Updated
Updated · Gizmodo · Jul 9
OpenAI's GPT-Live-1 Fails 2 Basic Letter-Counting Tests After Launch
Updated
Updated · Gizmodo · Jul 9

OpenAI's GPT-Live-1 Fails 2 Basic Letter-Counting Tests After Launch

3 articles · Updated · Gizmodo · Jul 9

Summary

  • A day after launch, GPT-Live-1 stumbled on two simple voice prompts, saying “seventeen” contains two Es and failing the standard “strawberry” letter-count test.
  • Those misses undercut OpenAI’s pitch of a smarter, faster, more natural voice model, even though the company’s headline upgrade was full-duplex conversation rather than text-style benchmarking.
  • User feedback also flagged a practical downside of that full-duplex design: the model interjects with acknowledgments like “mhm” and “yeah,” creating an interruptive feel during conversations.
  • The clips spread quickly on X and TikTok, with OpenAI developer Jason Liu reposting one with an expletive, extending a pattern of viral stress tests that have previously reached CEO Sam Altman.

Insights

OpenAI's voice AI aims for natural conversation, so why does its core feature constantly interrupt and frustrate users?
If a 'smarter' AI handles reasoning, why does OpenAI's new voice model fail at tasks a child can solve?
Is the biggest barrier to advancing conversational AI not flawed models, but the lack of high-quality human data to train them?