As more firms start using GenAI, it’s thrown up some questions around accuracy and power use. It’s undoubtedly useful, and small business owners say it has made their lives easier and extended their abilities. It’s saving them time, and helps with things like summarisation and drawing insights from text. It’s also making things easier for enterprises, and some are using it to write code and are apparently training their own models. 

With all this excitement, AI stocks have continued their ascent into the stratosphere, propelled on the hype that it’s going to transform our lives. It’s mind-bending technology, but the valuation of the hyperscalers, chipmakers NVIDIA and Arm, and AI companies OpenAI and Anthropic, among many others, are based on the promise of benefits to come, which seem to be at least half a decade away. Besides, this is technology that still needs adult supervision. It’s flaky at the moment, and as everyone knows, GenAI models can spit out unpredictable or biased responses, or responses that are just plain wrong. One reason why this happens is that the model may have been left unsupervised in its training, or the supervision was inadequate, or it was trained on noisy data. And any amount of testing doesn’t seem to help. As Avivah Litan, Gartner distinguished research analyst, told SIliconANGLE, she’s heard horror stories from clients who had spent a lot of time and money rooting errors out of their models. “They test and test, then put it into production, and six weeks later, it starts hallucinating.”