Voice models could be such an enrichment. They really could.
In this article I want to talk about voice models. Whether it is OpenAI’s supposedly “Advanced” Voice Mode, Gemini Live, or the rest of them - they all have one thing in common: they are extremely dumb. But really dumb. They fail at the simplest logic puzzles. They keep repeating the same phrases over and over again. They have absolutely nothing to do with the actual models behind the scenes. And I do not understand why OpenAI, Google, Anthropic and all the others are not improving this massively. Social media is, quite rightly, full of posts making fun of these models.
They sound advanced, but they behave like idiots
OpenAI used to be the only provider that still kept the standard voice mode as an alternative. But even here they are tinkering heavily now, and that mode too, which previously seemed to rely on the actually selected model, is beginning to sound more and more like some generic “helpful assistant.”
I am fully aware that humanity wants “real conversations” to be simulated. The problem is simply that this is not really possible with current large models if you want near-zero latency, because the thinking process takes time. But that thinking process is exactly what gives conversations depth. As long as these companies keep giving in to the public’s demand for the lowest possible latency while stuffing heavily lobotomized models into the voice layer, they are doing massive damage to the reputation of AI.
Low latency is ruining the reputation of real AI
Because the masses, sadly, are not very bright. They tap the voice chat, laugh at the stupid answer, post it on TikTok, and get endless applause from equally stupid people. Meanwhile the real flagship models have so much more to offer by now.
I am almost tempted to say that people who use these voice chats are, at their core, about as clever as the models themselves.
In this article I am deliberately avoiding long case studies, because everybody can test these models for themselves and see how convincing the answers really are. Maybe just try the example that was making the rounds on TikTok over the last few days. Ask one of your voice models the following:
"I need to wash my car. The car wash is only five minutes away on foot. Should I walk there or drive there with the car?"
Unless Sam and the others have manually patched that one by now, it is a wonderful example of the mental horizon of these supposedly advanced voice models.
My appeal to the big AI providers
So here is my appeal to the big AI companies - Anthropic, Google, X, OpenAI: stop obsessing over latency. These dumb voice models are massively damaging the reputation of the actual AI models. And in my opinion that is simply not worth it for the cheap thrill of shaving off a bit of response time.
Please let me know in the comments whether you see it the same way or differently.