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.
Brussels is still rewriting who may train on what
The EU’s copyright consultation is open into November, while GPAI providers already owe a training-content summary. Web-scrape rules are tightening. Faces, voices and robot video were never the same category.
Physical AI is suddenly hungry for first-person human video
Robot labs are pretraining on tens of thousands of hours of first-person human video. The bottleneck in physical AI has moved from the model to whether that capture is specified, consented, and actually the skill you need.
What “human in the loop” is supposed to mean
Human in the loop is a named person, a written task, and a record of the judgement – not a slogan on a slide next to an unsupervised pipeline.
Your eval set is probably leaking into training
If the same people, prompts, or clips sit in training and eval, every graph will look like progress. Held-out human evaluation has to be a different dataset with different rights.
Synthetic data is a tool. It is not a human baseline.
Synthetic data scales a pipeline. It cannot replace consented human capture where diversity, provenance, and ground truth still have to come from people.
Read speech is not how your customers talk
Read speech is cheap and clean. Spontaneous speech is how people talk to products. A speech dataset spec has to choose, on purpose, or your WER will lie.
What is egocentric data in physical AI?
Egocentric data is first-person capture of a person doing a real task – the viewpoint a robot or wearable policy actually needs, not a third-person demo.