Speech doesn't have to be the response a voice model gives back. @charlierguo, a Developer Experience Engineer at OpenAI, builds "Voice Agents Can Just Do Things" around that idea, and @aiDotEngineer has it posted on YouTube. The talk lays out a framework for deciding when a voice agent should actually talk, act, or just wait.
- Three modes, not one. He frames voice design as speech-to-speech, speech-to-action (the model uses tools), and event-to-speech (the model reacts to something and speaks), and argues the best products mix all three rather than picking one.
- Speech-to-action is the underused one. Form filling, creative tools, and computer use (he points to the ghost-cursor animation that clicks around the screen for you) are all cases of voice acting on your behalf instead of just replying.
- Event-to-speech is still exploratory. Hands-free moments, like cooking, and proactive notifications that escalate from a visual cue up to actual speech, are the two patterns he points to here.
- Native audio over the chained pipeline. OpenAI's Realtime model family skips transcription and works on audio tokens directly, since, as he notes, only a small fraction of face-to-face communication is carried in the words themselves.
- GPT Realtime 2. The newest Realtime model reasons before it speaks, can delegate parallel tool calls, and uses preambles so it can tell you it's about to go do something, the way a travel agent might say they need a couple seconds to check flight prices.
- Accessibility as the throughline. He points to developers who lost hand mobility and dexterity, and who now generate far more code with the help of voice and coding agents.
I'm working through the published talks from AI Engineer World's Fair sharing summaries and takeaways. Follow for more!
The stock market is punishing strong economic data.
A blowout jobs report is normally good news, yet several strong reports have sent stocks lower within hours instead.
In the analysis below, we break down six S&P selloffs after strong US payrolls since 2015, along with three times the market rallied instead.
The trigger isn't inflation alone, it's how the data shifts expectations for the Fed.
AI is expanding cybersecurity beyond enterprise networks into agent guardrails and counter-drone software.
We interviewed CyCraft (7823 TT), a security vendor with 300+ customers, 90%+ of 2025 revenue from subscriptions and an 86.3% gross margin in 1Q26. https://t.co/Mx1Q6ypYLQ
If a model can one-shot almost anything, the scarce thing stops being implementation and becomes having an idea worth building. @eve_bouff, Head of Design at Y Combinator, builds an entire session out of that premise in "Imagination Engineering: Live in the future and then build what's missing," on @aiDotEngineer's YouTube. Most of it is a live tour of things she made for herself inside a week, which is the useful part: a concrete picture of what software on demand looks like when you stop rationing it.
- Thinking in public, not just building in public. She opened a Slack channel for her raw stream of consciousness, inspired by a former colleague's "Lewis Thoughts" channel at Conductor, and gave a few trusted people access.
- A week of brain dumps compiled into a personal site. She pointed Opus 4.8 at the channel and asked for a website from it. It pulled out the books that changed her life, the quotes she keeps, the projects she's shipping at YC, the tools she uses daily.
- voice.md. She had a model read everything of hers it could find published online, tweets included, and distill her tone into a file it can then write from.
- The design language maintains itself. A design system doc for the site updates automatically whenever she adds a new component.
- Shape of Minds. Built that same morning starting at 10am: hover over history's most prolific minds and compare them across thinking, obsessions, quirks, origins, routine, worldview, habits. Then she asked for the commonalities. A lot of them argued constantly, were childless, took naps, and barely ate.
- A glossary file for the agents. The same information the page displays, written as markdown, because handing an agent the file is cheaper than making it parse the rendered page.
- Tools built for an audience of one. A Slack emoji search because the native picker never surfaced what she wanted. A countdown timer to hold herself to a self-imposed 10am deadline. Short reports on things like Renaissance patronage and the ten principles of good design, sized to read on the way to work.
- Design is heading toward sliders. Her site lets you swap art, apply shaders from Paper, translate the whole thing, and retune font, typeface and column width. Her read on where this goes: a lot of future design work is parameter adjustment.
I'm working through the published talks from AI Engineer World's Fair sharing summaries and takeaways. Follow for more!