The successful implementation of AI for business purposes is reliant on how a particular model or algorithm is trained. As with any other productivity tool, incorporating it into your business without training or a clear view of its uses and limitations can be counterproductive.
An interesting example of this in action is found in a unique experiment undertaken by Robert Caruso which involved the popular general-purpose AI tool ChatGPT 4o, a console chess game from 1979 and a very unexpected lesson.
Mr Caruso, when discussing chess computers and their influence on AI development, noticed that the chatbot claimed to be a strong chess-playing computer that could easily defeat Video Chess, a chess video game running on the Atari 2600.
The Atari 2600 (at the time known as the VCS) could not ordinarily even display the pieces of a conventional chess board and was believed to only be able to predict a move or two ahead.
At the insistence of the ChatGPT instance according to Mr Caruso, he set up a game using an emulator and was surprised at how easily the significantly more advanced AI system was defeated by a system that was considered primitive even in the late 1970s.
It initially struggled to recognise the abstract piece designs necessitated by the Atari hardware, but even once that was changed to standard notation, the newer system struggled to maintain board awareness, confused pieces, attempted to use captured pieces and sacrificed pieces for no gain.
It eventually conceded, but provided an important lesson not only when it comes to board games but also when it comes to specialisation and training.
It revealed limitations in some currently popular tools that necessitate additional training, validation and limitations in their use to tasks that it is provably efficient and effective at undertaking.
Much like how a business should seek proof and evidence of an employee’s capabilities before giving them a task, the same is true of an AI system, as proven through a rather fascinating challenge that went badly wrong.
