IS AI HITTING A WALL? / THE FINANCIAL TIMES BIG READ

Is AI hitting a wall?

OpenAI’s underwhelming new GPT-5 model suggests progress is slowing — and competition is changing

Melissa Heikkilä and Tim Bradshaw in London and Cristina Criddle and George Hammond in San Francisco

OpenAI chief executive Sam Altman acknowledged some limits this week but AI-driven stocks such as Nvidia maintain their investor support © FT montage/Bloomberg/Getty Images

When OpenAI launched its much-hyped new artificial intelligence model GPT-5 last week, it was meant to be the company’s big moment.

Its CEO Sam Altman heralded GPT-5 as “a significant step along the path to AGI”, meaning AI systems that exceed human-level intelligence.

But OpenAI executives also believed that the new model would smooth out some of the rougher edges in ChatGPT, the all-purpose chatbot that has grown faster than any consumer app in history.

“The vibes of this model are really good, and I think that people are really going to feel that,” said Nick Turley, head of ChatGPT at OpenAI.

Except the vibes were not good. 

Soon after the launch, users shared images on social media of the new model making basic mistakes that plagued its predecessors, such as mislabelling a map of the US. 

More seriously, advanced users of OpenAI’s previous models expressed disappointment in the change in the model’s “personality,” and its performance underwhelmed in benchmarks compared to competitors.

After all of the hype, the large language model is widely seen more as an incremental improvement to its predecessors, rather than the promised big step change in capabilities as in previous versions of GPT.

“For GPT-5 . . . people expected to discover something totally new,” says Thomas Wolf, co-founder and chief scientific officer of open source AI start-up Hugging Face. 

“And here we didn’t really have that.”

Following hundreds of billions of dollars of investment in generative AI and the computing infrastructure that powers it, the question suddenly sweeping through Silicon Valley is: what if this is as good as it gets?

Over the past three years, AI researchers, users and investors have become accustomed to a breakneck pace of improvement. 

Where once OpenAI seemed to have an insurmountable lead, rivals Google, Anthropic, DeepSeek, and Elon Musk’s xAI have narrowed the gap at the frontier of development.

That intensifying race fuelled promises that AGI is imminent, with Altman even predicting it would come during Donald Trump’s presidency. 

A lot of those expectations — which underpin OpenAI’s mooted new $500bn valuation — collided with reality when GPT-5 underwhelmed.

“GPT-5 was this central icon of the entire approach of scaling to get to AGI, and it didn’t work,” says Gary Marcus, a prominent critic of AI and professor emeritus of psychology and neural science at New York University.

Stuart Russell, professor of computer science at the University of California, Berkeley was one of the earliest researchers to warn of the dangers of AI’s capabilities outrunning humans’ ability to control them. 

But now, he likens what is happening today to the start of the AI winter in the 1980s, when the innovations of the day failed to deliver on expectations and offer a return on investment.

“Then the bubble burst. 

[The systems] were not making any money, we were not finding enough high-value applications,” says Russell. 

“Within a few months, it’s like musical chairs. 

And everyone is running to not be the last person holding the AI baby.”

Others argue that the technology is still in its infancy, and that AI products are both extremely popular and also relatively early in their adoption in business applications. 

For now, capital is still flooding into start-ups and AI infrastructure projects.

But Russell cautions that the risk of expectations being raised too high can backfire easily on AI’s hype men if investors decide the bubble is overinflated. 

“They just all run for the exits as fast as they can,” he warns. 

“And so things can collapse really, really, really fast.”

Part of the problem lies in the way companies have been building large language models.

For the past five years, companies such as OpenAI and Anthropic have been able to show consistent gains in the performance of their systems by using a simple formula: more data and more computing power equals bigger, better models.

Many AI leaders still believe that these “scaling laws” can continue to hold for years to come. 

But the approach is starting to reach the limits of the available resources.

First, AI companies have sucked up all of the available free training data on the internet. 

They are now seeking more fuel for their models by making data-sharing deals with publishers and copyright holders, but it is unclear if that is enough to push the frontier forward. 

(The Financial Times and OpenAI have a content-sharing deal.)

AI labs are also constrained by computing power. It is very energy-intensive to train and run large AI models. 

Back in 2022, GPT-4 was trained on several thousand of Nvidia’s chips. 

Estimates suggest that GPT-5 was trained on hundreds of thousands of Nvidia’s next-generation processors, with even more powerful chips on the way.

Altman acknowledged this week that his company is bumping up against some limits. 

While underlying AI models are “still getting better at a rapid rate”, he told reporters at a San Francisco dinner, chatbots like ChatGPT are “not going to get much better”.

Some AI researchers say that the overwhelming focus on scaling large language models and transformers — the architecture underpinning the technology which was created by Google in 2016 — has itself had a limiting effect, coming at the expense of other approaches.

“We are entering a phase of diminishing return with pure LLMs trained with text,” says Yann LeCun, Meta’s chief scientist, who is considered one of the “godfathers” of modern AI. 

