On the Long and Costly Road from GenAI to AGI
I read this week that equity strategists are now vigorously debating the market potential of genAI. It's about time. Universal acclamation is always suspect, just as rational debate is always welcome.
As it stands, genAI is performing parlor tricks, especially on Elon Musk’s X, where Grok (from Musk’s x.ai) serves as a fount for all manner of low-rent deep fakes and trite memes. Is that what genAI is for? Is it simply a tool for mild amusement and personal vendettas? The market maestros, cloud giants, and enterprise customers that span a wide swath of industries better hope the creative scope of genAI extends a lot further than cheap, ephemeral entertainment.
Immense costs are associated with the development and production of genAI, including expenditures on datacenter real estate, AI expertise, data sets, training models, IT infrastructure, and energy and cooling technologies. On a relative basis, some of these costs are likely to decline with the passage of time and the advent of innovation, but some costs are likely to prove more intractable than others. That’s why it’s difficult to envision a near-term horizon when genAI will be relatively inexpensive to produce and deliver.
That means genAI has to grow up and move past its adolescent stage of hijinks and practical jokes, maturing into adulthood sooner rather than later. It doesn’t want to become the information age’s answer to 3D movies, a novelty that burns brightly before fizzling like a damp squib.
We know what most CFOs and other C-level executives want from genAI. They want it to cut costs, quantitively boosting productivity and efficiency to the point that employees who perform back-office jobs and manage recurring closed-loop processes can be relieved of their duties. We all know this is what most businesses want from genAI, but we rarely speak of it because it’s taboo in the technology industry to acknowledge that any type of intelligent automation puts people out of jobs.
Inconvenient truths are to be avoided. You’re not allowed to admit, for example, that genAI, which is an incremental improvement on more rudimentary, discrete forms of intelligent automation, is attractive to many corporate buyers because of its perceived ability to make certain human business functions (ie jobs) superfluous.
But let’s have the courage here to speak the truth. If C-level executives are candid, they will concede that their expectation is that genAI will allow them to run leaner, meaner organizations. I think the disconnect is that their cost-cutting expectations are way ahead of the technology’s capabilities. The vast majority of employees, even those in the technology industry (which is currently “eating its own dogfood”), are safe from the depredations of genAI and automated bots.
Even so, the truth remains that corporate buyers of genAI want to use the technology first, if not foremost, to cut costs by making categories of employees redundant. That’s the plan, but genAI figurative and literally cannot do the job, or, at least, not enough jobs to justify the expense. If all that genAI does is moderately boost the productivity of most back-office and other process-oriented employees, I doubt it ever will ever fulfill the colossal expectations of enterprise buyers; nor will it provide a compelling return on investment for the cloud hyperscalers who are investing billions of dollars in the datacenters and underlying IT infrastructure that are prerequisites for the propagation of genAI services.
Escaping the Novelty Era
If you invest prodigiously in genAI infrastructure, you’re counting on genAI to deliver prodigiously on its promise. At the risk of belaboring the self-evident, if your spending on a given technology initiative massively exceeds the return on investment over a reasonable period of amortization, you will find yourself drowning in a pool of red ink rather than floating merrily in a tranquil sea of the black (or green) ink.
Yes, I grant that it’s still possible for genAI to redeem itself, for it to grow out of its prankster period and become a responsible adult, but, despite the happy talk of vendors and VCs, that outcome is not a foregone conclusion.
It’s here that I must make a distinction between genAI and artificial general intelligence (AGI). GenAI is capable of generating text, images, videos, or other data using generative models, frequently in response to a prompt from a user or process. GenAI’s models (such as LLMs) learn patterns and structures from training data and subsequently generate new probabilistic data with similar characteristics. Reading between the lines, you will reasonably deduce that genAI yields nothing more than a halting, partial intelligence, investing it with a necessarily limited scope of commercially valuable applicability.
Meanwhile, the still-aspirational AGI has been defined variously and rather broadly – as one would expect from a technology that has yet to transition from theory to practice – as having “an intelligence that may be comparable to, match, differ from, or even appear alien-like relative to human intelligence, encompassing a spectrum of possible cognitive architectures and capabilities that includes the spectrum of human-level intelligence.” (The quoted text is from Wikipedia, but it’s a relatively serviceable approximation of the slippery signpost that academics and data scientists have provisionally installed.)
When it comes to fruition (assuming that it does so), AGI, through its ability to reason at or beyond the levels of at least some humans, would have a profound and permanent effect on nearly every aspect of commercial and recreational human existence. It bears repeating, though, that today’s or tomorrow’s genAI is not and will not be AGI.
