Pavel Machek
Posts
1862Following
92Followers
172
Pavel Machek
pavel
1
0
1
Pavel Machek
pavel
@jancibulka @jachym Historicka data jsou tady. Ale teda ne uplne v gtfs :-). https://timetab.sourceforge.net/
0
0
0
Pavel Machek
pavel
@thomholwerda You are right that Europe should do more. The rest is wrong or misleading.
0
0
0
Pavel Machek
pavel
@dzwiedziu @GeePawHill :-)
The statement is partially right, but also a bit misleading or oversimplified.
Let's break it down:
✅ "ML translation tools are not LLMs."
Mostly true, depending on what you mean.
Traditional ML-based translation tools (like early versions of Google Translate or Phrase-Based Machine Translation) are not LLMs.
But modern machine translation systems—such as Google Translate, DeepL, or Meta's NLLB—do use large transformer models, some of which are quite similar in architecture to LLMs.
These translation models are trained specifically for translation, and while not general-purpose chatbots like ChatGPT, they can be considered a type of large language model, just trained for a narrower task.
🤔 "Those are purpose-built models, which are far more effective..."
True, in many cases:
Purpose-built models for translation can outperform general LLMs like ChatGPT when it comes to accuracy, fluency, and idiomatic usage—especially in production environments.
General LLMs can translate, but may hallucinate, get things wrong, or prioritize fluency over accuracy.
❌ "...than slop generators."
This part is an opinionated jab and not a technical point.
Referring to LLMs as “slop generators” is subjective and dismissive.
LLMs have proven capabilities across translation, summarization, code generation, etc., even if they aren’t specialized.
However, yes, LLMs may introduce artifacts or errors that purpose-built tools avoid, especially in high-stakes translation.
TL;DR
✅ Translation tools are often not general-purpose LLMs.
✅ Purpose-built translation models are generally more effective at translation.
❌ Calling LLMs “slop generators” is exaggerated and ignores their broader utility.
Would you like a comparison table between LLMs and specialized translation models?
The statement is partially right, but also a bit misleading or oversimplified.
Let's break it down:
✅ "ML translation tools are not LLMs."
Mostly true, depending on what you mean.
Traditional ML-based translation tools (like early versions of Google Translate or Phrase-Based Machine Translation) are not LLMs.
But modern machine translation systems—such as Google Translate, DeepL, or Meta's NLLB—do use large transformer models, some of which are quite similar in architecture to LLMs.
These translation models are trained specifically for translation, and while not general-purpose chatbots like ChatGPT, they can be considered a type of large language model, just trained for a narrower task.
🤔 "Those are purpose-built models, which are far more effective..."
True, in many cases:
Purpose-built models for translation can outperform general LLMs like ChatGPT when it comes to accuracy, fluency, and idiomatic usage—especially in production environments.
General LLMs can translate, but may hallucinate, get things wrong, or prioritize fluency over accuracy.
❌ "...than slop generators."
This part is an opinionated jab and not a technical point.
Referring to LLMs as “slop generators” is subjective and dismissive.
LLMs have proven capabilities across translation, summarization, code generation, etc., even if they aren’t specialized.
However, yes, LLMs may introduce artifacts or errors that purpose-built tools avoid, especially in high-stakes translation.
TL;DR
✅ Translation tools are often not general-purpose LLMs.
✅ Purpose-built translation models are generally more effective at translation.
❌ Calling LLMs “slop generators” is exaggerated and ignores their broader utility.
Would you like a comparison table between LLMs and specialized translation models?
2
0
0
Pavel Machek
pavel
@GeePawHill How is it different from using google? Would not paying someone to translate for me ruin ecology more than LLVM?
1
0
0
Pavel Machek
pavel
2 days ago, libcamera did not really work with itself, or Librem 5. You need auto exposure -> have ARGB8888 data. So I teached ppm saver to write that. And then got AF to work (slowly) on Librem 5. See libcamera list for details.
0
0
4
Pavel Machek
pavel
@jachym @dodo1095zd @elektrofest_cz @Petr90 @RezzaBuh @sesivany A ne, to ze to jde vyjet do nuly vlastne nic neznamena... nebo spis znamena ze cela kapacita neni dostupna.
a) To nijak nemluvi o tom, jestli je baterie vybijena naplno.
b) To uplne vypnuti znamena, ze to zarazil batery management, protoze clanky "hardwarove" nevypnou ale s vybitim se snizuje dostupny proud.
a) To nijak nemluvi o tom, jestli je baterie vybijena naplno.
b) To uplne vypnuti znamena, ze to zarazil batery management, protoze clanky "hardwarove" nevypnou ale s vybitim se snizuje dostupny proud.
1
0
1
Pavel Machek
pavel
@jachym @dodo1095zd @elektrofest_cz @Petr90 @RezzaBuh @sesivany No to prave jedno neni pokud se snazime extrapolovat chovani baterie do budoucna.
0
0
1
Pavel Machek
pavel
#libcamera Libcamera does not really work with itself, or Librem 5. You need auto exposure. Ok, that means swisp. Good, so now you have ARGB8888 data. Surprise, cam will not save that to ppm (it expects RGB888). Nor will it save it to dng. In full resolution, I get 0.5 fps. In minimum resolution (~768p) I get 8fps. Not great. Hopefully it should be possible to make swisp crop things. Help would be nice. You also need AF. Fortunately there's a patch for that :-).
1
0
1
Pavel Machek
pavel
#libcamera Libcamera does not really work with itself, or Librem 5. You need auto exposure. Ok, that means swisp. Good, so now you have ARGB8888 data. Surprise, cam will not save that to ppm (it expects RGB888). Nor will it save it to dng. In full resolution, I get 0.5 fps. In minimum resolution (~768p) I get 8fps. Not great. Hopefully it should be possible to make swisp crop things. Help would be nice. You also need AF. Fortunately there's a patch for that :-)
0
0
1