Codez

Content creator | AI researcher & builder | AI insights from 2030 | @beyond_xai

[Codez @0xCodez]

https://t.co/eDBs67VMq2

[Codez @0xCodez]

this is impressive…

I can’t believe this small open-source model by @thewebAI is so good.

A 2.24 GB model just ran 2.4x faster than Qwen 9B.

No GPU. Just an Intel Xeon with 15 GiB RAM.

→ TwIL-LM3-Pro: 2.24 GB, 10.47 tok/s
→ Qwen3.5-9B: 5.67 GB, 4.36 tok/s

Same CPU, same memory limit. TwIL is faster just because its weights are smaller.

So I made them race in Snake

Same board. Same apples. 60 seconds. Real speeds.

TwIL: 450 pts
Qwen 9B: 210 pts

The model you can fit beats the model you can’t.

[Codez @0xCodez]

Holy sh*t, I turned my Meta Ray-Ban glasses into a pro golf coach with Opus 5.5

Claude Opus 5.5 just reverse-engineered my Meta glasses...

1 prompt + 20 min = Claude Code agents built an AR putting coach that floats over the green.

here’s how Opus did it, step by step:

step 1 → it sniffed the Bluetooth traffic and decompiled the Meta AI app.

the loophole: whatever the phone renders, the lens shows

step 2 → it split the work across 4 agents in tmux: camera stream, vision, putt physics, UI

step 3 → SLAM + depth AI turn camera frames into a 3D mesh of the green, scaled by the ball (42.67 mm) and the cup (108 mm)

step 4 → it reads the slope of every 5 cm of grass and draws contour lines over the break

step 5 → it simulates hundreds of putts and keeps the one that finishes 17 inches past the cup.

blue = the ball’s path, red = where to aim

step 6 → hand tracking turns your fingertip into a mouse. tap a bubble, switch the view

the result: the glasses read the break, you just stroke the ball...

p.s: mine in the video is a practice green. Practice only, slope readers are banned in competition

Find full prompt below

Then read the article below on how to reverse-engineer anything

[Codez @0xCodez]

@0xMovez damn, this game looks wild

[Codez @0xCodez]

@thejayden @thewebAI yeah, true brother !

[Codez @0xCodez]

this is impressive…

I can’t believe this small open-source model by @thewebAI is so good.

A 2.24 GB model just ran 2.4x faster than Qwen 9B.

No GPU. Just an Intel Xeon with 15 GiB RAM.

→ TwIL-LM3-Pro: 2.24 GB, 10.47 tok/s
→ Qwen3.5-9B: 5.67 GB, 4.36 tok/s

Same CPU, same memory limit. TwIL is faster just because its weights are smaller.

So I made them race in Snake

Same board. Same apples. 60 seconds. Real speeds.

TwIL: 450 pts
Qwen 9B: 210 pts

The model you can fit beats the model you can’t.

[webAI @thewebAI]

We ran TwIL on a CPU with 15 GiB of RAM. No GPU. It generated tokens 98x faster than Qwen3.8-27B on the same hardware.

The test:
- Intel Xeon CPU, 4 threads
- 15 GiB RAM
- No GPU

Results:
- TwIL-LM3-Pro (2.24 GB weights): 10.47 tokens per second
- Qwen3.5-9B (5.67 GB weights): 4.36 tokens per second
- Qwen3.8-27B (16.45 GB weights): 0.107 tokens per second

TwIL and Qwen 9B both run at the RAM's memory limit of about 24 GB/s. Qwen 9B is 2.4x slower only because its weights are 2.5x larger.

Qwen 27B's 16.45 GB of weights don't fit in the 15 GiB of RAM. Parts have to be re-read from disk for every token. Size alone would make it 7.3x slower. The disk reads add another 13x. Total: 98x slower than TwIL.

With enough RAM, the gap would shrink to about 7x. But most people don't have unlimited RAM, and most people don't have a top-tier GPU. The model you can actually fit matters more than the one you can't.

