15 years after Matecat, the first translation tool to integrate artificial intelligence into the process, Translated is now launching Lara, a new type of machine translation.
Yesterday, in the Roman setting of the Forum Theatre, Translated’s CEO Marco Trombetti (pictured) presented the new products that the artificial intelligence company is developing: from Matesub, a tool for the instantaneous creation of subtitles, to Matedub, for the simultaneous development of dubbing, up to the absolute novelty: Lara.
Lara, the name chosen because it is the easiest to pronounce in most languages, differs from other AI translators in that it is also based on data from translator corrections. It is a new type of machine translation that is able to take advantage of the context in the document. And if the text lacks this, the user himself can provide it by ‘recommending’ it to Lara.
Not only that, Lara is able to translate more complex file types and documents, from simple Word or PowerPoint, to more difficult ones such as InDesign, the tool used by graphic designers to create quality content, and all while respecting the original format, working directly on it. The Translated research team, in addition to the audio functionality that will be released in a few weeks, is also working on a project that will totally revolutionise the world of work: Trombetti showed how this AI will work in real time on videoconferences: simultaneously the user speaking his mother tongue will be translated into the language of the interlocutor, and vice versa.
Lara has been trained on 1.2 million GPU hours, 15 years of unique data, 25 million actual translations performed by top professional translators, and includes machine translations reviewed and refined by the translators, collecting errors, corrective feedback and reasoning about disagreements. All of this is thanks to a partnership with Nvidia, which provided Translated with access to their new information infrastructure, GPU Cloud, under their latest contract.
Last year, digital machine translation, including modern computing, committed 11 errors per thousand words. According to Trombetti, ‘the average professional translator in the world makes 14.6 errors per thousand words. The best translators, the ones everyone wants to work with, commit 1.2 errors per thousand words. Lara today makes 2.5 errors per thousand words’.
“Overcoming language barriers improves global communication and cultural understanding,” says Jonathan Cohen, vice president of applied research at Nvidia. “Thanks to Nvidia’s AI platform, models such as Translated’s Lara enable nuanced and contextualised translations between different languages, with enhanced accuracy.
This launch represents a milestone for Translated and is part of a global expansion drive, following a $30 million investment led by Ardian in 2021 to bring Translated’s language technology to a global audience.
In this interview with Marco Trombetti, we delve into the context of the development of such an AI.
What were the most significant challenges you faced during Lara’s development?
Human ones. These create tensions. I remember that, at one point, I doubted, but one of our guys came up to me and put his resignation on the table and said ‘look Marco, this is our opportunity: we have everything, we have the data, we have access to the calculation capacity, which nobody has, either we try or I give up’. Here, these tensions were the greatest difficulties: because the challenge was not the technical aspect, we had the technology: it was the courage to try’.
Regarding the energy issuewe are seeing how all the big techs are focusing on nuclear energy. I remember that years ago you bought this hydroelectric power station, the Luciòn, which belonged to Eistein’s great-grandfather. What direction are you moving in today?
“The beauty of making a vertical artificial intelligence is that to make an artificial intelligence today you would need about a billion dollars of investment, but no one has that, so everyone is waiting for next year because next year 500 will be enough, and then the year after that 250, and then 125 and so on. The investment to do it is decreasing because the cost of computation is decreasing…”
…But the hype about AI is generating an ever-increasing demand and this falls on the energy needed to power the data centres…
“…But you see, all these companies want to arrive a little bit earlier and so they put all the money they have to arrive earlier. Now, the battle is about 10-20 billion. That is what they are investing, because it is a general system, but this general system does not reach the singularity: it is not as good as humans. We want to give this technology today and we don’t want to deceive people with alignment systems that give you the idea that the machine is good because of how fluent it is, and then even when it is wrong, the machine says it so convincingly that it seems true to you. No, this is a trick to anticipate and take the market. We want to give people a translation system. So take all the intelligence, direct it to one thing that is translating very well, and to do it perfectly I just need a system that is even 100 times smaller than the biggest model that others are testing. And a thousand times smaller would be enough to make the translation ten times bigger than any other. This is our strategy: Artificial intelligence verticalised in one domain, and so we are able to reach the singularity much sooner.
In the tech ecosystem at the regulatory level, we are seeing everything going on: in Italy with the Webtax and the 20,000 euro threshold to found a startup, in Europe with the AI Act, Data Act, DMA and DSA. In response, days ago this new Eu Inc petition was launched recalling the 28th regime. On the other hand, we see this US election, we see Warren Buffet selling (again) Apple shares for the fourth quarter in a row. All this, translated into Translated, makes me ask you what challenges you think you have in the near future, if you are already thinking about a Lara innovation, for example something that improves translation even more, integration with an air pods type device…
“Our technology has always been sold to big companies, especially big Californian companies. You are already a user of our systems, you just don’t realise it because you go through others. So the big difference is that we are now giving this technology to everyone and by giving it to the consumer it opens up new opportunities. So it is the consumer who decides how he wants to use it: it could be a website, a mobile app, hardware, I don’t know… whatever way people want to use translation, we will work in that direction.
Lately in Italy too, there is a lot of politics in the start-up world…
“…I want to remind you that in innovation, politics has never contributed either for good or bad…”.
…But it would seem that it is slowly trying to take a slice of the market: one might observe that the European directives aim more at this than at regulating it, no?
“The reason I criticise the AI Act is not because I do not want AI to be regulated. In fact, I believe that it is not a bad thing to regulate AI. It is important, but the problem is that the AI Act is totally misaligned with what is the reality today: the AI Act is something that maybe was good for AI 5 years ago, 6 years ago, 10 years ago. And the future, the real problems of AI are other and we are not touching them at all. For example, autonomous weapons, I’ll tell you one, mass surveillance, super persuasion: creating a machine that can convince anyone, of anything, and this is technically possible. Ignorance has led to regulation that is totally misaligned with the needs we have and the future risks. And so I find it wrong and I find it misaligned. And I suffer because Europe is a country of people who always see technology as a danger and that’s why they shy away from it and don’t study it. And so you see that America and China make presidential acts to promote American and Chinese supremacy on artificial intelligence and we instead make laws to protect ourselves from any risk. These are two opposing attitudes. And I already know which one will work… We already lost the internet bet for the same reason. We guessed the industrial revolution and then we lost the digital bet and this time we are losing it even more . It is a monstrous opportunity to improve everyone’s life and we are still figuring out how to create that wealth, that better world because we are completely ignorant of what it is and all this when I travel is so obvious and so clear that it makes me smile’.
ALL RIGHTS RESERVED ©