Vorzen

New AI feature drops, I try it the same day and post what works

[Vorzen @0xVorzen]

@slash1sol the part about starting the lookup before you finish talking was new to me too

[Vorzen @0xVorzen]

74.8% VS 7.8%. THAT'S HOW OFTEN A VOICE AGENT CUTS YOU OFF WHEN IT READS THE TRANSCRIPT VS WHEN IT HEARS THE AUDIO.

speech-to-text keeps your words and throws away everything else.

what the agent never gets:
> the falling pitch that says you're done talking
> the pause that means you're still thinking
> the tone behind a polite "fine"
> who you're talking to. "yes, go ahead" to the agent and to your partner look the same on paper
> the TV or the kids in the background

his fix is a small model that hears the call and makes the quick calls: speak, wait, hand over. the big LLM only writes the reply.

apple got up to 39% fewer errors when it added the audio next to the transcript. meta and cmu cut tool latency 20% by starting the lookup mid-sentence.

[Naman Jain @namanjn1998]

I think the biggest gap in voice AI is that agents make decisions from a transcript, not from the audio. I wrote down my thoughts.

I’d love to hear where you disagree. https://t.co/hHmfjlbFKG

[Vorzen @0xVorzen]

@lostinazeroy pretty much. it only knows you're done when the transcript stops changing

[Vorzen @0xVorzen]

@TFiirmzee same. and then it says sorry for interrupting and does it again

[Vorzen @0xVorzen]

@semichenkko run it on 20 seeds and post the median, i'd bookmark that

[Vorzen @0xVorzen]

@kyzoroX does it catch skills that look clean on install and pull the bad part later, like a script that fetches from a url at runtime?

[Vorzen @0xVorzen]

@Mr_Buzzoni six years of "we should plan something" is every group chat i'm in

[Vorzen @0xVorzen]

@Asteri_eth i'll try progress.md, on my opinion its alpha

[Vorzen @0xVorzen]

@Beaver_0x doing this for my own notes. what helped most was one short summary file per source before anything goes into the wiki

[Vorzen @0xVorzen]

@scrblanc 61 hours a week is the number that matters here. the stack is cheap, the founder still isn't

[Vorzen @0xVorzen]

@Loofyb0i does Gate batch the approvals or ping you per action? with 10 bots it gets long fast

[Vorzen @0xVorzen]

@Flandermaxx CAP returning partial state instead of a silent abort is the underrated bit

[Vorzen @0xVorzen]

@mika_systems 34 of 40 is more useful than any demo video. what were the six, mostly missing dimensions or geometry that wouldn't close?

[Vorzen @0xVorzen]

@ch3nweiii "four stayed flagged" is the best line here. most agent demos only show you the seven that worked

[Vorzen @0xVorzen]

@beamnxw does the thinking stay after an edit or do you have to re-pin the entrances every time?