When AI Works ⭐

With Big Tech pushing us onto the AI rollercoaster whether we want it or not, it’s time for a reality check.

I have my opinions about this, of course. We all do. But this is not about being pro AI or anti-AI. It’s about being practical. It’s about using AI when and where it makes sense and ignoring it when it’s not. Each of the use cases noted below is transformative in its own way, a new and better way to do something. Collectively, I think it’s pretty impressive. And it’s only going to get more impressive over time. With the obvious caveat that there will be forward mistakes as well as forward progress, just like anything else.

First, a couple of core principles that I feel like everyone should agree on. I know, good luck with that.

AI exists to help people, not to replace people.

In keeping with the above, you need to do some work: Yes, AI still hallucinates, and is in many ways the ultimate example of “garbage in, garbage out.” Again, it will only get better.

When successful, AI will save you time, save you money, or save you both time and money.

When successful, AI will enable capabilities that are often not possible otherwise.

Intent is the key to how we communicate with AI. That is, instead of using commanding-based user interfaces that require us to click a specific button or other item on a screen, our interactions with AI will involve natural language, spoken or typed, in which we describe an outcome. When successful, AI will figure out what we intend to do and then do that thing.

Memory is the key to personalizing an AI to our specific needs. Over time, AI will learn about the types of things we want and then remember to apply that learning for future interactions. This is essentially a new form of customization in that it happens automatically and over time.

These principles, like the AI use cases noted below, are by nature incomplete. This is a big topic, and many of these things will require follow-ups and more detail. I will forget things. Unlike AI, I’m human.

And before getting to what doesn’t work, the one thing in which the delta between hype and reality is the greatest.

What’s not there yet: AI agents

We’re being sold on AI agents that will “do things on our behalf,” and I’ve only half-jokingly retorted that there isn’t a single instance on this planet of that actually working. But the reality of AI and agents is that this requires a lot of work, and most of that work hasn’t even started yet.

I’ve described that work as “programmatic” apps and services because I wasn’t aware of a good term for it. The idea is that we have apps (in Windows, on mobile) and services (traditional online services and now AI agents and services), and the former, especially, have always been designed for direct interaction by people. But for AI agents to work seamlessly, these apps (and services) need to be accessible in more sophisticated ways, with purpose-built interfaces that are not designed for people.

Well, we do have a term for this, as it turns out: Semantic.

I guess I was always vaguely aware of this term, but I was listening to the ChatGPT Atlas episode of the OpenAI podcast the other day, when I heard the term semantic web and it clicked. The semantic web is essentially a machine readable version of the web. So you might have a website for people to read, like Thurrott.com, and then there could be a semantic version of that website that AI services can more easily parse. (Rather than doing an unsophisticated screen-scrape, which is what happens now.) Given this, I will start using the name semantic apps to describe those apps that can seamlessly interact with AI agents and services.

So we can add that to the core principles: AI agents will begin making sense when our industry evolves the web and apps into semantic web and semantic apps.

OK. So what does work right now, and usually reliably?

Summaries

AI is terrific at summarizing content. Videos, web-based articles, and documents of all kinds are perhaps the most obvious examples. But AI is also very good at summarizing meetings, and that can be true whether you attended them or not.

Asking questions, getting answers

AI is quickly replacing traditional web search because people most often just want an answer and not a list of links. But thanks to natural language interactions, AI is particularly good at answering questions. This could be general or very specific, like a health-related question (Why is my scalp dry?) or a question that pulls information from multiple places (What is the best DMV to go to in eastern Pennsylvania tomorrow morning?). Tied to the meeting summaries noted above, you can also ask about what happened (Did my name come up?). And as with search, AI is replacing the types of things we would ask on sites like Reddit (Can leather conditioner fix a polyester seat?)

Writing help

Everyone needs help writing, whether it’s spelling and grammar, rewriting existing content, getting started writing new content, or whatever else. This happens everywhere, on PCs and on phones, and in many apps. It’s as useful in a text messaging app as it is in a word processor.

Visual search

All phones have this capability now and it is stunningly useful: Point your camera at a business or other place, or any thing, or use a photo or screenshot and find out what that thing is. The next level is finding out where you can buy it, if that’s what you’re trying to do.

Automation

This is one of the dark horses, in that we’ve been promised these kinds of solutions for years, but they’ve never landed with consumers. These are no code, ideally solutions, or low-code otherwise, things like Google Workspace Studio or Copilot Studio in which you use natural language to explain what you want to AI. Things like email filtering, alerts of any kind and so on.

Translation and captioning

This one is so magical it’s a no brainer. There is live (on-the-fly) captioning of whatever audio is happening on just about any device these days. Live language translation capabilities, same. And then the combination of the two, enabling a Babelfish-like experience.

Coding help

If you’re a software developer, chances are good that you’re already doing this, as AI pair programming has emerged as the one near-universal success story for AI. This is the GitHub Copilot, Cursor, Anthropic/ChatGPT/whatever else coding solution in which you can modernize an existing app or even vibe code your way to the start of a brand new app of any kind.

Research

If you look past the Q&A stuff noted above, which you might think of as one-off questions, the next step is more advanced research-based querying that maybe happens over time and will involve some basic agent usage. This is the Google NotebookLM world, but you can do it with just about any AI now. Some key use cases include shopping scenarios, like when you’re looking for something but don’t have a product name or brand in mind, any price watch-type solution like flights, consumer electronics, and whatever else, and planning for a one-time trip to some location. Basically, I don’t do this thing a lot, I’m not going to become an expert in it, but I need it to be right this one time. I guess I would put educational and learning in here as well, for whitepaper-type document generation.

Entertainment

Yes, everyone is freaking out about AI-generated music and movies right now, but I’m thinking more along the lines of day-to-day solutions like AI-based Spotify playlists, online multiplayer or single player maps and locales, and even something like the recent Alexa capability where you can use natural language to jump to a specific scene in a TV show or movie. (Go to the part where the Tyrannosaurus Rex attacks the Ford Explorer.)

And probably a lot more I’ve forgotten or don’t know about.

More soon.