Here's what building AI is really like
Getting into the weeds of building custom AI for our immigration startup
Happy Monday folks,
Spent the weekend “back home” with my family before trekking back to Toronto for another big week in the business world. Always good to recharge with some family time.
Let’s jump in:
Story: the nuts and bolts of building AI
It’s been a fun and interesting month on the AI side for us at Visto, my immigration tech company. For those who don’t know we launched Canada’s first AI immigration chatbot a few weeks ago and are continuing to make it better.
One lesson I’ve learned so far: it’s been harder than expected, which isn’t a bad thing.
Most of you have probably used ChatGPT and/or something similar, and maybe you’ve even used or built a custom GPT or some other AI app built on top of something like OpenAI. I won’t go too deep down the AI rabbit hole from a tech perspective (largely because I wouldn’t be able to), but it seems like a lot of early AI apps - and especially custom GPTs - are made by slapping some custom data into ChatGPT and calling it a day.
Don’t get me wrong, there are tons of extremely successful companies - financially and otherwise - that have just done that. And/or made niche custom GPTs and slapped a paywall on top.
On our end, we knew from the beginning we wanted to build a custom database and custom algorithm, which has proven harder than just training a custom GPT ontop of OpenAI.
One of the reasons for that difficulty is that, from what I’ve learned so far, there’s an art and a science to building custom AI. The science is a lot of standard, relatively basic stuff. Building a database of information, organizing it well, building the app so it has the chat functions, training, etc.
Then there’s the art side, which from my experience so far, I’d say is the prompt engineering side. For those with as little tech experience as I, you can think of prompt engineering as basically figuring out the right prompt - aka the question or command to give the AI - so that your chatbot is as accuarate as possible.
The prompt is what sits between your user and your custom database, because when your user types in a question to your chatbot, the prompt is the instruction you are giving to the AI so it can respond as effectively as possible.
For example, let’s say you’re building a copywriting AI tool that can whip up SEO-optimized blog posts. Your prompt might be something like “you are an expert blog post writer with an exceptional understanding of SEO. When the user requests a blog post about a specific topic, make sure to first find the best SEO keywords, then draft the blog post between 500-700 words with those keywords in mind, using optimized headings and formatting…”
As you can imagine, these prompts can get really long, detailed and complex, but there’s also a ton of strategy you want to follow in how you structure the prompt and give AI these instructions. Not to mention, small changes in your prompt can make huge differences to the output from the AI.
Can’t go into much more detail because, quite frankly, I don’t understand much of it and our engineers do the heavy lifting - although I do help with fine-tuning our prompt, and also testing the output we get from it.
So there’s some science and strategy, but also an art to getting that prompt just right so your AI output is as effective as possible. And it’s the tough stuff that makes it a fun challenge we’re working on improving every single day.
This was a little more technical than most of my posts, what do you think - do you enjoy hearing about some of the technical AI stuff, or should I stick to my more traditional business topics like marketing, sales, operations, etc.?
Let me know…
Key takeaway: working on innovative tech can be tough, but not in a bad way
Initially when some of our AI releases took longer than expected, I was concerned. As a tech cofounder and CEO, I always want us to be launching and improving new features asap.
What I didn’t appreciate was that 1) some of this AI stuff is still so new that there are very few experts who can, for example, whip up a “perfect” prompt right away (as mentioned above, that part seems to be an art that can’t really be perfected anyway).
And 2) when the tech is harder to build, you know it will be more valuable when you get it right. Just like anything in business, if it’s easy to build/launch then you will probably have tons of competition and it will be difficult to differentiate.
But if something is harder to do - technically or otherwise - it means you’ll probably have less competition and offer even more valuable if you can crack it. A huge moat/differentiator for your business.
So if you believe in what you’re building, lean into and appreciate the difficulty. As they say, it’s the pressure that creates those diamonds.
Content for the week:
Vlog: Because I was travelling this weekend I didn’t have the time to record and post a vlog, sorry to say. Back in action with one next week.
Book(s): I finished reading Hustle Harder, Hustle Smarter by Curtis Jackson aka 50 Cent. If you’re like me, you grew up listening to some of his early 2000s bangers (yes, this white Jewish guy listened to rap growing up in case you didn’t know!), so when I saw the book I was initially conflicted.
A business book from a rapper whose “prime” was ~20 years ago?
I was pleasantly surprised because 1) it was a good read, and 2) he’s actually been doing a lot of business over the last few decades (including selling Vitamin Water, producing some hit shows, still performing and other stuff).
It wasn’t like your typical business biography, aka white man works his way up the corporate ladder or starts his own successful company, which was a breath of fresh air. It was also written really casually and conversationally, which made it an easy read. He held my interest the entire time and shared some really good business lessons too - highly recommended.
Next up: Success is Never Final: His Life and the Decisions that Built a Hotel Empire about Bill Marriott.
Thanks as always for following along on my journey, let me know if you have any thoughts, feedback or questions, and have an awesome week,
- Josh Schachnow
Canadian immigration lawyer, CEO at Visto.ai
If you’re looking for more of my help, you can check out:
My immigration tech platform, Visto, that helps Canadian immigration firms prep and file applications much quicker.
Make sure you’re subscribed on Youtube where I post the weekly vlog and am also posting shorter snippets of my tips/advice as Youtube Shorts!