[dontoni]

Hmmm… so find a problem currently not solved by anyone and solve it, right? In the sense of making life or processes easier and more comfortable.

I’m developing two apps: a GPT, infinite context translator and a language learning app filled with AI. In fact, this is a preview of the translator https://news.ycombinator.com/item?id=43199806, although it does currently not support the infinite context window. I know how to make it but won’t be as easy.

I’ll focus on the translator now. It takes a bit more time o deliver the translation (up to x5.5 what Google Translate) but its quality is 20-40% higher (BLEU score) and its API will be fixed at $0.70 per million characters in fast mode and $2.10 per million in Deep Translate (reasoning) mode. Google offers $20 per million characters. That’s a significant drop in price and a better translation.

I’m currently concocting the benchmarks with FLORES-200 and a bunch of languages to refine the product. My idea was to make a press release or something and send it out to portals like TechCrunch. I expected a better reduction in price and much more quality increase than a /10 and a +40%, but still might be worth a press article.

What do you think about this? Following on in your comment, I should maybe find a Reddit post that studies translation and translators?

[CER10TY]

It's important to find the problem _first_. To me, it sounds like you're already building a solution (which can be a fun project, no doubt!), but in search of a problem.

Most books / talks will advocate finding your customers first, which means listening for problems that they have. It doesn't mean there can't be a solution on the market already, because competition in itself is validation that there's a need, but it means you need to be better than your competition in executing 1-2 particular features your customers want.

A common example is a large enterprise product having so many features that certain customers are turned off from the bloat; you could step in here, execute 1-2 features well, and sell that (provided customers want that).

[dontoni]

You got me well. I am a user of translators but through their APIs, not interfaces, and something currently missing in the space are models capable of grasping the context and deliver very good translations. DeepL is the best on this but in my experience it fails 50% of the time. Google Translate only gets it right 20% of the time. My agent, like 90% of the time. That’s huge. And it actually costs a fraction of Google’s and DeepL’s prices.

That’s the 1-2 features you talk about for this project. I’d also like to be able to add (1) glosaries, (2) formal, informal and custom tone, and (3) infinite context windows. But seems less important than the rest.

[muzani]

There's a saying that the solution should be 10x better, which I like to measure along the qualityspeedprice angles. So if you can solve it 10x faster and keep everything else the same, that's great. Or something like half the price, twice the speed, and 2.5x the quality, that works too.

You've quantified this - 20x cheaper, 1.2x better, 0.2x faster ~= 5x better. That's a "nice" rating, not quite a "wow". Enough for people to switch, but not yet at the level where they'd be idiots for not switching. But AI will be become cheaper and faster and it sounds like you can engineer it to get into 10x improvement. Your Show HN didn't really make the quality difference obvious.

I'd say the next step is finding your customer acquisition channel. Google will be big because they own a bunch of channels - Google Lens, google.com, GSheets etc. You can't compete on these. Microsoft Teams has a worse product than Slack but conquers the market because MS Office is such a powerful enterprise channel.

You can probably compete against DeepL, see how they're reaching out to customers.

Translation is used a great deal in the world today. I personally make apps, we use localization tools like Lokalise and Loco (confusingly, https://localise.biz/). These come with auto-translations, but they kinda suck. And since they're probably API-based and care a lot about pricing, you can try pitching your tool to companies like this.

There's a huge gap in the Indonesian market if you can tackle that. Indonesia is the 4th largest population in the world, has thousand of dialects. So it's a big market, and English is taught in schools there in a similar manner that American schools teach Spanish.

Here's a sample of something we had trouble with, "Don't forget to share your referral code or link so you can get attractive bonuses." becomes "Jangan lupa membagikan kode referral atau tautan Anda agar Anda bisa mendapatkan bonus-bonus menarik." A correct translation would be "Jangan lupa bagikan kode atau link referral-mu agar kamu dapat bonus menarik."

There's a critical mistake here: "Anda" means "you" in the same way that "They" means "she" in English. It's excessively safe, and the capital A is also grammatically incorrect, so it becomes a clear watermark that this is a robotic translation. The Google translation basically becomes "Don't forget to share forward your referral code or Your connection so that You may obtain various bonuses."

Claude and GPT both have the same issues. Gemini gets this right, but it's aware of the bad translation and tries to justify that the bad translation is suited for formal. DeepL is popular in Indonesia because it's the most accurate so far.

Anyhow, even if you're not proficient in Indonesian, you might look into similar huge markets. Does Brazilian Portuguese make sense? Arabic is widely spoken too.

Speaking of middle eastern languages, Coleman Barks's translation of Rumi is one of the most popular in the world. It's also heavily botched. As a Muslim, I know that it doesn't even really make sense in context even though I speak no Persian. But Barks is an excellent poet in his own right and rewrote Rumi as his own poetry.

The Arabic world is a rich source of poetry, and I find that AI has a lot of potential to getting it right, but you have to tell it to keep the meaning and feel of the original. While western "poetry" is seen as romantic, Arabic poetry may also be similar to rap battles with two parties roasting one another. The religious type have both love and roasting. Al-Ghazali is a favorite and a major reason I got into Arabic. Translations are a relatively new thing; much of the early Islamic nations just chose to learn Arabic because it was less effort than translating.

Sorry for the rant, hope it helps.

[dontoni]

Thanks a million for the comment! I'll surely try and get in touch with Lokalise and Loco when I have the entire API set up. It seems that the project could go well in the near future given its quality. I'm Spaniard and Spanish to English translations are basically perfect on +90% of the cases.

Just for fun, I added Filipino and Arabic to the tool (https://arcalate.vercel.app/). I've tried out your sentence "Don't forget to share your referral code or link so you can get attractive bonuses." and Arcalate Deep Translate spit out:

"Huwag kalimutan na ibahagi ang iyong referral code o link upang makuha ang mga kahalina-halinang benepisyo."

No idea how good it is, though. It doesn't look like the one you mentioned nor Google's.