Download TwIL-LM3-Pro on @huggingface: https://t.co/pTRuIDyV0L

[Codez @0xCodez]

@AlterEgo_eth @thewebAI yeah bro, do it !

[Codez @0xCodez]

@lez99999 @thewebAI Yeah, I think these small open-source models are the future, especially for agentic AI.

[Codez @0xCodez]

@humzaakhalid @thewebAI Yeah, I’m personally using it a lot right now - definitely worth the switch, bro.

[Codez @0xCodez]

@0xCarnagee damn bro, it's looks crazy

[Codez @0xCodez]

Try it now:
https://t.co/uBt6OdmoXR

[Codez @0xCodez]

@0xMovez wow, Krea just keep pushing the limits !

[Codez @0xCodez]

Full Prompt:

You're a senior AR + computer vision engineer.

I own Meta Ray-Ban 2 glasses (photo attached) and an Android phone. Turn the glasses into a real-time AR putting coach.

STEP 1: RECON

- Walk me through enabling Bluetooth HCI snoop log, then parse btsnoop_hci.log with tshark
- Decompile the Meta AI app with jadx and map how the camera stream and display overlay are called
- Cross-check against Meta's Wearables Device Access Toolkit
- Output: a spec of how to pull camera frames to the phone and push rendered pixels back to the lens

STEP 2: AGENTS

Spin up 4 Claude Code agents in tmux with one shared SPEC.md:
- agent 1: camera stream + overlay link
- agent 2: vision
- agent 3: putt physics
- agent 4: UI
Each agent tests against recorded footage of a green before touching the live glasses.

STEP 3: SCAN THE GREEN

- Monocular SLAM for head pose + Depth Anything V2 for dense depth
- Metric scale from known objects: ball = 42.67 mm, cup = 108 mm
- Point cloud → Delaunay mesh → heightmap with 5 cm cells

STEP 4: READ THE BREAK

- Slope per cell from the heightmap gradient
- Contour line every 1 cm of elevation (marching squares)
- YOLO detects the ball, flagstick + dark ellipse = the hole

STEP 5: SIMULATE THE PUTT

- Rolling ball physics: 5/7 g along the slope, friction from green speed (default Stimp 10, editable)
- RK4 integration, 1 ms steps
- Search launch angle + speed so the ball finishes 17 inches past the cup
- Render: ideal line (blue), aim line (red), aim point, corridor of putts that still drop, yellow setup lines at my feet square to the aim line

STEP 6: HANDS-FREE UI

- MediaPipe Hands on the same camera feed, index fingertip on a bubble = tap
- Floating bubble menu: Ideal Line, Aiming Line, Contour Lines, Setup Lines, Corridor, Putt Info, Animation
- Actions: New Ball, New Target, New Putt

RULES

- Re-project every overlay into the current camera pose each frame, under 100 ms end to end
- Ask me before anything that needs a physical action (pairing, approving the connection)
- Ship a working build I can test on a practice green today

[Codez @0xCodez]

Holy sh*t, I turned my Meta Ray-Ban glasses into a pro golf coach with Opus 5.5

Claude Opus 5.5 just reverse-engineered my Meta glasses...

1 prompt + 20 min = Claude Code agents built an AR putting coach that floats over the green.

here’s how Opus did it, step by step:

step 1 → it sniffed the Bluetooth traffic and decompiled the Meta AI app.

the loophole: whatever the phone renders, the lens shows

step 2 → it split the work across 4 agents in tmux: camera stream, vision, putt physics, UI

step 3 → SLAM + depth AI turn camera frames into a 3D mesh of the green, scaled by the ball (42.67 mm) and the cup (108 mm)

step 4 → it reads the slope of every 5 cm of grass and draws contour lines over the break

step 5 → it simulates hundreds of putts and keeps the one that finishes 17 inches past the cup.

blue = the ball’s path, red = where to aim

step 6 → hand tracking turns your fingertip into a mouse. tap a bubble, switch the view

the result: the glasses read the break, you just stroke the ball...

p.s: mine in the video is a practice green. Practice only, slope readers are banned in competition

Find full prompt below

Then read the article below on how to reverse-engineer anything

[Movez @0xMovez]

https://t.co/LCLhoxByi3

[Codez @0xCodez]

@0xMovez Wow. What a clean setup, sending to my Claude, asap

[Codez @0xCodez]

@andreysuperior Yeah. So true

[Codez @0xCodez]

Holy sh*t… Wabi just made the App Store feel old.

