Hiltzik: Have we already passed the peak of artificial intelligence?

You only have to go back about three months to find the peak of AI hype on Wall Street. That was on June 18, when shares of Nvidia, the Santa Clara-based company that dominates the AI-related hardware market, peaked at $136.33.

Since then, shares of the maker of high-quality computer chips used in AI labs to develop their chatbots and other products have fallen more than 22 percent.

This includes a 9.5% decline on Tuesday that equated to a $279 billion drop in market value, the biggest single-day drop in value of a U.S. stock ever.

If you want, you can shed a tear for Nvidia founder and CEO Jensen Huang, whose net worth (on paper) dropped by nearly $10 billion that day, but pay closer attention to what the market movement might say about the status of artificial intelligence as a hot technology.

This is not a pretty picture. Companies that have rushed into the AI ​​market out of fear of missing out on useful new applications for their business have found that the benefits are hard to achieve.

The most pressing questions in AI development, such as how to stop chatbots from inventing answers to questions they cannot answer – “hallucinating,” as developers call it – have not been solved despite years of effort.

In fact, some experts in the field report that the bots are getting dumber with each iteration. Even OpenAI, the leading developer of AI bots, admitted last year that its GPT-4 chatbot performed “worse” on some tasks than its predecessors.

As for the prospect of AI enabling business users to do more with fewer people in the office or on the factory floor, the technology generates errors so frequently that users may add employees just to double check the bots' output.

A CEO whose company uses AI to analyze the statements executives promised investment analysts on previous earnings calls told the Wall Street Journal that the statements were “60 percent right and 40 percent wrong.” (The system even misstated his name.) He didn't say it, but a system that produces errors nearly half the time is simply worthless.

Business customers have reason to be concerned when deploying AI bots without human oversight. Even at relatively simple tasks like processing a routine legal file or answering simple customer service inquiries, AI has failed, sometimes spectacularly.

In a recent case, a marketing consultant used artificial intelligence to create a trailer for Francis Ford Coppola's critically-dumped new film, “Megalopolis,” by printing out critical reviews of his previous films, including “The Godfather.”

Variety reported that the bot fabricated negative statements and linked them to the names of critics who actually liked the earlier films. Last year, a New Zealand grocery chain's AI recipe bot recommended that users mix bleach and ammonia into a thirst-quenching drink. In fact, the combination is potentially deadly.

These are tasks where current AI models should excel: producing standard texts, advertising copy, customer service information that can be accessed at the touch of a button on a touch-tone phone.

That's because bots are developed by feeding them unimaginably large amounts of published works, Internet posts, and other mostly generic text materials. The developers then apply algorithms that allow the bots to provide answers that resemble human speech but are based on the probabilities of a particular word following another – one reason the bot action is often dismissed as “autocomplete on steroids.”

However, these systems are not “intelligent” in any sense. Although they produce simulacra of coherent thinking, they are often not powerful enough for tasks that require human judgment – for example, medical diagnosis and treatment.

In a British study published in July, 1,020 mammogram results were evaluated by both human experts and an AI system. The human experts detected 18 breast cancers that the AI ​​system had missed, while the AI ​​system detected only two that the human experts had missed. “Shouldn't AI/machine learning be better by 2024?” asked Michael Cembalest, chief market strategist at JPMorgan Asset Management, in an analysis of the AI ​​market this month.

Other studies suggest that AI can be helpful in diagnosing disease, but only when used as a technological tool under the supervision of human doctors. Diagnosing and treating patients requires “emotional intelligence and moral agency, qualities that AI may be able to mimic but never truly possess,” bioethicists at Yale and Cornell wrote earlier this month.

A persistent concern about AI is its potential for misuse for nefarious purposes, such as facilitating the shutdown of a power grid or the meltdown of the financial system, or producing deepfakes to deceive consumers or voters. This is the subject of Senate Bill 1047, a bill in California awaiting the signature of Governor Gavin Newsom (who has not yet said whether he will approve it).

