Intelligence as a Service: AI Functions That Think Like Humans, Run Like Code
MindStudio's Serverless AI Functions merge human-level intuition with the speed and power of software
Every time I build something that lets people sign up with usernames, I find myself reaching for a giant text file full of offensive words that I’ve had saved on my hard drive for years. It’s got all the things you would expect, as well as all the variants that do things like replace the letter o with the number 0, or put underscores between each of the letters. When someone signs up, I make sure that the code checks, line by line, to see whether their chosen username matches any of the bad words on the list. If it does, it throws an error and makes them pick a different one.
But it’s inevitable that some clever person comes up with something new and it slips through the cracks. If you showed their username to a human, it would be obvious that it’s offensive, but because it doesn’t match anything on the giant list, the computer thinks it's fine and lets it through. After all, it’s not like the computer has a concept of what is offensive and what isn’t—it’s just chugging along trying to match words against a list.
If you’ve spent enough time building apps, you know there are a lot of problems like this. Things where the answers are incredibly obvious and straightforward to human brains, but, annoyingly, are almost impossible to define in code. Things like identifying if an email is spam, or taking a list of user-submitted product reviews and summarizing them, or creating a personalized message for someone based on things you know about them.
To non-programmers, it might be surprising that some of these things are so difficult—because to any human looking it barely requires any thought to solve the problem. This is why if you build anything that relies on user-generated content, you inevitably also need to build moderation tools and find humans to manually review what comes through.
Magic Functions
Over time, the number of computer-hard/human-easy problems has decreased. Services like Twilio made connecting to the cellular network and sending an SMS as simple as writing sendSMS(). Stripe took accepting payments from something that used to involve negotiating a contract with a bank to just writing charge(). Imgix and Mux took image and video CDNs from giant datacenter projects and complex transcoding pipelines to simply making a slight edit to a URL.
These advancements—taking big, complex processes and saying to developers, “don’t worry about how it works, just treat it like any other function in your code”—have fundamentally changed the way we think about what it is possible to build.
I believe we should be thinking about Generative AI in the same way. You can go ask ChatGPT if a username is offensive, or if an email subject line is spam, and it will tell you. You can ask it to create a tag cloud from a list of product reviews, or look at a list of recent purchases and see if they seem like they might be fraud, and it will.
MindStudio’s Serverless AI Functions let you build AI workflows and turn them into bite-sized functions that you can then bring into your projects and treat the same way you would treat any other code. So in addition to your usual signup validation, you can now also add a check for isOffensive(username). From your project’s perspective, it’s just code—in the same way isUsernameAvailable(username) checks your database to make sure the username isn’t already registered. But behind the scenes, the function is executing a workflow you’ve created in MindStudio that is trained and tested to your specific sensitivities around what is and is not offensive. It then queries your choice of one or more large language models, and casts the result to true or false.
Serverless AI Functions are like having the ability to script a human to perform a task at the speed of a computer.
Cost and Accuracy
At the end of the day, most of these kinds of tasks are pretty basic. You don’t need the all-consuming genius of a giant flagship model like Claude Opus or GPT 4o to figure out if “b00bz69” is offensive or not—a small, inexpensive model like Llama 3 8B or GPT 4o Mini is more than capable of giving you a fast answer.
However, some tasks might require a bigger, more capable model. You might want to take hundreds of product reviews and generate a summary that appears for customers at the top of the reviews box, and at that point—yes—you might need to consider the cost of such a task and decide how often to run it. Thankfully, one of the most valuable parts of building these workflows in MindStudio is that you can understand the cost as you’re building, and test different models to find the one that works best for your use case. When new models come out or vendor prices change, you can easily swap models without redeploying your product.
Additionally, powerful logging and analysis tools make it easy to keep an eye on what’s happening. You can sample reports about the decisions your Functions make, and even introduce alerts or human review processes if the AI is unsure about something.
You Don’t Need To Wait For AGI
I believe that a lot of the “gaps” people are solving for in AI today are because we are thinking about AI as a tool for building giant systems. Indeed, much of today’s focus on “agents” comes from a desire to build systems capable of complex reasoning and decision-making (e.g., building a “SalesBot AI” that knows how and when to do everything from qualify leads, negotiate contracts, generate collateral, etc.). It’s true that today’s AI systems struggle to perform these sorts of things reliably. Hell, it’s hard even for most humans to do these things well.
However, today’s AI models are more than good enough to complete the kinds of basic tasks that we would otherwise need to outsource to content moderators, customer service associates, or interns. Not because these people are skilled in some specific way, but just because we need someone to perform the kinds of tasks that are hard for computers but easy for human brains to solve.
We Don’t Know What This Unlocks
In the same way that services like Stripe and Twilio unlocked entirely new kinds of applications, this is another one of those things. The examples I’ve identified here (content moderation, fraud, ranking, etc.) are things that I have personally struggled to build, and are therefore the things that are immediately exciting for me to now solve using AI Functions. But for others, there will be completely different kinds of things (NPC behavior in video games, or ultra-personalized ad copy created on-the-fly, maybe?). Junior developers might even find value in using AI Functions to do things that are possible in code, but just not easy or intuitive.
The point is that over time, by building hybrid systems that mix the power and rigor of traditional code with the magic of AI Functions, we will expand the type of software that it is possible to build. And we don’t need to wait for some superintelligence—we can get started today.
More Info
Official Website
Press Release
Demo Video
more about MindStudio: mindstudio.ai
follow me: instagram, linkedin