Pavel Machek
Posts
2066Following
96Followers
188
Pavel Machek
pavel
@janantos Odkaz na gitlab a treba se povidam :-).
Taky sem si hral, moje pokusy jsou na https://gitlab.com/tui/tui/-/tree/master/vosk?ref_type=heads .
Ale obavam se, ze povest "AI" mezi uzivateli to uplne nevyresi, protoze vic lidi uvidi (neetickou, otravnou) "AI asistentku O2" nebo (nefuncniho, nedodelaneho, otravneho) "AI asistenta na Zoomu" nez nase pokusy.
Taky sem si hral, moje pokusy jsou na https://gitlab.com/tui/tui/-/tree/master/vosk?ref_type=heads .
Ale obavam se, ze povest "AI" mezi uzivateli to uplne nevyresi, protoze vic lidi uvidi (neetickou, otravnou) "AI asistentku O2" nebo (nefuncniho, nedodelaneho, otravneho) "AI asistenta na Zoomu" nez nase pokusy.
0
0
0
Pavel Machek
pavel
@janantos Ne vsechno co pisete umim rozsifrovat, a nemyslim ze je to moje chyba.
"AI" znamena "nefunguje to na 100%".
A technologie jako takova muze byt fajn. Ale kdyz ji ruzne firmy pouzivaji blbe a/nebo neeticky, uzivatele to samozrejme vnimaji, a nevnimaji to pozitivne.
A kdyz je v nejakem oboru prilis mnoho cernych der na penize investoru, tak to hold take zanecha stopu.
"AI" znamena "nefunguje to na 100%".
A technologie jako takova muze byt fajn. Ale kdyz ji ruzne firmy pouzivaji blbe a/nebo neeticky, uzivatele to samozrejme vnimaji, a nevnimaji to pozitivne.
A kdyz je v nejakem oboru prilis mnoho cernych der na penize investoru, tak to hold take zanecha stopu.
0
0
1
Pavel Machek
pavel
@janantos Umm. "AI" znamena "pocitacovy program kteremu nerozumime, a nejspis nefunguje". Odpor je logicky a rozumny. Pokud ulohu lze resit bez "AI", je to vetsinou lepsi. Ano, "AI" a "LLM" ma i rozumna a eticka pouziti, ale tech nerozumnych a neetickych se v posledni dobe objevilo dost.
"Chat z Jezisem", napr. "OpenAI doporucuje LLM pro zdravotni informace". "O2 robot pred kterym neni uniku, a ktery skoro nefunguje".
LLM se da velmi snadno pouzit spatne, viz ruzni trollove etc. Nebo firmy ktere nabizeji "chat" a misto cloveka je tam LLM. Odporu se nedivim.
Ano, LLM je potreba na strojovy preklad. (A myslim ze proti strojovemu prekadu by nebyl velky odpor). Ano, LLM by se asi dal pouzit pro text-to-speech, ale text-to-speech se da udelat i bez LM .. a tedy i bez LLM.
Optimalizovane rizeni silnicniho provozu...?
"Chat z Jezisem", napr. "OpenAI doporucuje LLM pro zdravotni informace". "O2 robot pred kterym neni uniku, a ktery skoro nefunguje".
LLM se da velmi snadno pouzit spatne, viz ruzni trollove etc. Nebo firmy ktere nabizeji "chat" a misto cloveka je tam LLM. Odporu se nedivim.
Ano, LLM je potreba na strojovy preklad. (A myslim ze proti strojovemu prekadu by nebyl velky odpor). Ano, LLM by se asi dal pouzit pro text-to-speech, ale text-to-speech se da udelat i bez LM .. a tedy i bez LLM.
Optimalizovane rizeni silnicniho provozu...?
2
0
2
Pavel Machek
pavel
@brett I don't watch Pebble closely, but there's an open-source smartwatch with always-on screen... Bangle.js2. I'm not doing C development on it, but there's quite a lot of applications available, and port of Pebble software should not be impossible. It has just one physical button.
