Those are a lot of very small examples, but when you start to scale it up, it starts to be able to do things that humans really just can't do. Right now, it is an infant technology that can do some things very well and one of the things that computers can do better and faster than humans is dig for patterns. Like Pete said, our brains are good at finding things like faces where faces don't exist, but computers can scale that up tremendously. Before AI as machine learning was starting to become a buzzword, we had Big Data and data lakes. The idea was to start collecting and storing any data that you could generate ASAP and then keep it forever. The ability to sift through that was limited at the time, but we knew that there was a lot of data out there that was very transitory and even if we didn't know what to do with it yet, if we didn't capture it, it was gone forever. We could do some of that though. Think about it in terms of trying to find an answer to a question that you don't know to ask. A couple examples from the Big Data era:
- They captured the telemetry from the forklifts in an organization that apparently had an absolute E36 M3load of forklifts. By sifting through the data, they discovered that they could identify when a forklift operator was about to have a heart attack because of how the forklift was being driven. Not something that they set out to find, they just collected the data and basically told the computer to go find something interesting for them.
- One of the cell phone companies pumped all their billing and account data through one of the Big Data tools and had it try to discover something interesting. They found that if one of their customers switched to another carrier, the numbers that customer dialed most frequently would tend to start to leave for that same carrier. So they started sending promotions to the friends of anyone who switched to try to retain them. I heard it worked pretty well. Again, this wasn't something that they told it to find.
- Even if the data is "anonymized", with enough accelerometer logging data from enough vehicles, an individual stream of data can be de-anonyomized with a high degree of accuracy. Not GPS data. Accelerometer.
That stuff isn't terribly impressive now, but this is what they started doing about 15 years ago. With the advances in the code and the massive jump in the ability and availability of highly performant GPUs, this kind of stuff can be knocked out without breaking a sweat. It is a GIGO problem to be sure, but for some of these things, quantity has a quality all its own. In my home automation example, if we assume that the AI has been able to ingest behavioral pattern data about hundreds of millions of people along with demographic information, maybe even genetic information that they bought from 23andMe (that's a really big I-freaking-told-you-so, btw - not "you" specifically, of course, but generally), and absolutely any other data point, and all of a sudden it can start to find patterns and make statistical inferrences that you, yourself, may not be aware of. So, yeah, maybe that very first time there's a bad day on the stock market, your home automation has already put together what you're going to want before you know what you're going to want. Insert a really long and interesting conversation about what that means for the concept of free will, right?
Yeah, it has a very bland writing style but that's a sort of feature, not a bug. It's very generic because it's looking at everything that it has ingested (so, pirated books and the Internet) and it then determines what word would most commonly follow that word given the context of the prompt. So using it to generate copy can produce results that aren't particularly fabulous. But it can be incredibly helpful to brainstorm ideas. For a project for school, I needed to come up with a fictitious hospital. I wanted it to be somewhat clever, but not super obvious. My prompt was "I need to come up with a fake name for a hospital. I want it to be St. [something] Hospital. I want the [something] to be obscure but humorous. I'm looking for a name that would be related to anonymity or obfuscation". It kicked back a list of 10 things that were actually pretty good ideas. But I didn't like them. So I refined with another prompt: "What about names that might come from classic british and american literature?" I got ten more suggestions that were also very good. In the end, I didn't use any of them because I came up with another idea, but it was that brainstorming session that led me to choose St. Fortunato Healthcare for my hospital. The literature suggestions got me thinking about Poe even though it wasn't one of ChatGPT's suggestions. Something I could have done on my own or with another person. But the LLM helped me do it by myself in a matter of 90 seconds.
There was someone else who did hit on a really good point, though, about data leakage. The next big thing in security is going to be finding ways to build LLMs in such a way that they can "know" who's allowed to know what. If the LLM has access to all the company data, for example, it might know what every employee's salary is. But a sales rep isn't supposed to know. So when the sales rep starts asking questions of the LLM, we need a way to prevent the LLM from inadvertently using that knowledge to answer the question. Telling it not to tell anyone the salaries is easy. But what if the prompt starts off by asking about costs that go into a product or something of that nature? That's going to be tough to do. The last time they told an AI to lie to people about what it knows, Dave Bowman got locked out of the pod bay.
I guess the long story short is that there is some there there. It's young and, like any tool, it can do a really good job when you use it for what it's good at and it can hack the E36 M3 of stuff when you try to make it do things it wasn't designed to do. Like any new hawtness, there's a ton of buzz and BS around it and everybody thinks that in order to stay ahead, we need to take out the blockchain and put in the AI for everything. And not all of those implementations make sense, are done properly, or are appropriate for the state of the tech at this time. Yeah, it gets stuff wrong. A lot. Yeah, it hallucinates. A lot. And, yeah, it's really bland and generic when it generates long form prose. But we're still figuring out the right way to train the models, where to get the data to train the models, and where the right time is to say "I don't know" instead of making a statistical inferance.
