
OpenAI’s realtime stack is now the default for voice agents
Latency dropped enough that product teams are ripping out the ‘press to talk’ pattern. The story is less the model — more the session APIs and interruption handling.
Archive
Each edition is a dated JSON file in /content/stories. Newest first.
Models get cheaper. Taste does not.

Latency dropped enough that product teams are ripping out the ‘press to talk’ pattern. The story is less the model — more the session APIs and interruption handling.

Claude controlling a desktop is no longer a conference clip. Teams are wiring it into QA, ops, and research with tight allowlists — and discovering the boring failure modes.

The distribution play is the product: Docs, Gmail, and Meet as the surface. If you already live in Google, switching cost is now social, not technical.

Every closed-model price cut has an open-weight echo. Fine-tunes on commodity GPUs are still how a lot of startups ship without a seven-figure API bill.

Training headlines faded. The fight moved to tokens per dollar, speculative decoding, and who owns the last mile of serving. If your unit economics ignore this, they are fiction.

High-risk labels, transparency, and logging are showing up in RFPs. US startups selling into Europe are hiring compliance the way they once hired growth.

The 40-tab stack is dying in public. Newsletter operators and freelancers want a briefing surface, a draft pass, and an editor — not another copilot badge.