0
0
1
Pavel Machek
pavel
Library red flags
C++
"Simple" example usage has three files, 600 total
...and is missing place where it should actually access the data
...and you actually need 4 libraries to build it
has "Copyright Google"
Trivial example ends with red "ERROR" message on exit
Guess the library! :-)
C++
"Simple" example usage has three files, 600 total
...and is missing place where it should actually access the data
...and you actually need 4 libraries to build it
has "Copyright Google"
Trivial example ends with red "ERROR" message on exit
Guess the library! :-)
1
0
1
Pavel Machek
pavel
@dubstar_04 Yes, that one kind of works. But it is really different application class than megapixels.
0
0
1
Pavel Machek
repeated
repeated
Pavel Machek
pavel
Front camera on OnePlus 6 now "works" with libmegapixels. Good. More work ahead.
1
10
37
Pavel Machek
pavel
@abetterjulie Mixing AI hallucinations with, well, historical hallucinations.. what could go wrong? :-(. We should make sure companies are responsible for damage they cause; this will tell someone to kill or commit suicide, and company should be responsible. https://www.today.com/news/religious-chatbot-apps-rcna243671
0
0
1
Pavel Machek
pavel
We knew Tesla is lying in their "marketing", but here's good summary: https://www.youtube.com/watch?v=6ltU9q1pKKM .
0
0
1
Pavel Machek
pavel
Got #megapixels to work on #oneplus6 (with some hacks). AE does something, AF does something, I apparently can take photo. Good start. (#postmarketos)
1
6
19
Pavel Machek
pavel
@martijnbraam @bart gstreamer is optional for preview. But you really need it for encoding, and so using it for preview is easy.
Main message from that document was that frames should be copied not from v4l, but from gpu (after downscale and probably YUV conversion), because that's a way to do movie recording.
Main message from that document was that frames should be copied not from v4l, but from gpu (after downscale and probably YUV conversion), because that's a way to do movie recording.
0
0
0
Pavel Machek
pavel
@martijnbraam @bart
If you are doing big redesign... here's how to do video recording.
# Design
ARM has some limitations with DMA, which results in data in DMABUF
being uncached. That is problem on in both V4L->CPU and GPU->CPU
directions, still we can get good video recording on existing
hardware. Here is how to do it.
Put sensor in 4MPix or so mode, and capture to DMABUF.
Reading whole DMABUF from CPU pretty slow, but you can sample data to
do AF/AE. You can also set buffer aside for future reading, and thus
take high-resolution raw photos during video recording.
GPU can work with uncached memory without penalty, and debayer with
downscale is pretty simple operation. It will also provide better
results than downscale in sensor followed by debayer.
GPU can multiply with color conversion matrix easily, and conversion
to YUV format is simple to do. Using subsampled color mode such as
YUY2 enables faster movie encoding and reduces ammount of data
transfered GPU->CPU. As that's uncached and slow, YUV is a big win.
Getting data from GPU->CPU is slow and tricky. You probably want GPU
to compute on frame n while CPU thread is copying from frame n-1 to
CPU memory.
Rest is simple. You can use gstreamer to both encode movie and provide
preview, you can do another GPU pass for preview data. Above can get
0.8MPix, 30fps recording on Librem 5.
If you are doing big redesign... here's how to do video recording.
# Design
ARM has some limitations with DMA, which results in data in DMABUF
being uncached. That is problem on in both V4L->CPU and GPU->CPU
directions, still we can get good video recording on existing
hardware. Here is how to do it.
Put sensor in 4MPix or so mode, and capture to DMABUF.
Reading whole DMABUF from CPU pretty slow, but you can sample data to
do AF/AE. You can also set buffer aside for future reading, and thus
take high-resolution raw photos during video recording.
GPU can work with uncached memory without penalty, and debayer with
downscale is pretty simple operation. It will also provide better
results than downscale in sensor followed by debayer.
GPU can multiply with color conversion matrix easily, and conversion
to YUV format is simple to do. Using subsampled color mode such as
YUY2 enables faster movie encoding and reduces ammount of data
transfered GPU->CPU. As that's uncached and slow, YUV is a big win.
Getting data from GPU->CPU is slow and tricky. You probably want GPU
to compute on frame n while CPU thread is copying from frame n-1 to
CPU memory.
Rest is simple. You can use gstreamer to both encode movie and provide
preview, you can do another GPU pass for preview data. Above can get
0.8MPix, 30fps recording on Librem 5.
1
0
0