Generative AI: hype rather than hyperintelligence is lifting values

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

Generative artificial intelligence could produce magic in the sky, says Sam Altman founder of pioneering start-up OpenAI. Successful entrepreneurs are entitled to make visionary statements. But investors should keep their heads. Expectations for GAI are running way ahead of the inherent limitations that currently apply to it.

As investment in GAI grows, so does pressure to create new use cases. By 2027, data group IDC predicts that enterprise spending on GAI will reach $143bn, up from some $16bn this year.

AI hype has already helped to lift the tech-centric Nasdaq Composite index 36 per cent this year. Some of that increase rests on an overestimation of its capabilities.

OpenAI is itself hoping for more funding to pursue its goal of reaching human-like levels of artificial intelligence. It is worth remembering that when examining Altman’s plan to build “superintelligence”. This is generally defined as a capacity for thought exceeding that of humans.

Models predict, they do not comprehend. That limitation casts doubt on the possibility of AI achieving even human-like general intelligence.

Text generation produced by large language models depends for the moment on the data used to train those models. LLMs produce better results when they reflect recurring concepts. They struggle with new scenarios and tasks outside that envelope.

This helps to explain why Google DeepMind’s AI weather forecast model recently bested existing forecasting models. Most of the time, weather patterns are recurrent. Notably, DeepMind could not outperform prior models when attempting to spot unusual, extreme events.

LLMs meanwhile struggle to identify their own mistakes. Requesting a correction does not produce a more accurate answer. In a study of LLMs, Originality. AI found that every single one produced errors. OpenAI’s ChatGPT-4 offered inaccuracies in nearly a third of responses.  

Chief financial officers have more prosaic goals in mind as they hunt for ways to deploy its tools. These range from parsing employee performance reviews to scheduling waste collection times. 

So far, results have been patchy. A study of AI chatbot assistance in the workplace, conducted by the National Bureau of Economic Research, showed an encouraging 14 per cent improvement in productivity. But for customer support agents who took part, gains were limited to new and low-skilled workers. Those with experience showed little to no improvement. 

These limitations will become more obvious as generative AI tools roll out in 2024. That will put pressure on providers to address the final unanswered question: cost.

AI could add more than $4tn to corporate profits, according to McKinsey, a consultancy with a penchant for eye-catching predictions. But pricing clarity is lacking. Without it, companies cannot predict what financial gains AI can accomplish.

AI cannot predict that either, however accurately it forecasts the weather.

The Lex team is interested in hearing more from readers. Please tell us if you think GAI can produce superintelligence in the comments section below.