Well, China rattled investors in AI with its DeepSeek model. NASDAQ futures were down over 800 points at one point I saw on Monday, and of course this is surely do to short-selling by hedge funds. But you don’t short a market that’s not at least a bit over-bought, meaning that a lot of Street professionals thought there was a good chance that AI was over-hyped. That may be why there’s a move afoot to transition from “generative AI” stories to “agentic” and “inference” AI. The big question is whether this a flight from hype or just a search for a new balloon to grab onto.
Like AI overall, there isn’t much truth behind what you’re probably hearing about DeepSeek and inference AI. In my opinion, stories about DeepSeek conflate multiple models the company has released, mixing the latest inference model with past models that are more like generative AI tools we’ve had for some time. Add to this the fact that a lot of my sources say the company itself isn’t exactly forthcoming on details, and that many of the user comments on DeepSeek may be planted, and you don’t exactly have a strong basis for examination.
The market reaction is, again in my view, almost entirely due, as I said above, to hedge fund short-selling, something a Street resource indirectly indicated in a comment that many funds had already started to believe the generative AI bubble had perhaps gotten too big. It’s also interesting to note that the market dump had started to reverse itself even by end of day Monday, was up on Tuesday, and NASDAQ futures are up this morning. Lesson: We can’t really take a lot from what happened on Monday, so we need to get to AI basics to assess the situation.
AI, like any other technology, succeeds because it delivers enough value to users to generate revenue for providers. There are two markets operating that would have to deliver on this—the consumer market and the business market. Consumers want everything free, meaning ad-sponsored, and that means that new consumer AI would have to either break consumers of their online-equals-free habit, or compete with everything else that’s fighting for ad sponsorship. Google’s publicized fight against the use of ad blockers demonstrates that consumers are weary of the increase in the number of ads, and nothing has emerged so far to break the “free” habit. The other market is business, and for business you need a business case.
Generative AI has focused on being able to produce stuff people are interested in and can engage with, so they’re trained on a broad range of things and they can be made available to anyone online. It’s the approachability of this model that has raised its profile so much, but for businesses it’s proved problematic. The business case for generative AI is something like “you make your company better by helping employees individually with routine tasks, like emails and documents”. Enterprises have been skeptical of this from the first, and I think that it’s clear that AI providers know that, and know some successor concept is needed.
“Agentic” AI, if it means anything, means AI acting as a specialist agent rather than imitating a person verbally (a chatbot), or in producing an image or video, as generative AI has done. The specialist mission means that AI agents don’t need hundreds of GPUs to run on, and they can be integrated with applications more readily, running locally where data governance policies don’t restrict what they can operate on. Business analytics is an example of this, and so is most missions of AI in operations facilitation.
Agentic AI, though, is proving hard to hype up for two reasons. First, AI specialization is a topic that is interesting to a very few, so in a world dominated by SEO and clicks it’s a loser. Second, those who are interested in it are interested because generative AI has already failed them, which means they probably have a fairly good understanding of what they need and how it would have to work. A few glowing platitudes aren’t going to move their needles much.
“Inference” AI seems to be the next candidate to toss platitudes at. Proponents say that it’s different because it makes predictions, but enterprises so far are unmoved by that definition for a couple of reasons. First, recalling that business analytics is the AI target most believe in, and given that business analytics is designed to let management predict optimal changes based on past experiences, the inference AI presumption must be that management would cede this task to AI, which hardly anyone is prepared to do. Imagine you’re a CEO, telling your shareholders you intend to let AI run your company.
You may have noticed that “autonomy” is a common theme here. Generative AI does stuff people do. Inference AI does stuff planners do. Agentic AI’s classic definition includes autonomous operation within the agent’s range. The notion that to be useful, AI has to replace human functions rather than enhance them, to displace people rather than empower them, is a source of a lot of stories—AI is going to steal your job, if it doesn’t just kill you. The problem with this is that it raises the bar on acceptance significantly. CIOs don’t tell me that they, or company management, is looking for this sort of AI, but rather that they want AI do help workers along, workers at many levels. OK, in some cases this might actually end up fully automating some tasks, but computers have been doing that for seventy years. Think of AI as a new employee; do you just let them do everything or are they tried out under supervision? We need to introduce AI into business to learn what we can trust it to do, not let it take over on its first day.
How about “inference AI” and DeepSeek, which may or may not be an example of it? There’s a very fine line between “generative”, “inference”, and “agentic” even if we allow for the tendency to wash everything with popular tech terms. How many stories have run about someone asking generative AI what will happen this year, or at some point in the future? Even basic forms of machine learning can be used for simulation, which is surely a predictive function.
The real question may be practicality. To even argue whether inference AI is actually “predicting” or “seeing the future” is great for generating clicks, but what business plan isn’t aimed at that goal? Do we really think we’re going to let AI do business plans instead of people? Do we send a GPU to jail for duping us, or the model, or the AI entities in the company who created the model? Come on, people, this is just crazy.
The good news here is that the AI agent concept (I avoid the “agentic” tag because the term already carries too much baggage) aligns well with what enterprises have told me they wanted AI to do from the first. Recall that they liked AI integrated with existing applications, particularly business analytics. Of the productivity-related missions of generative AI, the ones with the most credibility are the “co-pilot” applications of AI that are integrated as agents for a specific task (of course, some of those with the least credibility, like document and email assistants, are likewise integrated, so the value of the application and the gains AI can support are important).
China’s DeepSeek is disruptive to stocks because short-sellers know there’s a lot of hype to AI valuations. However, it’s also a warning that it’s lightweight AI suitable for agent deployment, and not artificial general intelligence, that matters. The way that the rest of the AI players respond will answer a question I asked in the past; did they really believe in generative AI’s ability to return on investment, or were they simply riding the hype wave while they did something useful. I suspect we’ll get the definitive answer shortly.
