@iam_agg

Currently hacking on the Humane Ai Pin (previously Vision Pro and Analogue Pocket). Mastodon: https://xitter.cf/t.co/BX8F6i31Hu

github.com/agg23
Joined May 2021
  • Tweets1.2K
  • Following235
  • Followers2.5K
79Media
Big Humane Ai Pin milestone: Full preservation of the original experience, entirely on-device. I've applied Humane's UX work and polish so we can build whatever functionality we want Video: Ai Mic -> cosmOS -> Gemini -> live search Powered by PenumbraOS penumbraos.com/
6
5
3
32
3,436
Adam Gastineau retweeted
SNSnote
@snsnotecomOct 23
1984년 매킨토시로 변한 아이폰 듀오 애플의 첫 폴더블폰 〈iPhone Duo〉를 반쯤 접으면 1984년 매킨토시를 재현한 작은 노트북으로 변신합니다. 새롭게 공개된 〈Tiny Computers〉는 매킨토시부터 윈도우 95·XP까지 고전 컴퓨터 10종을 체험할 수 있는 앱입니다. 키보드 입력은 물론 그림 그리기와 간단한 게임도 즐길 수 있으며, 일반 아이폰과 아이패드에서도 이용할 수 있습니다. 🎥 idanbo
152
1,889
417
11,718
1,987,963
I'm sure people are buying a $2k+ iPhone Duo to... Play 40+ year old Game & Watch games! Now you can wish your fancy new phone into the 80s hands of a child
2
1
1
20
1,218
Adam Gastineau retweeted
Analogue
@analogueOct 16
Introducing Super Nt+ and Mega Sg+ The definitive way to play original Super Nintendo, Super Famicom, Sega Genesis, Mega Drive, and Master System. Beginning with Super Nt in 2017, the originals set a new standard for video game preservation. Nearly a decade later, they return. Super Nt+ & Mega Sg+. Definitive, all over again. Now with Bluetooth + 2.4G. Reworked industrial design, OS, packaging, and more. At the core: an FPGA architecture engineered in-house from scratch, backed by over a decade of development. Purpose-built 1080p for 2D gaming. No emulation. Zero lag. Total accuracy. Plug and play. Preorder Oct 19, 8am PDT. Ships Dec 2026. analogue.co
276
479
143
3,459
12,342,558
Have people settled on a location for sharing LLM speed benchmarks? The sites I find seem abandoned It's frustrating not being able to compare. I can't tell if M5 Ultra Q8, 32k, single stream at 185 t/s is good or not. Some places say 6000 Pro isn't typically hitting that at Q4
309
I wanted to try messing with (prompting) inference improvements with M5 Ultra. I'm rather surprised at the results and after a lot of checking I think they're accurate M5 Ultra 256GB, 80 core, Qwen3.8 Flash Next Q8 Up to 85% faster decode vs oMLX main: 378 t/s @ 32k, 8 streams
1
2
764
This largely is improving MTP performance on larger contexts or multiple streams. I don't know if people care much about this work, but I will be slowly trying to PR it upstream. I hate LLM code and LLM opened PRs, so I'm having to review, understand, and post PRs manually
1
177
If you found PenumbraOS (the jailbreak for the Ai Pin) to be too cumbersome, Anders' Luma is probably more like the tool you're looking for.
Anders ᯅ@AndersDoesTechOct 2
Your Ai Pin. Yours again. github.com/TheAndersMadsen/l…

GitHub - TheAndersMadsen/luma: Meet Luma: An independent self-hosted cloud, control center, and...

