And now for the main event…

When OpenAI launched its highly anticipated GPT-5 model last week, the company almost certainly didn’t anticipate the public reaction, which ranged from ambivalence to downright hostility—a rare setback for showman impresario C.E.O. Sam Altman. Over the past few days, social media has been inundated with viral posts chronicling the unfortunate screw-ups of Altman’s self-described, pocket-sized “Ph.D.-level” intelligence system —struggling with basic math, drawing wildly inaccurate maps of the world, a persistent inability to count the number of Bs in blueberry, etcetera. A petition calling for the reinstatement of the older GPT-4o model, which was replaced with GPT-5, got more than 3,000 signatures in about a day.

Reddit was full of people literally mourning the loss of the older models they had developed emotional attachments to. (Time to get a dog, maybe?) Others offered more serious criticisms. “This is one of the biggest disappointments in A.I. history,” said computational linguist Maria Sukhareva. “It’s not even that the model is bad, it’s that the unrealistic expectations, fueled by outright lies about what transformers can achieve, inevitably led to a loss of trust and frustration among those who had high hopes for this release.”

The deluge of criticism seems to have put OpenAI on the back foot, at least for now. Over the weekend, 4o was brought back as an option for subscribers to the Plus tier, which costs $20 per month (nifty sales trick, there). ChatGPT head Nick Turley said that the company plans to “monitor usage as we decide how long to offer legacy models.” Altman, meanwhile, found a way to spin the mess: “Long-term, this has reinforced that we really need good ways for different users to customize things. Some users really want cold logic and some want warmth and a different kind of emotional intelligence. I am confident we can offer way more customization than we do now while still encouraging healthy use.”

Critically, however, the perception that GPT-5 would essentially embody the heavily fetishized idea of artificial general intelligence—one of Altman’s talking points and an industry hobbyhorse—has all but vanished. Altman called this misadventure a “step along the path” to A.G.I., but later corrected himself, saying that this nebulous concept is “not a super useful term.” Of course, this was quite a turnaround from January, when Altman contended that OpenAI is “now confident we know how to build A.G.I. as we have traditionally understood it.”

For Gary Marcus, though, all this made for a pretty satisfying weekend. Marcus, a cognitive scientist who sold his machine learning company, Geometric Intelligence, to Uber and has written a half-dozen books on A.I., has been studying the technology for decades. As Chat and OpenAI have firmly established themselves in the public consciousness, Marcus has become one of the industry’s most recognizable skeptics, repeatedly pointing to the inherent limitations of L.L.M.s, and the unrealistic promises of the major labs—all while taking a pretty relentless stream of flak from the true believers. In 2022, he published an article declaring that deep learning—the technical approach that led to the models underlying ChatGPT and its competitors—was hitting a wall. Pre-training transformers, he wrote, had reached a point of diminishing returns.

The underwhelming release of GPT-5 validated many of Marcus’s arguments regarding the fundamental roadblocks around reliability, abstract reasoning, truthfulness, factuality, and hallucinations. It also seemed to bolster his longstanding argument that A.G.I. will not emerge from neural networks alone. While some did so reluctantly, plenty of industry insiders were quick to give Marcus credit, acknowledging that he might have been right after all.

When I caught up with Marcus on Monday, he told me that it was “nice for people to see what I’ve been seeing. Sam really overplayed his hand—by a lot. Now his bluff is apparent, and I’m the one who called the bluff. That’s what it is.” Echoing impressions I’ve heard from a number of sources, he said that OpenAI should have chosen a different name for this model (GPT-4.6, perhaps?) and saved “GPT-5” for something more august. The launch, Marcus said, was a “blunder” uncharacteristic of Altman, whom he ordinarily finds to be very calculating. “There must have been something that drove him to do it,” he added. “But from the outside, it’s really a puzzling decision.”

