Have we got AI all wrong?
The UK is the hardest working country in Europe - putting in 10 hours a week more than the rest of the continent - but productivity still lags behind Spain, Italy, France, Germany and the Netherlands.
Subscribe to The Fora Institute of WorkAI is at the heart of the Government's strategy to boost GDP, but does AI actually improve productivity? The evidence isn’t clear.
More technology, more problems: “We have more and more technology that we're embracing at quite a rapid rate, and we're not doing anything faster than wedid before,” explains Simon Allford, co-founder of architecture practice Allford HallMonaghan Morris. “We estimate, in my industry, that 50% of the cost of a project is spent on non-productive management of process, and when we build it's taking us probably twice as long. I expect AI to have the same effect.”
He's not wrong. There’s evidence that AI isn’t the solution:
In August, a report from MIT found that 95% of generative AI pilots didn’t increase a company’s profit or productivity.
In July, a study from the AI research group, METR, found developers took 19% longer to complete tasks using generative AI when they’d expected AI to speed them up by 24%.
Last year, research from recruiters, Upwork, found 77% of surveyed staff reporting AI tools decreasing productivity and adding to their workload.
But is he right? There’s evidence that AI is the solution.
In February, the Federal Reserve Bank of St Louis found that generative AIdelivered a 1.1% increase in productivity.
Then a study of Procter & Gamble employees found one person plus AIperformed as well as two people at a task. Meanwhile, Australia’s National Science Agency found that 70% of staff reported productivity benefits.
We’re going to need a different type of AI: We are reaching the limits of Generative AI like ChatGPT, Gemini and Claude, according to the impressively named Floridi Conjecture, developed by Yale’s Professor of Cognitive Science, Luciano Floridi. He argues there is a trade-off between accuracy and size. The bigger the dataset the AI is trained on, the more mistakes it makes.
Does your VC know about: the efficient compute frontier, or the point of diminishing returns where adding more resources to train an AI model yields progressively smaller improvements in performance.
Don’t burst my AI bubble: It already popped.
In August, Meta froze hiring in its Superintelligence Labs after luring more than 50 top researchers with nine‑figure packages. NASA and IBM's Surya foundation model, trained only on nine years of solar observations, predicted solar flares two hours in advance, improving forecast accuracy by 16%. Lab research AI model MIT's FastSolv, focused only on a highly specific database, accelerated research significantly. The future is small language models designed for a specific task, according to AI researcher Christopher Sanchez.
Make me sound clever: “I know people throwing info into an AI and asking it to make them an amazing looking auditorium that will get a client off their case,” says Allford. “AI should be the back of house stuff that doesn't require human intelligence. We should concentrate on where we bring value and create delightful places for people to enjoy themselves.”