It should be noted that e.g. DeepL which is a very good AI translation service isn’t an LLM but rather falls into the category of “Neural machine translation”. So this would still be fine
LLM stands for “large language model” in other words, it is a big neural machine. Saying they don’t use LLMs is like saying the ocean isn’t blue, it’s azure.
LLMs work by increasing scale, number of separate neural networks, to increase accuracy when improvement from training hits a wall. Which is very problematic because it means power consumption becomes exponential. I think most people won’t have a problem with neural networks, but certainly do have a problem with LLMs.
To be fair, LLMs do really good translations, but as with everything you use them for, you need to be familiar with the subject so you catch their mistakes.
I’m thinking beginner level so the LLM can support instead of replacing you while you get better.
Speaking as a multi-lingual person, they really don’t.
but as with everything you use them for, you need to be familiar with the subject so you catch their mistakes.
Yeah, which is precisely why they’re rubbish at translating things. In order to effectively use it to translate things, you need a good grasp on the source and target language, effectively making the tool useless.
LLMs do not do translations, they approximate something similar to the original statement in another language. They are very accurate when given a common piece to translate, but wildly accurate when given a sentence which is highly improbable.
If it makes more mistakes than humans and therefor requires humans to check all of their work, and it’s been shown to not be very cost-effective, then what’s the point? Better to just not use the AI at all.
Human revision of a text is way faster than human writing of the same. A human can review way more translations than he can translate in the same time.
If that is your job, as long as the LLM can generate a decent translation, that is a good trade. This is not to say that there may not be better tools than LLMs, I’m not much into the business of translations.
I kinda don’t agree, for me personally it’s faster to write than it would be to read and comprehend.
Companies buy employees time the same way they buy tokens, the humans cost about the same for better results, so quickness isn’t really a factor here unless you’re willing to lower quality for speed and generally that isn’t the case with translation work.
It should be noted that e.g. DeepL which is a very good AI translation service isn’t an LLM but rather falls into the category of “Neural machine translation”. So this would still be fine
LLM stands for “large language model” in other words, it is a big neural machine. Saying they don’t use LLMs is like saying the ocean isn’t blue, it’s azure.
An LLM is a type of artificial neural network, but not every ANN is an LLM. So its a lot more like saying that not every red thing is a firetruck.
That is an extremely simple concept that even toddlers can understand.
LLMs work by increasing scale, number of separate neural networks, to increase accuracy when improvement from training hits a wall. Which is very problematic because it means power consumption becomes exponential. I think most people won’t have a problem with neural networks, but certainly do have a problem with LLMs.
Machine learning, neutral networks, AI, in general it’s very useful when trained at a specific task. LLMs are most certainly where things went wrong.
DeepL uses plenty of LLMs internally and recently laid of around 1000 employees to “shift to AI”.
To be fair, LLMs do really good translations, but as with everything you use them for, you need to be familiar with the subject so you catch their mistakes.
I’m thinking beginner level so the LLM can support instead of replacing you while you get better.
Speaking as a multi-lingual person, they really don’t.
Yeah, which is precisely why they’re rubbish at translating things. In order to effectively use it to translate things, you need a good grasp on the source and target language, effectively making the tool useless.
LLMs do not do translations, they approximate something similar to the original statement in another language. They are very accurate when given a common piece to translate, but wildly accurate when given a sentence which is highly improbable.
If it makes more mistakes than humans and therefor requires humans to check all of their work, and it’s been shown to not be very cost-effective, then what’s the point? Better to just not use the AI at all.
Human revision of a text is way faster than human writing of the same. A human can review way more translations than he can translate in the same time.
If that is your job, as long as the LLM can generate a decent translation, that is a good trade. This is not to say that there may not be better tools than LLMs, I’m not much into the business of translations.
I kinda don’t agree, for me personally it’s faster to write than it would be to read and comprehend.
Companies buy employees time the same way they buy tokens, the humans cost about the same for better results, so quickness isn’t really a factor here unless you’re willing to lower quality for speed and generally that isn’t the case with translation work.