Identifying what AI you'll use next

One of the hardest things in assessing how AI can help your business is grasping how fast it will improve. Bill Gates recounts seeing GPT-2 in June, 2022, and being unimpressed. He challenged the team to come back when it could pass the AP Biology exam. To his surprise they were back in only 2 months with GPT-3. It not only aced the exam but went on to provide empathetic advice on other topics. ?Figuring out to apply AI requires insight into what it will be able to do next.

The pace of innovation in AI is astonishing, but the breakthroughs are not random. Here are three avenues where we can expect big improvements in the next year or so.

·?????? Visual information

·?????? Quantitative information

·?????? Taking actions

Visual information. Today’s GenAI deals mostly in words, though DallE and similar programs can create images based on word prompts. Thus far GenAI’s ability to take in information visually is limited, but signs are that is about to change.

Today language models can be coaxed into reading simple charts and graphs in a text document. However, in fields such as finance, engineering and science much of the critical information is presented in complex tables, diagrams and schematics. GPT-4V can go much farther. It can meaningfully describe multiple elements in a photograph and even infer some actions or causation (eg: girl blowing out candles on a birthday cake, barefoot getting ready to step on a nail.) Specially trained versions are learning to read X-Ray and CT scans. What this doctor’s helper lacks in bedside manner it will make up for in vigilance.

The evolution is vision is a good example of AI’s need to gestate, to get used in arenas where good enough is good enough. Two popular arenas for AI to improve are entertainment and accessibility for handicapped users. ?You can get insight into what’s next by looking at the AI in both.

·?????? Entertainment: It’s no accident that virtual and augmented reality are being deployed in gaming. Critical applications will come once the tech is trustworthy. ?Translation software was used in entertainment long before it was used at the UN. AI dubbing won’t win any Oscars soon, but it is being used to make an increasing amount of video content accessible in other languages.

·?????? Accessibility: Early transcription software automated closed captioning and let hearing impaired people participate in video calls in real time. Text-to-speech has allowed the blind to hear written information acting on their own. ?The American Foundation for the Blind’s list of AI-assistive tech products is a notable counterpoint to those who dis AI.

The importance of good enough is most apparent with self-driving cars. Autonomous driving has been just around the corner for at least 10 years. It turns out that filtering the massive visual input of a moving car for the one thing that requires attention is something humans are very good at, and computers are not. Our aging population and urban sprawl mean that self-driving cars will remain a tantalizing prize, but backers are having a tough time convincing governments that what they have now is good enough. In contrast autonomous moving robots are demonstrating they are good enough in warehouses.

Quantitative information. It’s counter intuitive that the cutting edge of AI is working with text vs. all of the numbers that are stored in computers. GenAI can be coaxed to do simple math, but nothing more. From everything I’m seeing I believe the main way we will use GenAI to access quantitative information is by making a chat layer over the existing databases, For example, if your sales data is linked to PowerBI you’ll be able to ask questions about it in natural language and have the AI translate those into the commands that PowerBI needs to pull the data. The GenAI will use it to create answers. In this way you will be able to feed it last month’s sales analysis and have it update it noting trends and anomalies. I’m pretty confident that the AI will take over the chore of pulling the charts and graphs together, leaving the analyst to focus on what it means.

Taking actions. Today most GenAI only deals with information, but the next step is taking actions. AI developers talk in terms of agents and assistants who can link to other programs to do things. This is when the rest of us get our shot at a Hollywood personal assistant.

It’s early days for this. A typical example is feeding the GenAI assistant a PDF of school holidays and ask it to add them to a family calendar. For those of us without software backgrounds it’s nice but not a wow. However, what gets the software pros excited is that the person creating this agent didn’t instruct the GenAI in how to deal with the calendar software. Once it was given permissions the AI wrote the program itself. The technical term is self-orchestrating and it starts to touch on the scary scenarios. We’re still a long way from HAL, but you can understand why some people are nervous.

Net: AI improves rapidly when it is used. If you want to see what’s next keep an eye on gaming and accessibility.