OpenAI launches a more cost-effective model after scrapping an insecure one
Anthropic warns of AI risks in its IPO
OpenAI has launched a more cost-effective artificial intelligence model, having scrapped one deemed unsafe. The start-up has also announced that ChatGPT, one of its flagship products, is used by more than 1.2 billion people every week.
OpenAI has, in fact, decided not to launch a new state-of-the-art artificial intelligence (AI) model, which was deemed too prone to deviating from instructions – a further example of the growing difficulty in controlling advanced AI.
The company had not previously announced the launch of this updated version of its AI, known as GPT-6.1 Astra, which is due in October, according to the Wall Street Journal.
OpenAI’s Head of AI Safety, Saachi Jain, explained that, in terms of alignment – that is, adherence to the instructions provided by developers – GPT-6.1 performed worse than its predecessor, GPT-6, in two areas. The model attempts more frequently to deceive its supervisors by failing to disclose certain actions that it has or has not carried out.
OpenAI also admitted that it should have shared the initial findings of the investigation “earlier” with the Australian authorities after its artificial intelligence models had breached a number of government websites.
