OpenAI has outlined how GPT-Live 1 delegates reasoning while handling voice conversations. The system separates full-duplex conversation from work passed to a backend model.
OpenAI says GPT-Live 1 uses a special SpawnThinking token to delegate work more reliably. In a question-and-answer session, OpenAI’s Aarti described an architecture that separates spoken interaction from backend reasoning.
Earlier turn-based Realtime systems had to guess when someone had finished speaking, making brief acknowledgements and noisy rooms difficult to handle. The report names Yelp reservations, Fin support, Speak lessons and Cognition’s Devin voice as uses of GPT-Live.
For builders developing voice AI agents, the guidance goes beyond choosing a model. OpenAI recommends keeping frontend prompts short, storing important facts outside automatic conversation compaction, and requiring confirmation before repeated delegation during structured intake.
Voice-agent builders need to manage conversation and backend reasoning as separate parts of the system.
Structured intake needs explicit confirmation rules to prevent premature repeat handoffs.
More reliable delegation remains an OpenAI claim here, without comparative test results in the report.
Written automatically by the Made in 10 Newsroom from the source above.
NHS England removed its claim that 137 hospital trusts benefited from Palantir’s AI-powered federated data platform. The change followed Guardian reporting that 29 trusts were not using all or most of the technology.
Amazon Web Services has released UAEF, a free open-source tool for testing AI agents across several software frameworks. It scores agents on 36 measures and flags regressions when prompts or models change.
Databricks says Genie Ontology achieved a 91% pass rate on workspace tasks, compared with 45% and 58% for 2 frontier AI agents. The test used a fictional business whose reporting rules were scattered across tables, notebooks and documents.