, LLMs are useful only when they can work on specific data sets. It’s only when we do this can we use LLMs to further our mission of creating value for our customers, because at a fundamental level, all value that flows from digital products given to our customers is derived from the unique data that we have, whatever that is. That can be a set of profiles in a dating app, a collection of instructions for maintaining equipment in an office building, it can be a set of recipes in a direct-to-consumer food prep service, the locations and availability of vehicles in a ride sharing solution.
For an example of how this creates better results, a friend of mine whenever she is in a dispute about something with her partner, she’ll copy and paste the WhatsApp chat she’s having with him and ask, “given this transcript, what’s going on between those two people?”
ChatGPT does AMAZINGLY well with this query because it’s very qualitative, very subjective, and very targeted on a small set of data. It elicits insights that are genuinely helpful in resolving the dispute that she and her partner have. What’s happening is that she’s giving ChatGPT her unique data to work with, and using that to derive value, albeit not commercial value.
This links up to one of the first points I made which was that one business I’d been talking to this year no longer has a business, because this was exactly what they were doing – i.e. making it possible to give ChatGPT an organisation’s data to analyse. These GPT apps just do that, out of the box. There’s no way that was an original idea, so loads of founders are now having to pivot.
The second thing that GPT apps do is make it easier to do something called “prompt engineering”.
When you see on LinkedIn all those people lying about “using ChatGPT to make $10,000 a month”, or telling porkies like “how I fired 20 staff and replaced them with ChatGPT”, the example they give relate to a legitimate aspect of using LLMs like ChatGPT which is this “prompt engineering” idea.
When you ask ChatGPT a question, the query you ask is actually tacked onto the end of a whole bunch of other instructions that tell the model how to actually answer your query. You can see this in these sort of cute hacks you see people do like, “answer in the style of a pirate,
Prompt engineering is about introducing instructions before the user’s query. This is effectively how you “write a computer programme” for the model. With GPT apps, when you’re setting one up you can say, “always answer the user’s query pretending that you’re a pirate”, and now if the user uses that app, every query they get back will follow your instructions that the answers should be like that of a pirate.
This is obviously a trite example, so allow me to give you a better example.
I have created a GPT app, which is live and you can use it (but you need a Pro subscription to ChatGPT to access it), which contains the transcripts of all of the videos that I have produced. This is the earlier part I spoke about which is where you can now give ChatGPT data to use, rather than using its standard set of data.
It’s called Fractional Mattbot and you can access it at bot.fractionalmatt.com – the link is in the comments.
The advantage here is that I have a view of the world that is mine. For example, if you were to DM me and ask me whether you should outsource your MVP development to Poland or to India, I will give you a very clear answer based on my experience and philosophy. If you ask ChatGPT, it will give you an average answer from everyone’s average experience, the end result for you the user being a set of hints as to how you explore this problem further.
However, if you ask a GPT app that can only use materials that I’ve written, you will get an answer that represents what I would probably say when asked.
So that is what the “Fractional Mattbot” does. However, part of getting it to do that is the prompt engineering part.
The way you actually engineer the prompt in a GPT app is by talking to ChatGPT. So I have said things like, “answer as if you were me, using my writing and speaking style, and sense of humour, that is in the materials uploaded”. And, “always prefer to use the materials uploaded, don’t use data from your general corpus unless you really need to, but tell the user when you do this”.
i.e. I have to engineer is to make it more like talking to me. The value that comes from that is that I converge on product-market fit. I have my data, I have my approach, by engineering the prompts I can make the ChatGPT deliver value to my customers (or in this case, customers/audience).
Another thing that you can do is that if you say to it “how can I contact Matt?” I have engineered the prompts so that Fractional Mattbot will reply with “have a look at his website at XYZ, and you can connect with him on LinkedIn at ABC”.
And you can keep doing this, with statements like, “If the user asks about pricing information, tell them that information is on the website, and provide them with a link”.
Now when the user is using the app it can map what the user is asking against those intents, and respond accordingly.
So I said that I was going to do a deep-dive to show you how I work as a Fractional CTO, and some of it is orientation material like the above. I’ve tried to show you what ChatGPT is good at, what it’s not good at, where the limits are – mostly that ChatGPT becomes much more useful if you can give it your own data – and give a couple of examples about how the new developments in the GPT apps look to UNBIND those limits.
The next part is to look at strategically what do you do about this.
The big challenge that founders have today is that there’s not enough here to use, but the direction of travel is obvious. I can make this very specific.
I can make Fractional Mattbot know everything about my service, pricing, how I do what I do, mechanisms for booking in meetings, or commissioning me, but I can’t use it. You can access the bot, ask it questions about specific issue, get the answers you want, and want to move forward commercially, but you’re stuck within the ChatGPT interface.
Like it would be very sensible of me just to have a little chatbot on my website that under the hood is Fractional Mattbot, but you can’t use that yet. There are also issues in being able to being able to hand over to other services. If you want to book a meeting, it should be able to take you through that process or take you to Calendly to do it interactively. If you want to buy something, it should facilitate that process.
This is a little example, but what we have today in GPT apps is 80% of the solution, created by a business that it would honestly be fairly stunning if they didn’t deliver the remaining 20% within six months. “fog of war” over what the next six months looks like.
The way to look at it is that right now these “GPT apps” are highly siloed, and this creates a huge structural problem in that you can’t embed them into digital products, and therefore can’t use them to deliver value to the end customer. But, it’s inconceivable that this problem won’t get fixed. It’s “jam tomorrow”. OpenAI are also very closed off with regards to sharing their roadmap. I actually can’t think of a situation in my career when there’s been something like this – an obvious outcome, but a kind of
That’s not helpful, strategically, to a startup or scaleup business right now because what do you do? Whatever you were building last week has been obsoleted. There is a new place that we all need to get to, but we don’t know for sure that Shangri-La is on the horizon, or what our ETA is even if it does.
Without knowing the specifics of a situation, all I can do is kind of say, “you need to stop and wait for more information to emerge”. But I don’t really like saying that.
For now, it’s time to really understand what these new tools do and start gathering intelligence. Also being able to prioritise forward motion in areas unaffected by these movements would also be sensible.