Meta will release open-source Llama 3 and OpenAI its proprietary GPT-5 shortly. Executives at the two tech firms say they’re working on ways of making models retain more information as they go about tasks and logically tackle sequences of actions. (Financial Times)

Talking point: The current generation of large language models do account for the context of the words they’re assessing or producing—that’s the “transformer” innovation for which the “T” in “ChatGPT” stands. But they can’t consider too many things at once, or retain them for too long. Meta is working to get its models to reason, plan and have memory, AI research vice-president Joëlle Pineau, who leads its work from Montreal, told the Financial Times. Both the social platform giant and OpenAI are trying to build artificial general intelligence, a self-teaching system that can figure out and do things on its own. Reasoning is a step on the path.