“Tarred with the same brush” is a common phrase used to suggest that sometimes stuff that’s only slightly connected is lumped together and dissed for a fault of a single member of the group. We’re now seeing issues with a specific model of AI tar not only that model, and loosely related competing AI models, but also the cloud. Microsoft didn’t report a bad quarter, but the financial markets were disappointed by its AI news, and that even called its cloud services into question. There’s even a question of whether “open source” AI models are really open source. What’s really going on?
One thing that’s clearly at the root of this is the hype cycle. Wall Street loves a bubble, because hedge funds know that individual investors will buy into it, and that hedge-fund short-selling can then prey on them. The media loves a bubble because it generates a lot of clicks, and let’s face it, we love a bubble because it’s more exciting to read about how AI is going to make mankind extinct than to read about a new version of Ethernet. We all got what we wanted.
What’s unclear, as is often the case with bubbles, is what reality (if any) is behind and beneath. Is AI an abject failure, or is our view of its success simply distorted by the hype? The latter seems true. Have the early beneficiaries of the hype prepared to jump off the hyperbole into a successful reality? That’s less certain. The Infoworld article quotes Google’s CEO as saying “We are driving deeper progress on unlocking value, which I’m very bullish will happen. But these things take time.” Is that optimism justified? It depends on how we interpret it.
I’ve been a naysayer about generative AI and its hype from the first, because I’ve been skeptical about the model. What we typically see as generative AI is an application of large-language-model technology to a kind of personal knowledge resource chatbot. We can interact with it, ask it questions in plain language, and get a very human-sounding response. This got a lot of attention because it was approachable, not because it was actually really valuable. There are situations where I’ve found it helpful in my own research, but I probably do a hundred searches for every AI question I ask. I get wrong answers regularly (hallucinations), and even when I get a right one I’m certain I could have gotten it without generative AI, maybe a few seconds later. Thus, I’d not pay for the service.
That’s the underlying issue here. Do we believe that consumers, individuals, would pay for AI? Why, when they don’t want to pay for other things that are clearly valuable. We accept ad sponsorship in our online publications and content. We embrace open-source software where we can. Now we pay for AI? I think most people realize that AI will be successful to the extent that it gets paid for, and it’s almost certainly businesses that will be paying, which means AI has to make a business case.
Enterprises say that so far, AI chatbots in support or presale missions can make a business case in almost 90% of cases. Business analytics and intelligence do nearly as well. “Copilot” applications, so far, make a business case in less than 20% of cases, and most of them are for things like code review or document checking, not for true “generative” missions, and involve specialized models/training rather than generalized tools. In the chatbot and BI applications, enterprises want in-house model hosting because critical business/competitive data is involved.
A shift to in-house hosting of LLMs would open up more business cases, and thus promote AI overall while accepting, and maybe even accelerating, the decline of AI as a service, in its broad and generative missions. It would also introduce some interesting shifts. One thing enterprises tell me is that they want an open-source (which to them means free) LLM, not one they have to pay for, particularly not one with regular subscription payments. Another thing is that they’d love to see someone with credibility offer a “GPUaaS” service that can handle training tasks with sovereignty protection for the data. The final, perhaps most interesting thing, is that they’d love to see an AI hierarchy evolve.
What? Well, about a fifth of enterprises say that their ideal AI is self-hosted in a data center and augmented with AI running on client devices. They want to see AI distributed across both personal and hosting elements, and if sovereignty issues could be resolved they’d accept GPUaaS hosting where additional GPU capacity is needed. It’s particularly interesting to me that the interest in a hosted form of LLM is linked to self-hosting and distributability to PCs rather than a desire to consume LLMs in hosted form for their regular use.
Enterprises also see the need for a more modest GPU to deploy in a personal device like a PC or even phone/tablet. Only 15% think that NVIDIA is likely to be the source for these personal GPUs; the majority (65%) think AMD would be the most likely supplier, with Intel and Broadcom coming in for 8% and 5% respectively (the rest had no opinion). To me, it’s less significant who enterprises think would be the NVIDIA alternative than that they think there needs to be an alternative. NVIDIA may be tarred with that generative brush.
So may Microsoft, Amazon, and Google. All three of the cloud giants have a vested interest in monetizing AI as an extension to cloud computing, and two of them (Microsoft and Google) have a search business to protect and expand. Too much attention on past investments ties you to the past when you need to face the future, in AI as in other areas. What exactly Meta might think it would do to monetize its LLM work if it doesn’t offer a chatbot service or license the models on a subscription basis isn’t clear yet, but surely some form of open-source-free LLM is essential for the next phase of AI, and that may be a step Meta is more willing to take than its rivals.
Well, there’s plenty of tar left on the brush.
