←
Paolo Rosson1,547 posts
Paolo Rosson avatar

Paolo Rosson

@redp314

Head of Applied AI at Dext. Building menuphotoai.com | Quantum physics PhD from Oxford | Former Italian blindfolded Rubik’s cube record holder

◷ Joined May 2015▧ 236 Photos and videos
Paolo Rosson avatar
Paolo Rosson@redp314·Aug 14

Dentist did a 3D X-ray of my jaw before a root canal and said I wouldn't be able to open the raw data, it needs specialized software. It's 800 DICOM files. Asked Claude Code to make me a viewer. Two prompts later... this is nicer than what he showed me on his screen.

0:13
Paolo Rosson avatar
Paolo Rosson@redp314·Nov 17

You might not need to train an image classifier anymore. Tested @liquidai’s d1-2B on my 5-year-old M1 Max: type a yes/no question, it sorts 8 photos in under 2 s. No labels, no fine-tuning, 0 tokens generated, it just reads its own yes/no probability. Real time, no cuts. Reply with a question and i’ll run it!

0:25
Paolo Rosson avatar
Paolo Rosson@redp314·Nov 12

Nothing can ever be as iconic as the actual discovery of this tomb back in 1922. I just got back from Egypt and wanted to visualize what it'd be like!

Paolo Rosson avatar
Paolo Rosson@redp314·Nov 12

I asked Fable 5.2 to recreate the iconic discovery of King Tut’s tomb by Carter in 1922. It rebuilt the full chamber with 160 individual objects, each placed from Carter’s original plan, then it checked it against the photo taken that week!

0:40
Paolo Rosson avatar
Paolo Rosson@redp314·Nov 11

Turns out the full extended screen of the @EvenRealities G2, 480 vertical pixels instead of 288, can be driven at 24 fps! That’s fast enough for real video. For now, a spinning moon drawn live on the glasses.

0:08
Paolo Rosson avatar
Paolo Rosson@redp314·Nov 11

Just managed to unlock the full screen on the @EvenRealities G2! Apps get a 576 x 288 window. Turns out the panel behind it is 640 x 480, and I can use all of it! So now my app draws the other rows itself: one 576 x 480 picture, 67% more pixels, almost a square!

0:06
Paolo Rosson avatar
Paolo Rosson@redp314·Nov 11

“So when I ask ChatGPT something, are you the one typing the answer?” A guide in Egypt asked me this when I told him I work in AI. For him, a human behind the screen was more believable than a computer doing it.

Paolo Rosson avatar
Paolo Rosson@redp314·Nov 10

Got Fable 5.2 to run the Claude Code CLI natively on my iPhone. So now, instead of a VPS, the phone is where Claude Code runs and I drive it from my laptop with Remote Control! It’s the actual Claude Code binary for arm64, on a stock iPhone. No jailbreak, no SSH. Ask it where it’s running and it tells you: Darwin 27.3.0, the kernel under iOS. Shell, file edits and search all work on the phone. 35 minutes locked and it still answered!

0:15
Paolo Rosson avatar
Paolo Rosson@redp314·Nov 10

Lost my Oura ring somewhere in my apartment but had no clue where. So I asked Claude Code to help me find it. It turned my MacBook into a Bluetooth hot/cold detector! Turns out the ring only sends a signal every few seconds, so walking around with a live meter was useless. So it switched to a survey: hold the Mac still at each spot for 20 s, rank the spots by the strongest signal. Survey said: the cardboard pile. Found it in a plastic bag right next to the boxes!

Post attachment
Paolo Rosson avatar
Paolo Rosson@redp314·Oct 31

Got NVIDIA’s Open-H3-Regenerate-2K running on my DGX Spark! It redraws a 768P video draft at 2K in 3 steps on MiniMax H3. Left: plain resize. Right: regenerate. Speed on the Spark: ~6 min per step, so ~18 min of refine for a 5 s clip, plus ~8 min to load. The model card does it in ~20 s on a B200, so this one is a quality pass, not a quick one

0:05
0:05
Paolo Rosson avatar
Paolo Rosson@redp314·Oct 30

Got Claude to direct a full short movie on my DGX Spark, with local models only! Script + shot list: Claude Storyboard: Qwen-Image 2.1 Video + sound effects: MiniMax-H3 via @pika_labs H3 Flex Music: Claude, written as Python code Edit: ffmpeg Crazy what can be done on one box under my desk!

0:25
Paolo Rosson avatar
Paolo Rosson@redp314·Oct 29

Pika H3 Flex from @pika_labs just dropped as open weights, so naturally i put it on my DGX Spark! Video + audio from a text prompt, all local: a cup by a window, a fox in the snow, wind, birds and crunching snow included. 95 s to generate a 5 s clip and the quality is amazing!

0:10
Paolo Rosson avatar
Paolo Rosson@redp314·Oct 28

Just managed to turn my @EvenRealities G2 into external monitors for my Mac, pinned to fixed spots in my room! Look at one and it snaps in, look away and it’s gone. Switching between them takes about half a second!

0:19
Paolo Rosson avatar
Paolo Rosson@redp314·Oct 27

Is this new? Almost got scared after seeing this hand pop up on X while making a post @elonmusk

0:06
Paolo Rosson avatar
Paolo Rosson@redp314·Oct 27

Amazing improvement! Getting 20% higher tps in my tests!

Mia@MiaAI_labOct 27

Do you like more speed? I know I do ⚡️ MASSIVE Qwen3.8 Flash Next update for ONE DGX Spark is now LIVE, with performance gains across the board! 🔥 Code tok/s gains: 1 stream: +62.7% 2-5 streams: +45-47.1% Prose, tok/s gain (non-greedy, thinking on): 1 stream: +10.1%, 2 streams: +29.8% 4 streams: +36.4%, 5 streams: +17.6% Prefill tok/s gains: 4k: +14.5%, 8k: +14.4% 16k: +13.5%, 32k: +12.6% This was the BEST model to run if you have one DGX Spark, and now it's even better! nothing else comes close in terms of quality, speed, and concurrency. Link & full tok/s and prefill numbers below.

Post attachment