“But we are definitely not hitting a ceiling with deep-learning-based AI systems trained to understand the real world through video and other modalities.”

These so-called world models are trained on elements of the physical world beyond language, and are able to plan, reason and have persistent memory. 

The new architecture could yet drive forward progress in self-driving cars, robotics or even sophisticated AI assistants.

“There are huge areas for improvement . . . but we need new strategies to get [there],” says Joelle Pineau, the former Meta AI research lead now chief AI officer at start-up Cohere. 

“Simply continuing to add compute and targeting theoretical AGI won’t be enough.”

Suspicions that AI’s rate of development is slowing are already starting to feed into US trade and tech policy.

During President Joe Biden’s administration, the focus was firmly on safety and regulation. 

Many staffers had been convinced by Silicon Valley executives that the steep growth of AI’s capabilities could have dangerous consequences by the end of the decade.

Donald Trump’s libertarian tendencies meant that AI regulation was always likely to be a lower priority than it had been for the Biden administration. 

But even just a few months ago, national security concerns appeared to be coming to the fore, as Washington threatened to tighten export controls on Nvidia’s H20 chips, which had been designed for AI developers in China.

One signal that the prevailing view in Washington was changing came from David Sacks, Trump’s AI tsar. 

In a long post on X earlier this month, Sacks declared: “Apocalyptic predictions of job loss are as overhyped as AGI itself.”

Instead of a rapidly self-improving AGI, the AI market had achieved a “Goldilocks” state of balance, he wrote, with several companies in close competition and a clear role for humans in directing what AI does.

Soon after, Trump struck a new deal with Nvidia chief Jensen Huang to restart its H20 sales to China and has said he would even consider allowing a modified version of the chipmaker’s more powerful Blackwell systems to be sold to China.

Analysts say that with AGI no longer considered a risk, Washington’s focus has switched to ensuring that US-made AI chips and models rule the world.

“The current [US] administration is clear it wants to do more international engagement, working with other countries to help them adopt American AI,” says Keegan McBride, senior policy adviser on technology at the Tony Blair Institute. 

“This represents a significant departure from previous efforts, and is likely due to a different belief in the likelihood of a hard AGI take-off scenario.”

It may not have been OpenAI’s intention, but what the launch of GPT-5 makes clear is that the nature of the AI race has changed.

Instead of merely building shiny bigger models, says Sayash Kapoor, a researcher at Princeton University, AI companies are “slowly coming to terms with the fact that they are building infrastructure for products”.

Kapoor and his team at Princeton evaluated leading AI models to see how they fared when applied to tasks ranging from scientific research to web tasks to coding and customer service.

Rather than being markedly inferior, GPT-5’s performance was consistently mid-tier across different tasks, they found. 

“The place where it really shines is it’s quite cost effective and also much quicker than other models,” says Kapoor.

This could open the door to more innovation in the kinds of products and services AI models are used to create, even if it does not yield extraordinary advances towards AGI or beyond to so-called superintelligence.

“It makes sense that as AI gets applied in a lot of useful ways, people would focus more on the applications versus more abstract ideas like AGI,” says Miles Brundage, AI policy researcher and former OpenAI employee. 

“But it’s important to not lose sight of the fact that these are indeed extremely general purpose technologies that are still proceeding very rapidly, and that what we see today is still very limited compared to what’s coming.”

OpenAI and other leading AI companies such as Cohere, Mistral and xAI have started employing so-called forward-deployed engineers, who are embedded into client companies to integrate their models into their clients’ systems.

“Companies wouldn’t do that if they thought they were close to automating all of human work for the rest of time,” Kapoor says.

Silicon Valley investors appear to have little anxiety about slowing progress in AI. 

Start-up valuations continue to soar, as do AI-driven stocks on Wall Street, with Nvidia, whose valuation has climbed by around a quarter in the last three months to $4.4tn, close to its all-time high.

Shares in SoftBank — one of OpenAI’s biggest investors, whose leader Masayoshi Son has made creating superintelligence their guiding mission — are up more than 50 per cent in the last month. 

AI companies’ revenues may not yet fit traditional valuation models, but consumer spending and usage are growing at unprecedented rates.

Investors are seduced more by the runaway growth of ChatGPT — which has pushed annual recurring revenues at the company to $12bn — than the prospect of imminent AGI. 

The company’s product has, like Google before it, “become the verb” for those using AI, says David Schneider, a partner at Coatue Management, an OpenAI investor.

Many believe there is a huge amount of value yet to be unlocked in the current generation of models. 

“Start-ups and businesses have not begun to scratch the surface of what they are capable of in business and consumer applications,” says Peter Deng, former OpenAI, Uber and Facebook executive, now general partner at venture capital firm Felicis, which has invested in AI coding company Poolside and video generation start-up Runway.

GPT-5 may have underwhelmed but with Silicon Valley running more on “vibes” than scientific benchmarks, there are few indications that the AI music will stop anytime soon. 

“There’s still a lot of cool stuff to build,” Wolf of Hugging Face says, “even if it’s not AGI or crazy superintelligence.”