Soumitra Dutta say reliability will be critical for AI's next phase
Updated: Sep 27
Nobody asks if a passenger plane is "intelligent". The only question is whether it is reliable every time it flies․ AI is starting to get assessed against the same standard․
When AI was trapped inside a chat window‚ an impressive demo was a decent stand-in for progress․ The model writing cleaner code‚ translating more accurately, etc. But once AI starts to matter outside the bubble of the chat window‚ reliability will be the thing that matters․
Soumitra Dutta‚ former dean of Oxford University's Saïd Business School and cofounder of AI companies NexiVerify and Caasaa‚ thinks the shift is already happening: "The real test of AI will not be how eloquently it generates language but how reliably it performs in complex environments․"
Language models are mostly evaluated on benchmarks․ Reasoning‚ memory‚ and coding capabilities․ Great‚ if what you're building is a system meant to generate information․ But a warehouse robot‚ an autonomous drone‚ or a surgical assistant is being evaluated for entirely different reasons․ The fact that it reaches the right decision is more important when the conditions on which that decision is based are constantly changing.
A warehouse robot may recognize a package‚ but may fail if the lighting conditions are different or if a human crosses in front of its sensors․ A surgical robot may know the steps of a procedure‚ but it may have to compensate for unforeseen variations in its movement or in the properties of tissues․ The job isn't producing the best possible response anymore․ It's staying safe or effective when the environment won't cooperate․ Soumitra Dutta, says we need to look for better outcomes, We need to take risks, so there will be failures, it’s moving so fast, we need to stay ahead of the changes, we don’t want to be left behind.
This is not a new problem‚ engineering-wise; we test aircraft well beyond normal limits all the time․ Likewise‚ medical devices are tested for years before they go into the clinic․ Bridges are built with safety factors‚ since lab tests may not replicate the real world․ AI systems moving into these spaces will likely be held to that same bar․
To be sure, bigger models and better benchmark numbers are still useful‚ but none of them guarantee that the system will behave reliably out in the wild․ Something that works 999 times out of 1,000 may be the wrong choice if the one failure hits someone‚ however accidentally․
Reliability is a much harder nut to crack than raw capability․ It's not just the algorithm․ It's the sensors‚ the hardware‚ the calibration‚ and whether the thing can recover gracefully when something goes wrong․ None of that shows up on a leaderboard․ But it's usually what decides whether people actually trust the system․
Comments