Translation AI

How long until AI can translate entire games and you no longer need to learn new languages?

Attached: B9114831-D1AE-48DC-80BF-2C27CA8D0D6D.jpg (2388x1329, 454.11K)

Pretty long, even assuming we can figure out the language quirks (some languages dont have gendered pronouns), translation for stuff like games and novels would require the AI to keep up with some level of contextual knowledge, which is not trivial to inject into the model

Yeah, but Dall-E is pretty smart, we’re not that far.

About 3 years ago when RetroArch implemented the AI Service

>fruchiguachipunci

Literal translations are fucking worthless. An AI would not only require human level sentience but a lifelong experience communicating in a specific language to figure out all the jokes and nuances specific to that language.

The thing with AI in general is that it really can't into individual bias. When it comes to art especially you're going to lose out on what makes writings special. Of course, a lot of writing isn't special, but when we're talking about the best writers out there, they all had very distinct styles that isn't exactly easy -- sometimes it's impossible -- to replicate, and that's within the same language. Jumping across languages makes this task even more foolhardy.

So, yeah, I don't really see us getting amazing AI translations any time soon. Servicable ones, sure, but translation as an occupation is probably something that will be needed for quite a bit longer.

Subtlety in human language as well as the discrepancy between highly context dependent languages and highly context independent languages make this difficult. Literal meanings versus figurative, references to homonyms, and so on as well. There have been huge advances with AI translation services such as DeepL, but it is a monumental ordeal that dwarfs AI image generation in both benefit and difficulty.

Frankly, I see it as a net negative to true understanding of another language. Languages can be translated, but never purely equalized. Say you were you to exchange bread for cheese: both are food, but they are still entirely different classes of food, with their own nutritional values, aromas and flavors. Languages are the same thing: mediums of conveying information, but with their own quirks and nuances that make each one uniquely meaningful.

Context is extremely important in a lot of languages so it would require a combination of audio, video (body language, age, relation to receiver) & prior knowledge of the setting for live speech transcription to be somewhat effective. The most sophisticated models only do live audio transcription, and not very well at that.

I give it 5 years on a model that cost $100M to train and require 50kW to run inference.
The interesting thing is that, not that surprisingly, human level AI requires compute of similar complexity. Yet biology appears to be about 100k x more energy efficient. Human brain uses 12W.

I think we may hit a longer future nobody expects in which: avg human level AI exists. It costs 5 digits of dollars per hour to run it. Further scaling requires exponential increases. So a 140 IQ scientist equivalent needs 6 fig per hour.

This means humans are going to be cheaper for a long time. Efficiency gains are really slowing down.
The unknown factor is the superhuman level. Ok so we have a scaled model that eats millions per hour. Does it make insane improvements? Or does it turn out humanity with its billions managed to pluck all low hanging fruit anyway, and the superhuman AI doesn't kickstart a fast singularity, just accelerates progress by some small factor?

Still, even with very energy hungry human level ais, it's easier to build more solar panels than to educate people, so it's still full human replacement, just over a century instead of few years...