cognit

vibe coding & sharing my learnings

[cognit @0x_cognit]

@UP_SHIFTED Wait until they pack it in a stick and add a hologram for visual. Bladerunner is set in 2048

[cognit @0x_cognit]

The team that cloned Joe Rogan’s voice in 2019 just built what Block calls the most human AI voice ever made.

Their fake Rogan was so convincing they refused to release it. Then Jack Dorsey’s company bought them, and six years later there’s Dessa.

Dessa learned to talk only by listening to people, without reading a single word of text.
It laughs mid-sentence, hears when your voice cracks, and knows when to interrupt you.
It runs on a laptop, with no cloud and no one listening in.

For now, only safety researchers can get near it.

When AI voices sound exactly like people, how will you know who’s really calling you?

[Alexander Hoff ❍ @axhoff]

please listen to dessa.

i have never heard anything artificial that sounded so… human.

if you ever doubted that human-ai-relationships were going to be a thing. well. this snippet will convince you otherwise.

[cognit @0x_cognit]

@0xVeyyro An important note here is that it can only run on m4 max and better, so still not super accessible

[cognit @0x_cognit]

@Klonzu Splitting the agent that reads from the agent that builds is the part worth copying. Great breakdown.

[cognit @0x_cognit]

@DarkFactorr The Community Note detail on the Adobe claim was a good

[cognit @0x_cognit]

@HashItOut_ Clear, specific, and honest about the catch. Bookmarking this for my 30 series card.

[cognit @0x_cognit]

@HashItOut_ I usually shoot on dji osmo pocket 4p and never before added iPhone footage. Good point on sharpness cutoff, I will explore that

[cognit @0x_cognit]

This is a f*cking game changer. Here is exactly how I save $3000 to $10000 a month but editing videos with Claude.

The core idea: Claude looks at frames, so you turn the footage into something it can judge, then let it make the choices.

1. Break the footage down
A script (ffmpeg + scene detection) splits every clip into shots and pulls 3 to 5 frames from each. It also measures motion, sharpness and brightness, so shaky, blurry or blown-out clips get flagged before Claude spends any effort on them. You skip the first pass of scrubbing through hours of junk.
2. Build contact sheets
Frames get stitched into one grid per clip. Claude reviews them the way an editor reviews a proxy: composition, light, subject, depth, variety.
3. Tag every shot
Establishing, detail, movement, portrait, transition. This stops the edit from turning into ten similar skyline shots in a row.
4. Approve a shortlist
Claude proposes the best moments and gives a reason for each. You keep, swap or reject by voice or text. “Best” is a matter of taste, so this step is what makes the edit yours.
5. Add music and cut to the beat
You pick the track, since Claude can’t hear the mood. Beat detection (librosa) finds the rhythm, and Claude places cuts on it. It can follow rules like “1.5 to 3 second shots, slower in the middle.”
7. Assemble
Claude builds the sequence with ffmpeg, or exports a timeline you can open in DaVinci and fine-tune by hand. Then grading, speed ramps, titles.
8. Save it as a skill
When an edit feels right, tell Claude to save the approach: your pacing, shot-type mix, grade, title style. The next vlog starts from your taste instead of from zero.

Where Opus 5.5 helps most: judging frames, tagging shots, and building motion graphics like titles, overlays, map lines and animated numbers. The creator found its animations much better than the other model he tried.
Expect 2 to 3 rounds of corrections. It’s still a big shift for volume content.

Post 20 shorts a month and a freelancer costs $3,000 to $10,000. One subscription costs $100 to $200.
Claude handles the sorting, assembly and graphics. You make the taste calls.

I am going to share some of the best skills I found in the comments.

[Alex @The_Alex]

Goodbye Davinci, goodbye Premiere. We don't need you anymore... You should be editing entirely in Codex or Claude Code.

Here's exactly how to do it.

[cognit @0x_cognit]

@HashItOut_ glad to help

[cognit @0x_cognit]

@alxqwrk Great breakdown. The “one clear message beats five small follow-ups” tip alone is worth a bookmark.

[cognit @0x_cognit]

@squishlab_ai 49/50 on inputs you picked yourself is promising, but the real test is whether it holds on edge cases nobody listed. Gotta try it

[cognit @0x_cognit]

@vylrex I always click approve anyway. Saves time

[cognit @0x_cognit]

@The_Alex I am editing my next travel vlog with Claude

[cognit @0x_cognit]

@r1VeN2k Seems like free money

[cognit @0x_cognit]

@r1VeN2k That's no longer a viking, it's whole ass Hulk

[cognit @0x_cognit]

@andreysuperior My bad, thanks for pointing it out. It's up now

[cognit @0x_cognit]

@AIDuckLabs Updated the post thank you

[cognit @0x_cognit]

This is impressive.
I gave Claude one prompt and got a working siege weapon. A working trebuchet.

Wind the winch, load the stone, pull the pin, let it fly. Change the weight and the stone either falls short or sails over the castle wall.

Claude wrote the physics model first: a 95 kg weight drops 0.97 m and stores 900 J. Every gauge, chip and bar on screen reads from that same model, so the numbers can’t disagree. 31.5 m/s at release, 42°, 4.3 s in the air, 100 m. The real swing takes 0.58 seconds, so the simulation plays it at ×1/6.

This article made it as simple as one prompt

[Elvar @elvar_official]

https://t.co/NAgH1EYziQ

[cognit @0x_cognit]

@coolerwave Good catch on 1,200 sites attempted versus 100+ actually breached.

[cognit @0x_cognit]

@PhnxArc Interactive demos beat static ones. Going to try the anatomy skill.