→ 2008: 500 apps.
→ 2009: "there's an app for that."
→ Today: millions of apps, and your home screen looks exactly like everyone else's.

Wabi flips it. You don't download an app, you ask for one:

→ "help me train for my half marathon"
→ it builds you a marathon coach: your race, your date, your 16-week plan
→ right inside the chat you already use with friends & family

There used to be an app for that.
Now there's an app for you.

[Eugenia Kuyda @ekuyda]

Wabi 2.0 is live!

personal agents are one person, one chat. but the parts of life that matter run on more than one person

Wabi is an agent you use together. you talk to it, it does stuff, builds the apps you need, and anyone can join the thread - your partner, your best friend, the preschool moms groupchat

school changes pickup time → Wabi updates the family calendar, everyone sees it.
a show you’d both love → Wabi finds a night that works, gets the tickets.
the trip everyone keeps talking about → Wabi picks the dates, builds the itinerary, books it

go start a thread with your people 👇

[Codez @0xCodez]

@ekuyda wabi is my daily tool right now !

[Codez @0xCodez]

Holy sh*t, I’m running Claude Code on CarPlay right in BMW dealership

Claude Opus 5.5 just reverse-engineered my CarPlay...

1 prompt + 20 min = 4 Claude Code agents shipping code live on my dash. Apple never approved a thing

here’s the full guide step by step:

step 1 → plug an Android CarPlay AI box into your car’s USB (Ottocast, Carlinkit Tbox)

step 2 → paste the prompt below into Claude Code on Opus 5.5 with a photo of your head unit

step 3 → Opus maps how CarPlay draws pixels and finds the loophole: whatever runs on the box, the car displays

step 4 → it spins up 4 Claude Code agents in tmux and streams them to your dash

step 5 → approve the connection on the screen once, the last time you lift a finger

step 6 → tap a pane to jump between agents, approve diffs, ship from the driver’s seat

the result: the laptop does the work, the dash is the control room...

p.s: works on almost any car with wireless CarPlay. Mine in the video is a BMW iDrive. Parked only

Find full prompt below

Then read the article below on how to reverse-engineer anything

[Codez @0xCodez]

@0xMovez wow, will test ego lite today !

[Codez @0xCodez]

@polyholyman lol, ultracode mode only 😆

[Codez @0xCodez]

Holy sh*t, I’m running Claude Code on CarPlay right in BMW dealership

Claude Opus 5.5 just reverse-engineered my CarPlay...

1 prompt + 20 min = 4 Claude Code agents shipping code live on my dash. Apple never approved a thing

here’s the full guide step by step:

step 1 → plug an Android CarPlay AI box into your car’s USB (Ottocast, Carlinkit Tbox)

step 2 → paste the prompt below into Claude Code on Opus 5.5 with a photo of your head unit

step 3 → Opus maps how CarPlay draws pixels and finds the loophole: whatever runs on the box, the car displays

step 4 → it spins up 4 Claude Code agents in tmux and streams them to your dash

step 5 → approve the connection on the screen once, the last time you lift a finger

step 6 → tap a pane to jump between agents, approve diffs, ship from the driver’s seat

the result: the laptop does the work, the dash is the control room...

p.s: works on almost any car with wireless CarPlay. Mine in the video is a BMW iDrive. Parked only

Find full prompt below

Then read the article below on how to reverse-engineer anything

[Movez @0xMovez]

https://t.co/LCLhoxByi3