Let’s take a moment to think about your kitchen. Big or small, rich or poor, your kitchen probably has a decent number of items in it. Some may have different appliances, like blenders, Instant Pots, toasters, microwaves, etc. Folks with more room may have things like air fryers, convection ranges, induction plates, and so on.
One of the things that cooking shows have drilled into us over the years is that unitaskers – single-purpose appliances – are bad in the kitchen. They consume a lot of room, a lot of counter space, and they’re impractical. A tool that has only one useful function just clutters things up, like barbecue meat shredding claws or a pancake batter dispenser. Regular forks and mixing bowls serve those functions just as well, and have plenty of other uses besides that.
There are, of course, some exceptions. My espresso machine is a unitasker. It really doesn’t do anything else except make espresso. However, that’s a case where the unitasker is so good at what it does that an attempt to substitute it with something else is going to create really subpar results. Could I just make regular coffee with coffee grounds and a cooking pot? Of course. Is it espresso? No. Will it be tasty? Maybe. But it’s also a lot of work.
Now, you may be asking, this is useful advice to a degree, but what does it have to do with anything I normally talk about? I bring all this kitchen talk up because this is how people are approaching artificial intelligence right now, and it’s causing issues.
Not a day goes by on LinkedIn or on Slack or Discord when someone asks, “Does anyone know a good AI tool for X?”, where X is a relatively common function. A tool to make a Powerpoint presentation, or a tool to make a specific kind of chart, or a tool to digest meeting notes or transcripts – these are the kinds of requests I see very frequently.
And the reality is that most of the time, you don’t need a specialized, single-purpose vendor for these requests, if you know how to use the foundation technologies well. For example, someone who wants an AI tool to take some data and turn it into a specific kind of radar chart? That’s something ChatGPT’s Advanced Data Analysis option can handle with ease. Something to process meeting notes? There are so many models that can handle that specific request. You don’t need to buy extra software to accomplish those tasks (though certainly vendors of that kind of software would argue that you do.)
This behavior isn’t limited to artificial intelligence. Every year, my friend Scott Brinker and his organization publish the MarTech Landscape, which shows the number of marketing technology solutions available in the market. This past year, it was around 11,000 different companies and solutions, which is absolutely nutty. The MarTech space is filled with point solutions – unitaskers – that do one thing, and companies’ accounting ledgers and credit card receipts are filled with dozens, perhaps even hundreds of single-purpose vendors.
Why does this happen? It happens for three reasons – time, skills, and culture. When something’s on fire and you need a solution right now, people typically do not invest a lot of time into evaluating the problem and determining the different options. They need something fixed, and they find and buy the fastest solution to their problem.
The second reason is skills and experience. If you’re technologically savvy, you have the necessary skills to build a solution, perhaps out of existing tech you already own, and thus you don’t need to buy anything. On the other hand, if you don’t have the necessary skills, you may default to buying something rather than acquire the skills needed to solve the problem (which is directly related to time).
The third reason is culture. Many folks who read this newsletter live in a culture of both capitalism and instant gratification. We have an app for nearly everything. We expect solutions, and we buy solutions as often as we can. Along the way, we’ve been dissuaded by corporations in both B2B and B2C contexts from solving our own problems; our default behavior is to buy rather than build – even when building may be the best possible choice.
So, what are the solutions here? How can you approach AI (especially generative AI) in a thoughtful manner without being overrun by hundreds of point solutions? The first and most important question you have to ask is, is the problem you’re trying to solve suited for AI?
Large language models – those engines that power services like ChatGPT – are really good at predicting words. As language models, they are proficient at working with language. Thus, if you have a language-based problem, they’re probably going to be pretty good at solving it. If you have a problem that is not language-based – like math – they’re going to be less good at solving those problems. The same is true of older forms of AI, like classical machine learning. If you have a problem that involves classifying data, you cannot and should not use regression models to do so. It’d be like trying to cook a steak with a blender. It’s just the wrong tool for the job.
Second, once you’ve identified that the problem you’re trying to solve can be solved by AI, the next most important question to ask yourself is, can I solve this myself with a foundation model? Foundation models are the starting point for AI solutions. These are tools like ChatGPT, Claude 2, Bing, Bard, and LM Studio, paired with models like GPT-4, Claude, PaLM 2, and the LLaMa 2 family of models. Foundation models are flexible; you can get them to perform many different tasks.
Contrast this with point solutions built around a very specific purpose, like transcription software, social media post generation software, NDA evaluation software, blog writing software, etc. where you may not even know the underlying model. These point solutions are inflexible and worse, may be bound to a foundation model that is significantly out of date. At the recent MAICON Conference in Cleveland, I talked to several engineers at vendors who sheepishly admitted they architected their solutions a couple of years ago to be hardcoded to the GPT-3 model from OpenAI, and now their solution significantly underperforms the newest foundation models.
Get good at working with the foundation models so that you know what AI is and is not capable of – and only then should you consider purchasing a point solution. When you do go down that road, ask tough questions about how the solution is architected. Ask the sales engineer to describe the abstraction layer inside the vendor’s software, and what kind of flexibility it has. A software solution with a good abstraction layer allows you to pull one model out and put another one in, like changing the heads on a stand mixer to switch among the paddle, the whisk, or the dough hook. Most software vendors do not build their solutions with this in mind, so the answer to this question can be quite telling in terms of how well the vendor will keep up with the rapidly changing AI landscape.
Just as there’s an app for everything, there’s an AI model for nearly everything, including things where AI models are simply the wrong solution. To avoid falling into the trap of dozens of AI point solutions cluttering up your productivity kitchen, follow the steps I outlined above and become proficient with the foundation models and tools. You’ll be more capable, understand better what the tools can and cannot do, and perhaps even invent a product or service of your own along the way.
By Christopher S. Penn
