Last year, chatbots including ChatGPT took the internet by storm with their ability to answer questions posed by humans with lifelike quality.
The breakthroughs were the result of years-long research in the field of generative artificial intelligence — where software creates content like texts or images based on descriptions — and came due to advances in math, computing power and new ways to train software.
This year, multiple AI experts said, people will probably see more of these public-facing products come out. AI companies could also move on from mimicking human language through text into speech, trying to build bots that could be marketed as smarter helplines or virtual assistants, they said.
Altman said part of this could be because the labs that create these products, such as the Elon Musk-backed OpenAI, can benefit from generating buzz so that large corporations try to license their technology and create in-house AI products for their own customers. “That’s their business plan,” Altman said of labs such as OpenAI.
Still, any attempts to release products widely will run headlong into issues — namely that chatbots are still wrong, racist and sexist at times — and require new training methods.
But it will take years to go from how they are currently trained, which is by ingesting large troves of text and using patterns to predict what word comes next, into a way he says is better: Instead of simply predicting what word comes next based on probability and patterns, teach bots to discern if those words are true based on data sets that are higher quality and from trusted sources. That would help avoid racism and sexism and be more accurate, he said.
“You have to teach it how to figure out truth and untruth,” Altman said. “And let me just say, that is profound because we know there are humans who don’t do this very well.”