The bill requires safety testing of advanced AI models and the introduction of “guardrails” to ensure they do not slip out of the control of their developers or users and cannot be used to produce “biological, chemical and nuclear weapons, as well as weapons with cyber offensive capabilities.” Some AI developers support the bill, but others condemn it, claiming the restrictions will drive AI developers out of California.

It is true that some of the risks considered seem unlikely in the foreseeable future, but the plan's initiator, Senator Scott Wiener (D-San Francisco), says the plan is not intended only for distant eventualities.

“The focus of this bill is how these models will be used in the near future,” Wiener told me. “The opposition regularly tries to belittle the bill by saying it's all about science fiction and the risks of 'Terminator.' But we're focusing on very real risks that most people can imagine and that are not futuristic.”

This leads us to doubts, not about the risks of AI, but about its actual benefits for business. These have spread throughout the industry as more companies try to use it and find that it is overrated. For example, a survey by the Boston Consulting Group last year found that “when solving business problems,” using the most advanced version of OpenAI's GPT chatbot “resulted in 23% lower performance than the control group.”

As the consultants noted, “it is not obvious when the new technology is appropriate (or not), and the persuasiveness of the tool makes it difficult to detect a discrepancy. … Even participants who were warned about the possibility of incorrect answers from the tool did not question its results.”

Some investment analysts say so much has already been invested in AI that the major developers like Microsoft, Meta and Google may not see any profits for years, if at all. In the coming years, Goldman Sachs analysts reported in June, “tech giants and others will spend more than a trillion dollars on AI… and so far little has come of it. So will this huge investment ever pay off?”

Nvidia's dominance of the AI ​​hardware market raises questions about the impact on financial markets if the company stumbles financially or technologically.

Cembalest of JP Morgan titled his analysis of the future of the market “A severe case of COVIDIA”. AI “is driving the [venture capital] ecosystem,” he noted, which produced over 40% of the new “unicorns” (startups valued at $1 billion or more) in the first half of this year and was responsible for over 60% of the valuation increases of venture-backed startups.

The instability this brings to investment markets was evident on Tuesday, when Nvidia's downtrend helped push the Nasdaq Composite Index down more than 577 points, or 3.26 percent.

Nvidia's decline was fueled by forecasts of a slowdown in its growth, which has been spectacular so far, and a report that U.S. regulators had served the company with a subpoena over antitrust concerns. (Nvidia later denied receiving a subpoena.) The overall market also declined, partly due to signs of slowing U.S. job growth.

Others warned that the hype surrounding artificial intelligence was over or that the potential of this technology was being exaggerated.

“At the top of the pile of inflated expectations in finance is generative AI,” warned technology consulting firm Gartner last month. “AI tools have generated tremendous attention for the technology over the past two years, but when finance functions adopt this technology, they may not find it as transformative as expected.”

That may be true of the projected economic benefits of AI in general. In a recent article, MIT economist Daron Acemoglu predicted that AI in the U.S. would only lead to about a 0.5 percent increase in productivity and about a 1 percent increase in gross domestic product over the next decade—just a fraction of typical economic projections.

In an interview for the Goldman Sachs report, Acemoglu noted that the potential social costs of AI are rarely considered by economic forecasters.

“Technology that has the potential to provide good information can also provide bad information and can be abused,” he said. “A trillion dollar investment in deepfakes would add a trillion dollars to GDP, but I don't think most people would be happy about it or benefit from it. … Too much optimism and hype can lead to premature adoption of technologies that are not yet mature.”

Discussions about the future of AI continue to be dominated by hype, largely generated by AI companies like OpenAI and their corporate sponsors like Microsoft and Google.

The vision of a world recreated by this seemingly magical technology has attracted hundreds of billions of dollars in investment. But if all of that suddenly evaporates because the vision turns out to be murky, it would be no surprise. It wouldn't be the first time something like this has happened, and it certainly won't be the last.