With that out of the way, we should note that OpenAI, the progenitor of ChatGPT, announced this week the launch of its new 'o1' model. While “o1” still slots under the rubric genAI, you can interpret it as a modest, incremental step in the direction of AGI. Make no mistake, OpenAI and its growing number of competitors have a long way to go before they get to AGI. Still, the fact that OpenAI is urgently trying to advance beyond the novelty era of genAI indicates an awareness among the vendor community that it’s time to get serious.
Cash-Munching AI Startups
Still, we’re not easily deceived. We know that OpenAI, like every other cash-munching AI entity, is perpetually on the fundraising circuit. That puts pressure on the executive teams at AI companies to pump up the volume and embellish the narrative in a bid to excite the enthusiastic interest of deep-pocketed investors. If you, as a genAI concern, can persuade existing and prospective investors that you’re striding along a non-hallucinated yellow-brick road toward a reasoning, thinking AI, well, how can the investors do anything other than thrown money at you?
Nonetheless, even allowing for the standard deviation of the genAI community’s carnival-barker bullshit, Sam Altman and OpenAI are absolutely correct in one critical sense: the last stop on the AI train, purely from a business perspective, cannot be genAI, which has been revealed as a mildly diverting destination with few attractions to hold the sustained interest of big-spending customers.
If AI is ever to prove commercially rewarding to parts of the supply and value chain that extend beyond the likes of “picks and shovels” purveyors such as Nvidia, then something beyond genAI, and closer to AGI, must be attained. Beyond the confines of religion and professional sports, suspension of disbelief can only be sustained for so long. Cloud giants will not pay indefinitely for infrastructure that does not provide a return on investment. One rearguard action against overspending on third-party GPUs involves crafting and deploying your own rather GPUs and AI accelerators rather than buying them from a supplier, but even that measure merely mitigates rather than eliminates the cost of infrastructure.
I anticipate that some of you might raise a reasonable objection, namely that genAI offers value propositions that extend beyond reductions in labor costs. Yes, it’s quite possible that genAI will result in process improvements that don’t involve labor displacement. In fact, I expect some of that to occur, but I don’t anticipate any of it to be sweeping or revolutionary. If you use genAI with guardrails that encourage information integrity and discourage confabulation or hallucination, you can save time on research tasks. (A quick aside: Neither of those terms – confabulation or hallucination – seems entirely accurate to me because of what seems a clear intent on the part of the owners of genAI to have the technology provide answers to questions, even if the answers are partial or complete fabrications. The nascent technology is being made to answer, even when it has no correct answer to offer.)
All that aside, my point is that the genAI is currently overmarketed and oversold to perform tasks it is woefully unsuited to address. The discrepancy between grand expectation and humble reality will jar unwary early adopters.
What about genAI’s capacity not only to displace jobs but to create new ones? All the VCs tout this scenario, but precious little has been written cogently about the specific roles and responsibilities of these future genAI jobs. Prompt engineer? How many people will be doing that as their exclusive job? It’s more likely that prompt engineering will become the responsibility of professionals already employed in existing roles in clearly demarcated areas of domain expertise. On the consumer side, no consumer is being asked to qualify as a “prompt engineer” to use Perplexity. If the jobs will come, we don’t yet know what they will be, or whether they will offer greater or lesser remuneration than the jobs that are certain to be displaced.
A Modest Prediction
What about genAI’s potential to produce new breakthroughs, in science, technology, medicine and healthcare, insurance, banking, or government? Anything is possible, I suppose, but find I hard to believe that the modest capabilities of today’s genAI are poised to strike the motherlode in any of those industries and their related disciplines.
So, now that we’ve done a lap around the circuitous genAI course, we’re back to the realization that any sustained and compelling payoff associated with AI will need to await the triumphant arrival of AGI. Until then, the customer value associated with genAI will be modest, incremental, and perhaps, given the market hysteria and bloated expectations that have preceded the technology’s arrival, commercially disappointing.
As of this week, OpenAI has put its cards on the table. It’s saying implicitly, if not explicitly, that genAI isn’t good enough to justify the hype it has generated and the enormous investment financing it continually solicits. The journey will be long, there will be many stops along the way, but the direction and the destination is clear.
My modest prediction is that OpenAI’s o1 launch will be followed in quick succession by similar salvos from other genAI purveyors indicating that they, too, are on a fast track to a more reasoning, thinking form of AI. When you look under the hood, you’ll likely see a genAI engine with enhancements, just enough of the latter to warrant, if the wind is blowing in the right direction and conditions are favorable, the new spin.
Make no mistake, however: What OpenAI told us this week is that AI, as a business proposition, must offer a lot more than what genAI has brought to the table.