I Talked to My Computer All Week and My Keyboard Feels Neglected

The Voice Upgrade Nobody Expected

GPT-Live shipped during general availability week and quietly stole the show from the models everyone was benchmarking. The difference from the old voice mode is not subtle:

• It lets me interrupt mid-sentence without losing the thread, which is how humans actually talk
• Latency dropped enough that pauses feel like thinking, not buffering
• It handles tangents gracefully instead of dragging us back to its outline

My test was practical rather than scientific. I did two walking brainstorm sessions for the TACOS backend refactor, talking through the design out loud while it asked clarifying questions. Both walks ended with usable written summaries in my notes. Zero typing involved.

Voice went from demo feature to actual thinking tool the moment interruption started working.

Agents That Finish Things

ChatGPT Work also graduated from preview, and I gave it a real job: compile my quarterly project expenses from receipt emails into a labeled sheet. It gathered across the inbox, built the spreadsheet, and flagged three transactions it was unsure about.

Forty minutes, mostly unattended. Two of forty-plus entries were miscategorized, which I only caught because I checked. That ratio is the whole story of agents right now: ninety-five percent reliable is transformative for chores and completely unacceptable for anything important.

Browsers That Browse For You

Perplexity's Comet browser opened up to Max subscribers, merging search, browsing, and an assistant into one Chromium shell. It answers questions about whatever page is open and chains multi-step tasks without the tab-juggling circus.

A friend screen-shared his setup and had it compare GPU prices across six retailer tabs while we talked. Impressive, occasionally slow, and weirdly relaxing to watch. Great for open-ended research, overkill when you already know exactly what you want.

Honest Limits After Seven Days

• Voice still fumbles long number strings, so dictating card numbers remains a bad idea
• Work-style agents need check-ins every few steps or confidence becomes hallucination
• Browser agents lose to plain old typing whenever the task is precise

The mouse survived the touchscreen era. After this week, I am less sure it survives this one.

My input split for the week ended up roughly forty percent voice, twenty percent delegated agent tasks, and the rest keyboard. That split would have sounded absurd in June.

Cheers, Yassen