Jonathan Spangler

Network /Systems Engineer , USMC Vet

[Jonathan Spangler @Spangler3000]

FSD 14.1.3 First Impressions . VERY IMPRESSIVE

All short distance drives , lengthy test tomorrow.
Drive 1: no brake stabbing or uncertainty at all smoother and more confident than v13
Drive 2: insanely impressive parking maneuvers and decision making.
Drive 3 : Slight brake stabbing when leaving parking spot . Possibly due to light fogging of windshield .

Overall , the limited initial testing has been extremely positive. From a technical perspective it’s honestly so impressive the reaction time AI4 is performing at .

The rate of improvement from v11 to 14 proves that autonomy is solved.
Insane work from @Tesla_AI and @aelluswamy for leading the AP team 👏

[Jonathan Spangler @Spangler3000]

Inference is hard.

[Jonathan Spangler @Spangler3000]

@jun_song A lot of cases traditional automation is still the right choice . Absolutely scale ai and a2a and agent loops but do it properly and not forced .

[Jonathan Spangler @Spangler3000]

@ivanfioravanti Me too while they are at it ! I’ll continue cluster ! Got Minimax m3 cranking ! Sharded over m5 max and m3u , 380 tok/s prefill and 34 tok/s decode .

[Jonathan Spangler @Spangler3000]

@theemozilla Ds4 flash and minimax m3

[Jonathan Spangler @Spangler3000]

@Prince_Canuma 💪💪

[Jonathan Spangler @Spangler3000]

@rtk86 @jun_song Finally beta release worthy ! For your testing if you want a faster Minimax m3: https://t.co/WeVvIQwbb8

[Jonathan Spangler @Spangler3000]

@altruisticsoni @MiniMax_AI @Gradient_HQ @tryParallax @GA_agent_ai Sorry for the delay , Finally beta release worthy ! For your testing if you want a faster Minimax m3: https://t.co/WeVvIQwbb8

[Jonathan Spangler @Spangler3000]

@MiniMax_AI @Gradient_HQ @tryParallax @GA_agent_ai Finally beta release worthy ! For your testing if you want a faster Minimax m3: https://t.co/WeVvIQwbb8

[Jonathan Spangler @Spangler3000]

@MiniMax_AI @Gradient_HQ @tryParallax @GA_agent_ai I got it running local over rdma sharded across 2 Macs this weekend . Tuning in progress . 🔥

[Jonathan Spangler @Spangler3000]

@digitalix For your testing if you want a faster Minimax m3: https://t.co/WeVvIQwbb8

[Jonathan Spangler @Spangler3000]

@digitalix Yep ! Distributed inference on mlx at this early stage is 🔥🔥

[Jonathan Spangler @Spangler3000]

@Prince_Canuma https://t.co/WeVvIQwbb8 , first stable release worth point !!! Lots of work into this . Very specific but happy to buildout distributed mlx !

[Jonathan Spangler @Spangler3000]

@jun_song Already an mlx port?

[Jonathan Spangler @Spangler3000]

@kaiostephens @JAKEVIN16 Use Zcoder , much better support .

[Jonathan Spangler @Spangler3000]

@Prince_Canuma @ivanfioravanti @kernelpool @angeloskath Got it stabilized! Full disclosure this was 90% optimized by agents as it began as honestly a proof of concept to myself of agents in full local loops . Turned into a lot more . Sent inv to repo before public launch . Cheers .

[Jonathan Spangler @Spangler3000]

@jun_song It’s so bad

[Jonathan Spangler @Spangler3000]

@Prince_Canuma @ivanfioravanti @kernelpool @angeloskath 🫡🫡🔥 I have not given up yet ,but will do. M5 max prefill flies past . The m5 max prefill is so much faster . Beats my m3u by quite a bit .

[Jonathan Spangler @Spangler3000]

@Prince_Canuma @ivanfioravanti @kernelpool @angeloskath Hit maybe a MSA issue or mlx vlm distribution rank issue . Trying to stabilize it now . Causing constant memory orphan issues. Lmk if you have any tips but I did get Minimax to 34 tok/s decode and 403 tok/s prefill , 377 at 150-200k context . To specify , again across 2 mac’s

[Jonathan Spangler @Spangler3000]

@Prince_Canuma M5 Ultra 😂

[Jonathan Spangler @Spangler3000]

@kimmonismus Yeah it sucks at optimizing models too . It broke so much of what gpt and ds4 local were able to optimize.