Summer News Roundup: Storm Clouds on the AI Horizon

Continuing the recent trend of financial introspection, HSBC has recoiled from the ominous precincts of private-credit lending. HSBC is not the first major bank to reduce exposure to increasing risky private credit, and it won’t be the last. As noted in a Reuters article that builds on an earlier Financial Times report:

The FT report underscores the growing scrutiny ​of private ⁠credit portfolios, with regulators worldwide becoming increasingly concerned about banks' exposure to the $3.5 trillion private credit industry.
Wealthy investors have queued up to withdraw their money from private credit vehicles ⁠in ​recent months amid worries about weakening lending standards ​and fears of AI-driven disruption at software companies that have borrowed from direct lenders.

I am not trying to frighten you like a slasher-film schlockmeister, but the private-credit situation is scary. The dam could break in weeks, days, or months, and if it does, many tech companies — including those you might think are immune from the effects — will be impacted. If your portfolio is heavily weighted toward tech, as is likely given the disposition of major funds and many ETFs, you might want to diversify toward more prosaic investment vehicles. Just as you seek an uneventful flight, you probably want to avoid the protracted and stomach-churning turbulence of what appears to await us on the markets.

AI apparently has a lot in common with many people: it identifies extraterrestrial life where it does not exist. From a recent article in Popular Science:

In recent years, many researchers—including some at NASA—have advocated incorporating machine learning and artificial intelligence in their search for organisms beyond Earth. Some of these approaches may show promise, but new research indicates much of today’s AI is even more easily duped by false positives than their human operators. 
“No matter what sequence of commands we started with, we were able to fool the AI 100 percent of the time,” Ankit Gupta, a Michigan State University (MSU) computer science engineer, said in a statement.”

AI’s Pattern Misclassifications, Meta’s Creepy Glasses

You might excuse AI for its shortcomings in extraterrestrial identification, but these inaccuracies have profound implications. AI’s susceptibility to misclassify patterns, described as an “Achilles' heel” in the article, results in potentially serious false positives in realms such as facial-recognition software, self-driving vehicle, and medical imaging. Ankit Gupta, who is quoted above, concludes the article by saying that humans need to be in the loop to confirm the veracity of AI’s work. This seems sage advice, but it does diminish the value proposition of push-button AI autonomy that many vendors are selling.

I suppose the advent of AI doesn’t mean that Meta has given up entirely on the creepy surveillance dystopia of the metaverse. The company seems determined not only to kill the last vestiges of privacy, but to drive a stake through the corpse’s heart. I don’t know about you, but I don’t want Meta or the social-media masses to see and hear surreptitiously recorded private conversations and interactions. Meta has also demonstrated repeatedly that it is not to be trusted with wearables, as the following excerpt from this TechCrunch article notes:

All the while, the company is facing multiple investigations and lawsuits over Meta AI glasses privacy violations. One lawsuit comes after Meta notably canceled a contract with an outsourced tech firm after some of its Kenyan workers alleged they had to view graphic content, like sex, nudity, and people using the toilet, while training Meta’s AI using people’s Meta AI glasses’ videos.
These are hardly Meta’s first scrapes with privacy violations or safety measures, either.
Arguably, Meta’s reputation on privacy has been tainted for years after numerous leaks and lost lawsuits about its alleged lack of child safety measures and desire for growth at all costs. There are books by whistleblowers documenting its alleged abuses, not to mention previous large-scale privacy disasters, like the Cambridge Analytica data scandal and others.

While I concede that change is always possible and that past performance is not a guarantee of future results, the well-documented form in Meta’s privacy-and-trust book does not inspire confidence.

I am merely one of many who foresaw that Chinese companies would inevitably seek to reduce, if not eliminate, their dependence on U.S. technology. It’s the only logical response to export restrictions and threats of crippling supply embargoes. As such, it should come as no surprise to learn that DeepSeek is developing its own inference chips.

ChatGPT Goes to Work, Agentic AI Engulfs the All-You-Can-Eat Energy Buffet

In other news, OpenAI has launched ChatGPT Work, the last word presumably added to distinguish it from ChatGPT Just Screwing Around. The enterprise journey will now deliberate on whether the all-in cost of ChatGPT Work, including proliferating token taxes, is equal to or surpasses the output it produces. The verdict will be a fateful one, and not just for ChatGPT.

Chatbots are last year’s model, to paraphrase Elvis Costello, and agentic AI is now all the rage. But did you know that agentic AI’s appetite for energy makes genAI chatbots look like abstemious monks? An article in Gizmodo, citing research from the Korea Advanced Institute of Science and Technology, suggests that AI agents can consume up to 136.5 times more energy per query than generative AI models. The article offers an explanation for the stark energy-consumption discrepancy:

Typically, LLM requests are a call-and-response: a person enters a query, and the model responds. But agentic AI typically requires multiple steps to execute a command. To do that, the researchers said, the agent must continuously ping its model to generate a new response as it reasons through all of the steps of its given task.

Agentic AI overindulges in other areas, too. It’s complex call-and-response processes mean that “agentic AI can take 153.7 times longer than a standard query.” This represents a resource-consumption problem as well as a potential performance bottleneck. Fortunately, optimized network architectures, and enhances to underlying network infrastructure, can at least partly mitigate latency issues.

The energy gluttony of agentic AI is a more intractable concern. Without major advances in the energy efficiency of agentic AI, according to the Korean researchers, agentic AI could soon generate 13.7 billion requests per day, comparable to the volume of daily queries handled by Google’s search. The problem is, unlike Google search, agentic AI’s voracious energy appetite would require 198.9 gigawatts of power, or approximately half the current electricity consumption of the U.S. Presumably, if agentic AI proves as appealing as its purveyors believe it will be, the energy demand will only grow from there. Does that seem sustainable, or even desirable, to you?

Jaws featured the memorable line, “You’re gonna need a bigger boat.” Well, more energy, and bigger datacenters, isn’t the right answer to the agentic-AI conundrum. At some point, if agentic AI retains its shiny halo as the next big thing, we will desperately need a means of consuming a lot less energy while the agents make their circuitous rounds.