#asahi-re · 2026-06-30 03:21
nicolas17but it turns out it's not a student, the author *is* an "assistant professor"#asahi-re · 2026-06-30 03:21
JamesCalligeros[m]>:)#asahi-re · 2026-06-30 03:21
nicolas17same, and I also thought this student's professor would like to know that the paper is BS#asahi-re · 2026-06-30 03:20
JamesCalligeros[m]my theory is that someone wanted a scholarship/funding#asahi-re · 2026-06-30 03:20
JamesCalligeros[m]yeah its not a big deal dont apologise :P#asahi-re · 2026-06-30 03:18
nicolas17amarioguy: that's the problem with LLMs, it's not *that* obvious#asahi-re · 2026-06-30 03:16
amarioguywouldn't have linked it if i found that out...i really need to stop giving the benefit of the doubt in this situation#asahi-re · 2026-06-30 03:15
amarioguybruh....so sorry for not realizing the paper was slop assisted #asahi-re · 2026-06-30 02:25
JamesCalligeros[m]if not entirely slop then definitely slop-assisted#asahi-re · 2026-06-30 02:25
JamesCalligeros[m]things are not referenced properly and there are silly things in there#asahi-re · 2026-06-30 02:25
JamesCalligeros[m]yeah i got a few pages in and stopped reading#asahi-re · 2026-06-30 02:23
nicolas17the whole h13/A13/M1 confusion isn't even self-consistent within the paper#asahi-re · 2026-06-30 02:22
nicolas17more evidence has been found in hackdifferent yeah#asahi-re · 2026-06-30 02:22
nickchanthat paper has to be slop#asahi-re · 2026-06-29 22:24
chaos_princessin case of coreml, there isn't even a "force ane" option#asahi-re · 2026-06-29 22:24
chaos_princessyep#asahi-re · 2026-06-29 22:23
yukaAnd the MLX stuff probably distributes the work on cpu/gpu/ane depending on what's best suited or able to accelerate the workload?#asahi-re · 2026-06-29 22:21
chaos_princessbasically, it can run convolutional networks pretty well, you can offload encoders from multimodal models on it, and well, that's it#asahi-re · 2026-06-29 22:18
yukaah, fair#asahi-re · 2026-06-29 22:17
chaos_princessit is. ane can only run the encoder, and the model runtime is dominated by the decoder#asahi-re · 2026-06-29 22:15
yukathis doesn't seem useless?#asahi-re · 2026-06-29 22:15
yukahttps://github.com/ggml-org/whisper.cpp/pull/1021#asahi-re · 2026-06-29 22:14
chaos_princessalso, ane is kinda useless on !macos#asahi-re · 2026-06-29 22:13
chaos_princessthere isn't a "real" linux ane driver#asahi-re · 2026-06-29 22:05
yukais more understanding necessary to use more features or improve the performance?#asahi-re · 2026-06-29 22:05
yukahow relevant is this anyways, given there's already an ane driver?#asahi-re · 2026-06-29 22:03
amarioguynickchan: yep....kinda disappointed and puts a good bit of the paper's credibility under question...#asahi-re · 2026-06-29 17:29
nickchanthere's some other mistakes like saying the host page size is 4KB as well#asahi-re · 2026-06-29 17:24
NickChan[m]already spotted errors in 12.1#asahi-re · 2026-06-29 17:14
amarioguynote that part of the methodology did involve binary RE hence why it's here#asahi-re · 2026-06-29 17:14
amarioguyhttps://arxiv.org/pdf/2606.22283#asahi-re · 2026-06-29 17:14
amarioguybut interesting paper came out recently on ANE, someone wrote a 300 page research paper on RE'ing the ANE#asahi-re · 2026-06-29 17:13
amarioguyjust realized i was disconnected before sending last time....#asahi-re · 2026-06-18 22:56
StanfordAppliedCyber[m]literally just struct A, struct B, struct C, etc, is probably safest#asahi-re · 2026-06-18 22:56
StanfordAppliedCyber[m]which is actually better from a copyright perspective, just don't include anything tainted in whatever docs you write#asahi-re · 2026-06-18 22:56
StanfordAppliedCyber[m]i don't think we have full field info from the included symbols, so you'll have to infer based on us + surrounding function names#asahi-re · 2026-06-18 22:54
StanfordAppliedCyber[m]_oftc_jailbreakhat[m]: the most helpful thing you could do for me, imo, to get the gpu driver working, is just try to dump all the structs that comprise the firmware abi#asahi-re · 2026-06-18 07:50
JamesCalligeros[m]The pcores are pretty high up the curve too. Plot the opp-microwatt values ;)#asahi-re · 2026-06-18 07:47
nickchantop of the line performance per watt is always going to suck but it's still possible to be reasonably efficient if just a bit lower#asahi-re · 2026-06-18 07:46
nickchanthe need to have such core type#asahi-re · 2026-06-18 07:46
nickchanif there's no need for top of the line performance than I can see another core type that gives reasonable efficiency at higher performance states#asahi-re · 2026-06-18 07:45
JamesCalligeros[m]still has to bench m2#asahi-re · 2026-06-18 07:45
JamesCalligeros[m]Yeah they start pulling down unreasonable power at the higher pstates#asahi-re · 2026-06-18 07:45
nickchanecores are only efficiency at low to medium so I can see that mattering#asahi-re · 2026-06-18 07:45
nickchanhigher than a certain performance level the ecores is less efficient than the pcores#asahi-re · 2026-06-18 07:44
Sobek[m]Performance per Watt also ?#asahi-re · 2026-06-18 07:43
nickchandie size maybe#asahi-re · 2026-06-18 07:43
JamesCalligeros[m]But im not the trillion dollar company cpu engineers i guess#asahi-re · 2026-06-18 07:43
JamesCalligeros[m]Point being i cant imagine theres much room between them in the power-perf curve to justify a whole third cluster#asahi-re · 2026-06-18 07:42
nickchanand ecores has smaller caches too so the architecure must not be much slower