News / AI series: Be careful, or your AI might hallucinate

The final article in our series on AI in logistics examines how companies can avoid AI hallucinations with ‘non-generative’ solutions

Global supply chains may be experiencing something of an “AI” gold rush, but according to a co-founder of Raft, anyone peddling fool’s gold will be forced “out in the wash”.

Co-founder and CTO of the self-proclaimed “world’s largest logistics AI platform” Raft, Nisarg Mehta, said the surging number of logistics platforms highlighting a purported AI operability was to be expected, “because it is easy”.

Mr Mehta told The Loadstar: “Scale is a really important word here, however, because it’s very easy to make something work with 10, 100, even 5,000 data points.

“But when you get into the territory where the number of data points can be measured in the millions and the billions, that’s when the game kind of changes. The difference between the scale we do it at and those doing a demo or quick little solution is night and day.”

Raft’s model is predicated on its being trained to interpret unstructured and semi-structured data, provided by customers, to reduce time spent on business processes.

As an example, Mr Mehta noted that one customer was using the system to speed up the time it takes to process document packages, with the AI-led system reducing the time spent from “seven, eight, or even nine days down to an hour or two”.

One of the hurdles that has been raised surrounding AI’s use in real-world situations was the increasingly prevalence of “hallucinations”.

An AI hallucination results from the system having been trained on bad data and uncritically reproducing garbage, exemplified by Google’s AI having informed an AP reporter that Buzz Aldrin “deployed cats on the Apollo 11 Mission”.

Raft’s CTO said part of the problem surrounding AI concerned its public perception and the over-visibility of what he described as one of its “two streams”.

Differentiating these streams, Mr Mehta said that there are the generative AI models, “things like ChatGPT and other chatbots” but there is also non-generative AI, which he said could best be understood as “supervised learning models”, with Raft utilising both streams.

“Generative AI can result in what is called ‘hallucinations’, which result from its efforts to create the next sentence without entirely understanding what it is,” he continued.

“The resultant experience is it spewing some sort of nonsense. At Raft we use generative AI, but we mix it with the non-generative form – which does not suffer hallucinations because it is not trying to generate text, it is just trying to classify something or make a prediction.”

Importantly, he added, the key is that the Raft system works under a “human-in-the-loop” model, with the classification/prediction sent to that human for use.

If there is a situation where the AI does not recognise or understand any of the data points, he noted that while it will “try its best”, it is also programmed to inform the human in the loop that it does not understand that piece of data.

He added: “We have developed a track record with our customers, and the people we work with, to show that we have an AI system that can work for the logistics sector, and at scale.

“Our platform was built around AI from day one, we have not just sprinkled it on at the last minute for a bit of marketing dust, it has been front and centre and is a key part of the value-add we offer our customers. In the end, it’ll all come out in the wash.”