For the past year or so, Marcus represented a minority voice within the A.I. industry. The “wall” he prophesied in 2022 was seemingly breached in 2024, when OpenAI launched its first “reasoning” model. This approach later became ubiquitous across Silicon Valley. These models, which arrived just as insiders were getting nervous about the diminishing returns of stand-alone neural networks, are made possible through the integration of so-called “chain-of-thought” (CoT) prompting, which forces a language model to produce multiple strings of tokens before outputting a final answer. The result was a boost in capability—alongside an increase in energy intensity—and an industry-wide maneuvering around Marcus’s wall. At least, that was the idea.

But the fundamental limitations remained; capability, as measured by a benchmark, and real-world efficacy are not the same thing. “They found another technique. It brought a little bit of life,” Marcus told me. “But it didn’t get them that far. The things that I was pointing out—which are failures to reliably reason, hallucinations, lack of groundedness in the truth—those they really have not made substantive progress on. Yes, you can get it to work in certain cases, but those problems remain unsolved.”

This has borne out across a number of recent studies: an Apple paper on the “illusion of thinking” in L.L.M.s; a Salesforce paper on the “substantial gap between current L.L.M. capabilities and enterprise demands”; and, most recently, a preprint out of Arizona State University, which found specifically that model CoT “becomes fragile and prone to failure” when dealing with patterns that were not encountered during training. “I think people got confused because there’s progress on some dimensions,” Marcus said. “And that’s true—graphics are much better. But there are a certain set of things we don’t know how to get past. What GPT-5 shows us is, hope is not a solution.”

While OpenAI pursued CoT, Marcus has long been an advocate for a neurosymbolic approach, which would combine the strengths of neural networks (pattern recognition) with those of symbolic techniques to add a much-needed dose of robust logical reasoning. “We need to start over from scratch,” he told me. But an industry pivot doesn’t seem likely, considering the extraordinary sums of money these companies have gambled on the current, brute-force approach. Altman, discussing the importance of investing in more and more compute, recently told CNBC he’s willing to “run at a loss for quite a while,” asserting that “it’s nice to not be public.”

Indeed, the notion that a large enough language model will somehow become superintelligent might be too compelling to abandon. There’s also a certain pragmatism to telling that story, since it helps keep the venture dollars flowing. (OpenAI, which loses billions every year, has a larger market cap than Shell…) But combine these factors with the reality that GPT-5 was not the breakthrough that was promised, and Marcus believes it could lead to a “wake-up call” in the A.I. race. “People might say, Yeah, we better stop promising that the next big model is really going to be the answer, maybe become more transparent about what we’re doing, and try to rescale expectations,” he said. “The problem is, the money to build this system is so enormous that realistic expectations may not merit the money they need.”

The A.I. industry, after all, is largely sustained by an internal loop. Startups invest in expensive chips with vital V.C. money, and hyperscalers invest in chips with preexisting revenue streams. And yet, the money flowing into the chipmakers and cloud providers is at least partially predicated on a fierce belief that A.G.I. is around the corner. GPT-5 was an indication that maybe it’s not. “None of the math here makes sense unless it’s A.G.I.,” Marcus said. “The whole thing is [based on] the myth that we’re going to get to A.G.I.—which maybe someday happens, but won’t happen soon. The problem is that none of these systems is reliable enough, and so nobody’s actually making that kind of money out of the systems.”

Yes, Marcus has developed a reputation as a sort of industry Cassandra, but what if he’s even half-right? What if our collective A.G.I. fantasies are just that, fantasies? What if L.L.M.s can’t get more than 95 percent reliable? What if those fundamental limitations remain, well, fundamental? L.L.M.s themselves won’t just vanish; people can certainly find some utility for them, especially in lower-stakes SaaS and enterprise applications. But the feverish push for the nationwide adoption of generative A.I., Marcus believes, will eventually be walked back, since it’s “premised on GPT-5 being magic.” Hard to imagine Altman and others getting too excited about that…

That’s all for today. I’ll see you on Thursday.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St, New York, NY 10006