Meet Luma: An independent self-hosted cloud, control center, and device runtime for an operator-owned Humane Ai Pin. - TheAndersMadsen/luma

github.com
3
449
If you're an iOS dev, you may have noticed that Device Hub doesn't support multitouch gestures with Option + Mouse, unlike Simulator has for a long long time (maybe forever?). This simple tool restores the lost behavior. github.com/agg23/DeviceHubMo…
1
1
394
GLM 5.3 Flash, Q4, 1 stream, 128k context, code: oMLX (latest main): 69.8 tok/s decode MTP (53.3 without), 2288 tok/s prefill Prefill barely unaffected by context size
Adam Gastineau@iam_aggSep 29
Batch 2 of M5 Ultra benches. Things are changing fast Qwen 3.8 Flash Next, Q4, MTP, 1 stream, 128k context, code: mlx-serve 26.9.6: 141 tok/s decode, 3783 tok/s prefill oMLX main (unreleased): 112 tok/s, 5331 tok/s prefill oMLX 0.7.0rc1: 105 tok/s, 3624 tok/s prefill
2
2
706
For those with a 256GB Mac who wants to try it, you have to merge the MTP head from a different model against the current oMLX quants
194
Batch 2 of M5 Ultra benches. Things are changing fast Qwen 3.8 Flash Next, Q4, MTP, 1 stream, 128k context, code: mlx-serve 26.9.6: 141 tok/s decode, 3783 tok/s prefill oMLX main (unreleased): 112 tok/s, 5331 tok/s prefill oMLX 0.7.0rc1: 105 tok/s, 3624 tok/s prefill
Adam Gastineau@iam_aggSep 28
Rough M5 Ultra 80 core GPU 256GB benches: Qwen 3.8 Flash Next, Q4 variants, MTP, 1 stream, 32k context, averaged code/prose: TensorFold 0.3.4.1: 152 tok/s decode, 2086 tok/s prefill mlx-serve 26.9.6: 124 tok/s, 3863 tok/s prefill oMLX 0.7.0rc1: 105 tok/s, 3694 tok/s prefill
2
1
6
1,857
more replies
Q8 continues to perform just as well with the new changes. I had to submit a small patch to fix a OOM bug, so I imagine few people have tried Q8 on Macs.
1
1
194
I'm hoping to do GLM 5.3 Flash next (it is running right now). The models are large enough that juggling them and trying to maintain a clean benchmark environment is hard, plus I'm just inexperienced. I'm learning (slowly)
2
138
Rough M5 Ultra 80 core GPU 256GB benches: Qwen 3.8 Flash Next, Q4 variants, MTP, 1 stream, 32k context, averaged code/prose: TensorFold 0.3.4.1: 152 tok/s decode, 2086 tok/s prefill mlx-serve 26.9.6: 124 tok/s, 3863 tok/s prefill oMLX 0.7.0rc1: 105 tok/s, 3694 tok/s prefill
10
1
5
50
17,163
Somewhat similar TensorFold numbers against M3 Ultra xitter.cf/ivanfioravanti/status/…
Ivan Fioravanti
@ivanfioravantiSep 27
Long Context Benchmark: TensorFold 0.3.4.1 + PR 29 and PR 30 on M3 Ultra with Qwen 3.8 Flash Next 4bit. 🔥 Now prefill is great and there are no more issues with 64K+ context on first decode request. Here a comparison 0.3.4, 0.3.4.1 and this version. Peak: - 0.3.4.2 130 t/s - 0.3.4 128 t/s - 0.3.4.1 119 t/s
1
618
It is pretty hard to approach this scientifically, especially since I am not familiar with the field. I would not be surprised if the choice of quants messed up the engine comparisons significantly. TensorFold in particular is built for a very specific model + quant at the moment
3
323
Could only get oMLX to run Q8, so here's a Q4 vs Q8 comparison for oMLX only. Seems to be a ~12-14% cost to moving to Q8 ad the moment. 95 tok/s Q8 vs 105 tok/s Q4 @ 32k
2
2
634
I think there's a lot of headroom to be squeezed out still, and it's obviously early in the chip's life. A rough test shows the GPU can draw ~162W. GPU power, decode / prefill: TensorFold: 97 W / 123 W mlx-serve: 87 W / 130 W oMLX: 71 W / 150 W Decode maxes out at 430-460 GB/s
3
462
Load more