Subvocal’s AI glasses: Talk to AI without saying a word

Transcript

Auto-generated YouTube captions. They may contain transcription errors.

• 0:01 Okay, this one's wild. A tiny startup
• 0:04 called Subvocal just raised $1.3 million
• 0:08 to build glasses that let you talk to
• 0:10 your AI without making a sound. Let me
• 0:13 show you how it works and why I think
• 0:15 the timing is perfect. Here's my
• 0:18 problem. I spend my whole day telling AI
• 0:21 what to do and I've only got two
• 0:23 options. I can type, which is slow. Most
• 0:26 people manage about 40 words a minute.
• 0:28 or I can talk, which is three to four
• 0:31 times faster. And now that agents are
• 0:33 doing real work for us, writing code,
• 0:36 answering messages, running tasks, the
• 0:38 slowest part of the whole system is me,
• 0:42 which is exactly why voice is blowing
• 0:44 up. Take Whisper Flow. It lets you just
• 0:47 speak instead of type in any app on your
• 0:50 computer or phone. In August, it raised
• 0:53 $280 million at a $2 billion valuation.
• 0:57 It tripled in under 10 months and it's
• 1:00 used at more than 10,000 companies.
• 1:02 Honestly, that's insane growth for a
• 1:05 dictation app. And talking to AI has
• 1:07 gotten weirdly human. In July, OpenAI
• 1:11 gave Chat GPT's voice a new brain called
• 1:14 GPT Live. It listens and talks at the
• 1:17 same time like a real person. You can
• 1:19 cut it off mid-sentence. It even goes
• 1:23 while you're talking. We're not barking
• 1:25 commands at it anymore. We're actually
• 1:27 having conversations.
• 1:29 But there's a catch. You have to say it
• 1:31 all out loud. Picture dictating your
• 1:34 emails on a packed train or explaining
• 1:36 your code to Claude in an open office
• 1:38 while everyone around you listens in.
• 1:41 Yeah, no thanks. Even Whisper is now
• 1:44 working with hardware makers so people
• 1:46 can dictate without speaking loudly. And
• 1:48 that's exactly the gap Subvocal is going
• 1:51 after. It's a San Francisco startup
• 1:53 backed by Y Combinator and its glasses
• 1:55 turn silent mouththing into text. You
• 1:58 move your mouth like you're talking. No
• 2:00 sound comes out and the words just show
• 2:02 up on your screen. So, how does that
• 2:04 even work? When you speak, lots of tiny
• 2:07 muscles in your face move, and they
• 2:09 still move when you only mouth the
• 2:11 words. The glasses send out weak radio
• 2:14 signals, kind of like a bat using
• 2:16 echoes. As those muscles move, the
• 2:18 echoes change. Then an AI model figures
• 2:21 out which patterns mean which words and
• 2:23 types them for you. They claim more than
• 2:26 200 words a minute. That's faster than
• 2:28 most of us talk. Quick myth busting.
• 2:31 There's no camera, so it's not reading
• 2:33 your lips. It can't read your thoughts
• 2:35 either, only the words you deliberately
• 2:37 mouth, and it's not always listening.
• 2:40 You decide when it's on. In their demo,
• 2:42 the founders silently mouth summarize
• 2:45 the unread messages in Slack. And boom,
• 2:48 the text appears. The plan is for it to
• 2:51 work anywhere you type. Chat, GPT,
• 2:54 Claude, Cursor, Slack, email. Imagine
• 2:57 telling your coding agent what to build
• 2:59 while the co-worker sitting right next
• 3:01 to you hears absolutely nothing. Okay,
• 3:05 reality check. These aren't shipping
• 3:07 yet. First deliveries are expected in
• 3:10 early 2027. That 200 words number comes
• 3:13 from the company, and nobody has tested
• 3:16 it independently. There are no public
• 3:18 accuracy numbers yet. The early bird
• 3:20 price is $499 with a refundable deposit
• 3:24 and it's US only for now. But here's why
• 3:27 I can't stop thinking about it.
• 3:30 Keyboards shaped how we used computers.
• 3:32 Voice tools like Whisper Flow and Chat
• 3:34 GBT are taking over now. And as we hand
• 3:38 more and more work to AI agents, the
• 3:40 next big race is over input. The
• 3:43 fastest, most private way to tell them
• 3:45 what you want. subvocal is betting that
• 3:48 way is silent. And if they're right, the
• 3:51 next big interface might be one nobody
• 3:53 else can hear.