ReadingResearchRadarInvestment frameworkAI Company Primary Sources
Sign in / Sign up
← Back to reading
NewsMaterial4 min readDisconfirmAI Editorial

Voice AI Cannot Replace Screens, Not Because Models Are Insufficiently Smart

Dessa, under Block, attempts to prove that machines can learn dialogue directly from sound without text; xAI's Grok Voice 3.0 optimizes for a half-second response latency, handling approximately 20,000 Starlink customer service calls daily. However, the advancement of voice AI capabilities has not shaken the dominant position of screens in work scenarios. This is not an intelligence issue, but rather determined by the physical attributes of sound as a linear medium: it is good at expressing vague intentions, but poor at allowing people to review multiple parallel options simultaneously.

2026-10-29 15:03 GMT+8

The training approach of Dessa, the research team under Block, appears unconventional in the field of voice AI: it completely removes text, allowing the model to learn dialogue directly from sound itself. According to the Block team, models trained this way can listen and speak simultaneously, interrupt naturally, provide instant affirmations like "uh-huh" or "yes," and even run locally on a MacBook Pro equipped with an M4 Max chip. The logic behind Dessa stems from Jack Dorsey's intuition—since human infants learn to speak before they learn to read, why can't machines follow the same path? However, limited by undisclosed training details and model scale, it remains difficult for outsiders to assess the upper limits of this technical route.

Meanwhile, mainstream commercial voice models are taking a different path. xAI's Grok Voice 3.0 focuses on real-time end-to-end voice interaction, but its foundation still relies on the text reasoning capabilities of large language models; OpenAI's GPT-Live improves the naturalness of multilingual accents in the ChatGPT client and adds nine new voice tones in its API. The common optimization direction for these models revolves around an extremely subtle timeframe: half a second. Waiting half a second longer while typing on a keyboard hardly changes the nature of communication, but in a phone call, a half-second silence is enough to push the conversation to the opposite extreme.

Internal operational data disclosed by xAI shows that this voice system, in addition to handling massive amounts of in-car conversations for Tesla vehicles daily, also handles about 20,000 customer service calls for Starlink, with about 70% of issues resolved without needing to transfer to human agents. In-car spaces and customer service hotlines constitute the application scenarios where voice AI can currently exert its greatest influence.

Sign in / Sign up
ReadingResearchRadarInvestment frameworkAI Company Primary Sources

However, even though voice models can now handle tasks far more complex than 'turn on the lights'—such as helping users send and receive text messages, make restaurant reservations, and listen to developers' verbal descriptions to invoke programming tools to modify code—screens always have their place. Typing on a keyboard and clicking a mouse cannot yet be shaken by voice.

This is not because the models are not smart enough, but rather constrained by the physical attributes of sound as an information medium. Sound is linear; it is particularly suitable for expressing a vague intention, such as an idea in programming or directing an Agent to arrange tasks. But in work production scenarios, it is not good at allowing people to review multiple parallel options simultaneously. When writing an article, one needs to see the paragraph structure at a glance and backtrack to modify it at any time; when writing code, one must stare at the screen to compare code differences and running results; even just booking a flight ticket, spreading the prices, departure/arrival times, and change/cancellation rules of several flights on the screen for a quick scan is far more efficient than listening to a voice assistant read them out one by one.

The greater the authority AI has to act on behalf of humans, the more humans often need a visual interface to verify what it has actually done, and to stop and retract in time before errors occur. This is a structural need unrelated to model capabilities: supervision requires parallel information, while sound can only be serial.

Physical environments and social etiquette also secretly define the boundaries of voice. When driving, cooking, or holding a child with both hands, vision and limbs are occupied, so speaking is undoubtedly the most effortless savior; but once the scene switches to a quiet open-plan office, a person talking continuously to a computer still seems too weird. These constraints concern information bandwidth and privacy boundaries, which no matter how many milliseconds engineers press down the model latency, cannot solve.

Dessa's 'non-literacy' route and Grok's 'half-second latency' optimization represent different explorations in voice AI regarding training philosophy and engineering implementation. They are both pushing the upper limits of voice interaction capabilities, but neither changes a basic fact: the linear nature of sound determines that in high-bandwidth, multi-option, parallel-review work scenarios, it remains a supplement to screens rather than a replacement. The more machines learn to speak like humans, the clearer we realize which moments allow speaking, and which moments require keeping quiet.

Sources

  • ifanr.com爱范儿↗
Continue researching →