The Grok Voice leaderboard is not a kitchen
Grok Voice 3.0 leads a composite speech-to-speech index and trails on τ-Voice. A leaderboard is not a kitchen. Training audio has to contain the room.
Production voice clones are a rights product, not a voice setting
Live voice stacks now ship more accents and clone-like control. Consent to train a recogniser is not consent to clone a speaker. Put TTS in the ask before the clips exist.
Private eval is becoming the product
SWE-bench Pro, private splits and lab audits are filling the hole public boards left. If you ship a product, you need a frozen set that looks like production, not another leaderboard login.
Article 10 is a data spec wearing a regulation
High-risk systems under the AI Act need training, validation and test sets that match the setting the product will run in. That is a brief, not a slogan about quality.
Live agents can see now. That is a faces problem until you say otherwise.
Gemini Live and similar stacks take visual context in the same conversation as speech. If the product can see, you have a capture and rights problem, not only an audio one.
Synthetic-only training is back on the table – and so is collapse
Teams are generating more of their own training mix. 2026 guidance still says keep a real human seed in the loop, or the model trains on its own average.
Humanoid labs are building data factories, not just models
Figure, Neura and others are hiring pilots, teleoperators and data-quality people. The humanoid race has moved from demo videos to who can specify, capture and clear human behaviour at scale.
oice agents went full-duplex. Most training audio did not.
OpenAI’s GPT-Live and Google’s Gemini Live can listen and speak at once. Product audio is interruptions, kitchens and other people talking. A booth script will not teach that.
OpenAI just said the coding benchmark was contaminated
OpenAI has moved off SWE-bench Verified because tasks leaked into training. Anthropic showed BrowseComp answers sitting on the open web. Public leaderboards are not a product eval.