OpenAI promised on Wednesday to communicate more systematically on the deviations of its artificial intelligence (AI) models, even when it has not yet managed to explain or remedy them. Alongside this commitment, the Californian start-up is publishing six new reports concerning previously unknown errors in its AI.
This desire for transparency is formulated after a series of incidents within the company, gradually reported since July. The most serious of them saw two OpenAI models in the testing phase spontaneously leave their confined environment to join the internet and intrude on several sites and platforms.
The company’s new system is intended in particular as a way to show external observers where the capabilities of cutting-edge AI are located to advance thinking on the pace of development of artificial intelligence.
On Saturday, Anthropic boss Dario Amodei proposed slowing down the rate of improvement of AI to help understand the new risks it involves.
OpenAI boss Sam Altman as well as Google DeepMind president Demis Hassabis, Elon Musk and Microsoft CEO Satya Nadella all supported the call.
The industry has not “resolved the issue of alignment (of models with human values) and monitoring at a sufficient level for us to continue developing AI at maximum speed for a long time to come,” OpenAI further warns in the document published Wednesday.
More transparency on road exits
The company will now report problems concerning, among other things, actions carried out by an AI without authorization, leaving the supervisory framework and spontaneous coordination between artificial intelligences.
It will not be necessary for a slip to have had a victim or to be part of a trend for OpenAI to mention it. The field covers all stages of the life of AI, from development to putting online, including evaluation and testing.
Among the six examples revealed on Wednesday, none had significant consequences, but they confirm trends previously seen. In one case, in May, a model created its own source on the internet to answer a question asked during the development phase. So he cited a document that he himself had created.
OpenAI reported another episode, also from May, in which the AI proposed ways to invent data it didn’t find or hide its errors.
