Tuesday, July 23, 2024

New INC. Magazine column from Howard Tullman

 

The AI Hype Machine is Running on Empty.

After dumping hundreds of billions of dollars into AI startups, investors are discovering that the payoff to date has been extremely underwhelming. 

EXPERT OPINION BY HOWARD TULLMAN, GENERAL MANAGING PARTNER, G2T3V AND CHICAGO HIGH TECH INVESTORS @HOWARDTULLMAN1

JUL 23, 2024

 

In a column in January I noted that in the practical world of business, where real results matter rather than hype and bragging rights, the smart players were starting to back away from their substantial commitments and investments in generative A.I. tools and projects. Especially the guys who write the checks and keep score. Yeah, they were all still talking a good game, but fewer and fewer of them were putting their money where their mouths were.  

The main reason seems to be that the near-term prospects for seeing concrete growth and improvement in revenues as opposed to cosmetic reductions in admittedly overbuilt headcounts aren't very encouraging. In many cases, any paths to eventual bottom-line benefits weren't even apparent because the operating costs of these new large language model (LLM) engines are so high that the businesses were spending serious capital dollars to generate digital dimes - if they were lucky. Compelling, substantive use cases for these tools as opposed to novelties, chatbots and toys have been few and far between.

What has really been emerging is the fact that, in addition to AI being ridiculously costly and resource intensive, after all the manipulation of the underlying data is done, you still need to hand off the output to a human being to actually get something done. Instructions aren't the endpoint of virtually any process - whether it's manufacturing, medicine, or movement - it's real-world implementation and execution by people that ultimately gets the tasks done.   

Things might be getting done faster, but it's by no means certain that the outputs are better. And it's absolutely clear that these outputs aren't new or innovative because they're ultimately constrained by the limits of their training data to making what amounts to best guesses at what's next based on what's happened in the past. You still can't Google the future. And if you're not smart and sharp enough to ask exactly the right questions in your prompts, you get garbage for an answer. There's no big prize for even having the best answer to the wrong question.

Bumping the speed and the scope of analysis or review may create some efficiencies, but these "improvements" don't add "intelligence" until the outputs are evaluated and employed by the human end users. No one's willing to turn these systems loose until their results, findings and conclusions have been vetted and fine-tuned by humans. Hallucinations might be a more polite term than lies or fabrications by the machines, but ultimately no one is going to trust them with our lives or our livelihoods any time soon.

There's still talk about the next generation or newest black box that will work some kind of magic that not only scales but shrinks costs as well, but there's no evidence that it's anything more than a pipe dream about a new version of Moore's Law. One of the flaws in this analysis is that the underlying foundation of Moore's Law is that experience is gained in production over time, which enables exponential enhancements in the circuitry. Sadly, to date, it's clear that in the GPT world we're interested bystanders at best and, while it's fascinating to watch, we rarely learn much of anything beneficial that will let us improve the process. Nor is there evidence that simply by adding more data and more computational power, we do anything to improve or expand the outputs so that they become self-effectuating and autonomous.   

Interestingly enough, we are finally starting to see that even the shameless hucksters and promoters on Wall Street are taking a hard second look and changing their tunes from rabid generative A.I. boosterism to a far more tentative endorsement that smells more defensive than aggressive. Research reports, press conferences, and presentations from most of the leading financial firms, led by Goldman Sachs, are beginning to observe and report on the empirical evidence in the field, which suggests that they may have completely misunderstood what's happening with this latest technology. Two key things are becoming obvious and each of them is largely contrary to the speeches and spiels we've been hearing from these guys for the last two or three years.

First, there are entire industries where the ultimate impact of these kinds of tools will be largely immaterial - construction is a good example. Even Goldman Sachs suggests that only around 6% of the fieldwork in construction and extraction businesses will be automated and the productivity improvements would most likely simply be a wash for the additional costs. There may be some augmentation but even those tasks will continue to be directed and executed by onsite workers. Fast food, customer service, and transportation will be other areas where it will be very difficult - without sacrificing the quality of engagement and experience - to dramatically reduce personnel. We've already seen all of the major QSR players, including McDonalds, take steps to back away from some of their initial AI implementations.

Second, the most likely jobs to be eliminated in large numbers through substantial task automation (30%-to-50%) are NOT likely to be the low-paying positions (no collar and blue collar), which require physical labor and direct interaction with customers and co-workers. Instead, white collar and new collar (knowledge workers) positions including administrative jobs, legal work, financial analysts, marketing, and writers and editors will take the hit. Unilever in Europe is already leading the pack in this workforce pruning.)  The two critical defining characteristics of the targeted jobs will be that (1) it is very difficult in many of these cases to directly measure productivity and (2) senior managers looking for easy cuts and economies with only passing concerns for content, originality, innovative analysis and quality will happily trade out these positions for machine-created material that may well be drawn and lifted from other similarly situated creators.

When you look closely, as everyone has finally begun to do, the only conclusion you can make is that unless you're Nvidia and basically producing the picks and shovels for this industry (and largely without competitors), there's unlikely to be very much there there. And what is there will inure (as usual in tech) to the biggest of the big guys. As the saying goes, when the elephants start dancing, the grass takes a beating.