The AI landscape is changing. It’s not really obvious at this point because most people get their AI views by reading online stories, and they’re not tracking the reality of the shift at this point. Maybe they never will. Whether they do or not, we have to remember that AI adoption, the profits from AI, depend on its ability to deliver value proportionate to investment. It’s been clear for over a year that AI can deliver interest, entertainment. Value? That’s becoming clear now too.
What is AI? Artificial intelligence, which means an artificial form of natural (human) intelligence. We don’t really have that yet, and we may or may not reach it, so obviously we need another definition or we’d consign the whole thing to hype. Let’s approach it another way, and ask what the difference is between AI and a program. Most of you probably don’t remember the old chat-psychiatrist program so old that it was available before there was an IBM PC. It would ask you something, you’d answer, and you could have a sort-of-conversation. At the time, it seemed to many like the computer was emulating a real person, with real intelligence. Obviously, given the limitations of microprocessors in the 1970s, that wasn’t AI. We’ve also had chatbots for decades now, and most people can’t really tell whether they’re chatting with a real human or with a chatbot, even without “AI”. So what’s the difference?
Answer: Learning. AI has the ability to learn from experience, to learn from the real world and to then extrapolate what it learned into new situations it’s presented with. Yes, I know that it’s popular to separate “artificial intelligence” and “machine learning”, but I think that separation has lead us a bit astray. Any useful form of AI is really based on rules and not program languages, meaning that software creates a mechanism to absorb, to learn, rules and then to apply them in real-world situations. This means that AI has the ability to react to things other than the things a software developer explicitly coded in. It’s what makes, for example, an AI chatbot better than that 1970s program.
To most people, AI really means generative AI, which is AI designed to create something specific, like an image, music, or an answer to a question. In real-world terms, it’s a difficult association to map out because the expectations are set from the public-model examples from Google, Microsoft, and OpenAI. These models are designed to be essentially super-chatbots that can accept various (but usually textual) inputs and generate various outputs. They’re compelling to play with, but the problem enterprises continually gripe to me about is that giving a worker a thirty-buck-a-month tool to write better emails or press releases isn’t profound enough to change their business fortunes. Would they buy it? Probably if the worker’s productivity justified it, but they wouldn’t expect it to be more than an aid. I remember companies buying programmers any mechanical pencil they wanted, back at a time when writing code meant really writing it. Same sort of thing.
Public-model generative AI is trained on enormous reams of data, usually from the Internet. As I’m fond of saying, the Internet contains every possible truth and every possible lie, so there are obviously risks in using it as the font of all knowledge. One way to avoid the risk is to select sources considered credible in the training. We can assume that happens in the public-model applications. However, that opens the real generative AI question, a question that’s being asked by almost every enterprise who seriously evaluates AI, but isn’t talked about. If you train on selective material, can generative AI be made specialized? And if so, might the requirements for running it be much more modest? Could there be another kind of AI out there?
It’s getting into tax season, so let’s think of the question of filing tax returns. Suppose you’re a small business, and you have, or can create, the usual reports like a balance sheet, P&L, and so forth. These reports are fairly standardized in terms of content. Suppose that you could train a generative model on the instructions on filling out taxes, and on some accounting material that relates tax forms to business reports. Could you create a generative AI model to do taxes? My expert friends all say that you could. Would the process of training the model, or running it on current business data, require a data center full of GPUs? My friends say it would not, and so we can see that using a generative-AI approach to specialized business analysis can produce valuable results with limited resources.
This isn’t a brilliant personal insight (but you can call it that if you want to). IBM has seen this all along. Meta and Google have awakened to it as well. You don’t need to field questions from high-school kids on states and capitals, or write their homework for them, or help a bored office worker answer an email. You need to analyze stuff and make comments and recommendations. That’s actually a lot easier, because the knowledge base needed to train an LLM is much smaller and the resources needed are as well.
Google’s Gemma is a new open LLM that can run on a laptop, even on a phone. We can see consumeristic AI models developing, and we can also see that this form of AI is simply a feature of the device. Will phones with AI chips be more expensive? We already know they are not, but they might be more marketable. They might hold their value a bit longer in a world of smartphones where differentiation is increasingly difficult, and commoditization increasingly likely.
How about business AI? Trivial AI features in applications, even the integration of AI with Microsoft’s 365 suite, aren’t a revolution, they’re just a feature, not all that different from spell-checking. The real progress in AI is the progress in defining how a more scalable LLM could be applied to the specifics of business intelligence. Yes, a business has to empower workers to be more productive, but only a radical change in how work is done would likely make this sort of incremental, per-worker, gain enough to change the business fortunes. What’s needed to do that is a set of radical insights about how the business should function as a collective, a tool that would analyze all that’s available, all that could be done, all that’s known about markets and marketing and sales, and distill out a set of insights that would then change basic operations, That’s not a feature, it’s an application, and it could be a game-changing one.
How difficult do you think it would be for a sales manager to pick up a sales performance report from a random company and pick out the top players? For an accountant to assess the “books” of any random small company? Not at all difficult, because this sort of stuff is based on a very narrow set of data. But surely less difficult than it would be to assess the health of a big company, to diagnose problems in a global network, or to assess opportunities in a major market based on market demographic data and a company’s own information. The really hot AI stuff is probably not at the laptop-level, nor is it in the domain of the big public models. It’s in the middle, which of course is where most opportunities are usually found in the end.
This is good news for the chip vendors. Public-model and public-cloud AI are both justified by economies of scale, and GPUs and AI chips of any sort are part of the infrastructure they aim to optimize. If every enterprise built out private AI clusters, the total GPU consumption, by my estimate, would be well over twice what it would be if everyone relied on public infrastructure. But even that’s meaningless, because my data from enterprises says that the total usage of AI by enterprises would be at least three times greater if they were able to secure their data on premises. So GPU upside is six times what we’d see without that middle-ground AI, which is a darn good upside.
But a better one is lurking higher on the food chain, closer to the things that perform the critical function of justifying AI investment. All of these depend on realizing the benefits of AI, really realizing them. Right now, AI is a consumer fad that’s eased its way into becoming a worker fad because workers are people too. It’s transforming into something more, but it’s not there yet, and in fact it hasn’t even transformed enough for us to be certain about what will drive it.
Right now, the business world is focused on cutting costs. Raising revenue is not an option for most of them, at least not a realistic option. Cutting costs means cutting spending, which means both capital spending and operations, which is people. When companies cut capex, they’re cutting each other’s revenues. That raises the pressure to cut costs. When they cut people, they cut consumer spending, which cuts revenues, and so it goes on. In other words, we are in a consolidation cycle. Early on, but in.
The alternative is to focus on raising revenues. Find new things people and companies want to buy. Since all of IT is about efficiency, productivity, that means getting more production out of the same resources, or at least getting a larger production gain than the gain in resource costs. The challenge is that this is not only a shift in a mindset that’s settled in for two decades, it’s also bucking the results of that mindset on buyer behavior.
Is AI a consumer gimmick, a kind of faceless Facebook? Or is it an actual valuable aid in our lives. The thing that’s different about AI now, in February of 2024, is that more and more players of various types are recognizing how important that choice is, and they’ve made the second of the two already. I think that the results of this shift will become clear by fall.
