Is an AI winter upon us? There seems a chill in the air

GPT-5, the long-awaited new model from Sam Altman’s OpenAI, was released earlier this month to a tepid reception. If it’s a step toward artificial general intelligence (AGI), as the company repeatedly said it would be, it’s a tiny one indeed. The model was so poorly received by some ChatGPT diehards that OpenAI was forced into an embarrassing rollback, making older models available again. Altman’s claim that GPT-5 was like talking to a “PhD-level" expert quickly became a joke.

Also Read: Outrage over AI is pointless if we’re clueless about AI models

At the same time, CoreWeave, one of the few pure-play AI stocks, plummeted more than 25% last week after guidance that spooked investors: Revenue growth is expected to be enormously outpaced by capital expenditure increases. And its IPO lock-up was coming to an end, which didn’t help either.

And while it’s hard to pin down just how beneficial AI has been (or will be) in the business world, one piece of research from McKinsey & Company should give everyone pause. While eight of out 10 companies surveyed said they were implementing Generative AI in their business, the consultancy group observed, just as many said there had been “no significant bottom-line impact." Gulp.

The reaction to GPT-5 in particular has longtime AI sceptics taking a victory lap. Author and journalist Brian Merchant noted that Altman seemed less willing to use the phrase ‘artificial general intelligence’ now that his latest AI tool is still so far from that. “I think it’s not a super useful term," Altman told CNBC. Merchant pointed out that it’s a term Altman has used often, including in February on his personal blog. It has been handy in raising billions of dollars.

Also Read: Siddharth Pai: Don’t be naive, Agentic AI won’t eliminate agency costs

Altman has had some other telling things to say over the past few days. “Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes," he told a group of reporters last week. “You should expect OpenAI to spend trillions of dollars," was another remark. My favourite, from a CNBC interview the week before, was: “It’s nice not to be public."

I’ll bet! Throughout this, I have been wondering how investors would have reacted to GPT-5 had OpenAI been a publicly traded company. I suspect at the very least, OpenAI would have had a CoreWeave-like week once investors looked at the user revolt, the backtrack-of-sorts on super-intelligence and the prediction that “trillions" of more dollars would be needed. Altman said he was confident the company could invent a “new kind of financial instrument for finance and compute" to fund its rapid industrial expansion. I guess we now know at least one book in the GPT-5 training data. 

Then again, the fallout hasn’t extended to other stocks tied closely to OpenAI’s fortunes, such as Microsoft or Nvidia. This suggests investor nerves have not yet been frayed.  One argument is that the embarrassing viral failings, such as not being able to spell ‘blueberry,’ are trivial stunts that miss the bigger picture: GPT-5 is more sophisticated in picking the appropriate model for a task, which, though it seems unremarkable, is actually practical and useful. Another view is that AI capabilities have improved sufficiently and groundbreaking AI agents can go off and carry out certain tasks that are just in reach—and the return on AI investment will finally kick in at that point. 

Also Read: What OpenAI has up its sleeve: An AI gadget mightier than the sword?

Or it may not, and this month will be seen as the beginning of something significant. I don’t quite think we can call it an AI winter yet—but there’s no question that a sudden chill is in the air. 

My takeaway from the GPT-5 launch has been that while AI companies can tout overall performance on various benchmarks, these are becoming increasingly less relevant. Impenetrable to anyone other than AI researchers, these scores mean little to the end user, be it the consumer or a CEO.

What sets the narrative around AI progress (or lack thereof) is its practical application; it’s here where all AI firms are still falling short. GPT-3.5 was around for months before the demonstration of ChatGPT wowed the world—until we all discovered its shortcomings, not through lab testing of millions of queries, but with our own eyes. AI agents might be the next ‘ChatGPT moment,’ if they do as promised. Call it the ‘Blueberry Benchmark’ of real-world usefulness. Better scores are needed urgently, or investors could be in for a frigid time. ©Bloomberg

The author is Bloomberg Opinion’s US technology columnist.