SJ

Make AI run faster on your robots. Co-founder & CBO @zeticai_

[SJ @sj_zetic]

@zetic_ai @NVIDIARobotics @zeticai_

[SJ @sj_zetic]

We redesigned how Cosmos Policy runs on Jetson Thor:

1.83s → 275ms
p50 per 16-action chunk.

LIBERO-Spatial sim;
Official skip-decode baseline.

I’d love to hear what’s slowing your robot down.
DM me your model + hardware.

[SJ @sj_zetic]

@JasonMa2020 Boring companies are the best!

[SJ @sj_zetic]

@roseajohnson @Techweek_ Exciting week ahead ! 💪

[SJ @sj_zetic]

@Thom_Wolf @huggingface Still waiting for mine to be shipped! 🐣

[SJ @sj_zetic]

@Techweek_ Gonna be fantastic

[SJ @sj_zetic]

Grab a spot here: https://t.co/OVWRvh9VB3

Mission leaders: me + @vivmarquez from Liquid AI's DevRel team. Astronaut attire optional, but encouraged.

#ZETIC #OnDeviceAI #LFM #LiquidAI

[SJ @sj_zetic]

48 hours to launch 🚀

Tue 10/6, 5:30 PM in SF: an on-device AI workshop with @liquidai for #SFTechWeek.

Get models running on your own phone, no cloud required.

Bring a laptop, cable, and device. We close with telescopes and an astronomer.

RSVP 👇

[SJ @sj_zetic]

And back in the office 💪
#PhysicalAI #ZETIC

[SJ @sj_zetic]

Sunday morning run with the team before #SFTechWeek #TechWeek

[Yeon Kim @zetic_ai]

@physical_int
pi0.5 policy inference just went sub-20 ms on Jetson Thor by us.

p50: 19.16 ms / p95: 19.20 ms
37.3x vs our original 714 ms PyTorch reference.

Runtime optimization with Execution Design.
Closed-loop success: 448/500 (89.6%).

Make robot to not to be like a sloth anymore!

[SJ @sj_zetic]

𝗪𝗲 𝗰𝘂𝘁 π𝟬.𝟱 𝗽𝗼𝗹𝗶𝗰𝘆 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗳𝗿𝗼𝗺 𝟳𝟭𝟱 𝗺𝘀 𝘁𝗼 𝟭𝟵.𝟭𝟲 𝗺𝘀.
𝗡𝗩𝗜𝗗𝗜𝗔 𝗝𝗲𝘁𝘀𝗼𝗻 𝗔𝗚𝗫 𝗧𝗵𝗼𝗿.

37.3x faster policy inference at p95.
• p50: 715 ms → 19.16 ms.
• p95: 715 ms → 19.20 ms.

#ZETIC #PhysicalAI #Robotics #VLA #EdgeAI

[SJ @sj_zetic]

Jev-style decision AI, on-device. 🎮

KEV-4B on Jetson Thor.
Nine plays at 1,000 km/h simulated speed.

❌ Baseline: 0/9 - all Game Over
✅ ZETIC optimized: 9/9 still running !

We make AI run faster on your hardware.
DM me.

https://t.co/yQ6SMShGDJ

#Jev #DecisionModel #ZETIC

[SJ @sj_zetic]

Want to make your AI run faster on your robots? Send me a DM.

Visit https://t.co/yQ6SMShGDJ

[SJ @sj_zetic]

𝗪𝗲 𝗰𝘂𝘁 π𝟬.𝟱 𝗽𝗼𝗹𝗶𝗰𝘆 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗳𝗿𝗼𝗺 𝟳𝟭𝟱 𝗺𝘀 𝘁𝗼 𝟭𝟵.𝟭𝟲 𝗺𝘀.
𝗡𝗩𝗜𝗗𝗜𝗔 𝗝𝗲𝘁𝘀𝗼𝗻 𝗔𝗚𝗫 𝗧𝗵𝗼𝗿.

37.3x faster policy inference at p95.
• p50: 715 ms → 19.16 ms.
• p95: 715 ms → 19.20 ms.

#ZETIC #PhysicalAI #Robotics #VLA #EdgeAI

[SJ @sj_zetic]

For context, @nvidia's Jetson AI Lab tutorial: ~48 ms with TensorRT FP8 + NVFP4.
https://t.co/tWxwDm08LC

Our 19.20 ms p95 is ~60% below that published figure, a 2.5x ratio between the reported latencies.

With an acceptable 89.6% Closed loop accuracy.

[old account @oldaccountkr]

https://t.co/eEn83HIFud

We aren't going anywhere before we get out of the Monday range. Personally I'm with the bears 🐻

[old account @oldaccountkr]

When you get lost zoom out a bit and you'll see a clearer picture. The bulls had several chances to change the picture but failed to do so. The daily and weekly looks like shit

[Dakota Smith @daksmitty]

I'm not in this one but this is one of my favorite patterns to trade. Ideally, we start closing above these grey boxes. Look for a retest of the grey's for entries if you are not already in. $cream