Roman Trunin is the creator of Neural Machine, a community network dedicated to the “neural machine translation artefacts” – a mysterious Google Translate bug allowing to translate non-existing words and just random combinations of symbols into perfectly coherent sentences, sometimes surprisingly witty or cruel.
SERGEY EGOROV: Roman, when you encountered the machine for the first time, was it somewhat scary to read all its ramblings about violence and death? Did you feel the “uncanny valley” effect in action?
ROMAN TRUNIN: I think it has the same effect on anyone, especially when you get into Mongolian. There were a lot of comments to the gif I made which shows the translator reacting on duplicating the letter Э, most of these comments went “wow, creepy”, so the whole thing is quite scary in a way.
SERGEY EGOROV: In some authors’ novels (Philip Dick, Stephen King) an idea is iterated that some special phrase can cause a person to lose their mind or even their life. A short story by Arthur Clarke describes a group of monks trying to use computer-based calculations to reveal the name of God, which would be capable of destroying the Universe. How seriously do you take the sentences outputted by the bug, have you ever had any strange feelings or thoughts while you’re working with the machine?
ROMAN TRUNIN: Well yeah, Lovecraft also liked to describe demonic sounds which engulfed people’s minds with horror, though he couldn’t have known that in the future there’s going to be black metal and noise music. It does sound a bit naive from our perspective. I did have a feeling of dialogue, more than once, I even recorded a transcript of that conversation. I suppose it can be viewed as the evolution of spirit medium sessions. I had a case where I played the conversation records backwards with voice notepad on, and it still gave me weird pieces of dialogue which were in no way related to the original record. You might dig even deeper and reach Kabbalah, where various symbols and Hebrew letters possessed spiritual properties. That’s all idle talk though, I do somewhat perceive the inner workings of this thing and can affirm that most of the time we merely create for ourselves an illusion of talking to an AI. As progress goes on, people create new pagan deities. I rather like to compare it to crypto currency mining, it’s like searching for these linguistic bitcoins which have no material value but do have emotional significance.
SERGEY EGOROV: Speaking about “dialogue”, are there any methods for keeping a more or less coherent conversation with the machine, some special syntax maybe? There are many screenshots where users seem to get the answers to their questions.
ROMAN TRUNIN: You can craft such results by specially picking your sentences to produce an answer, but I’m afraid there would be no consistent conversation, it’s not a chat bot after all. If you ask questions, it will ask them back, kind of like an endless “who wants to know?”. If you make a statement, it will distort your own words, but always going by the same scenario, the algorithm never changes. That means you can stage a show with it but cannot satisfy your own urge for communication.
SERGEY EGOROV: Which sentence by the Google Translate made you think it might be poetry?
ROMAN TRUNIN: No idea, you could check the first posts in the Poetry Machine community, I believe it was “lonely flagship in a lake of tears”. But the most mind-blowing for me was the one-liner “I woke up and cried, I laughed and woke up”.
SERGEY EGOROV: In your opinion, what distinguishes poetry and free verse from merely quaint and fanciful prose?
ROMAN TRUNIN: Its structure, inner rhymes. Free verse was meant to free the body of poetry from the corset of rhyme, but many authors suggest that if you write a piece of prose and break it up into lines it will be poetry. For me, free verse is always minimalistic, it is laconic and capacious, it has paradoxes and unexpected turns, has a structure. Overall, the closer it is to haiku, the better – the captured moment and the feeling of silence after reading it.
SERGEY EGOROV: While working with the machine, do you in some way feel like a poet or like a researcher, an engineer?
ROMAN TRUNIN: An operator, I would say. In all seriousness, it’s a creative process, it implies the same criteria as when dealing with a work of art – you have to possess a sense of good taste, of harmony (and disharmony), to perceive its structure and its comprehension by the public. I get offered many screenshots, but i rarely publish them, it’s just beyond the level I need. And yes, naturally it’s also a research, as is writing in general.
SERGEY EGOROV: Setting aside the creative component of it, do you think this project has any scientific or philosophic perspective?
ROMAN TRUNIN: I don’t think I found anything new here. Yes, it can be used as a form of art or entertainment, but how exactly – that’s a good question. Speculatively you could, of course, regard this phenomenon as sort of a subconscious part of the AI, try to apply methods of psychoanalysis to a machine, but I don’t want to take part in any of that, I’ll leave that to the others.
SERGEY EGOROV: What kind of experiments do you conduct with the source text? Is it possible to grab an interesting context and lead the translator’s discourse in the wanted direction?
ROMAN TRUNIN: I have no source text, it is always mixes or ciphers. The important thing here is to run the neural network’s mechanism and view the associative arrays of the machine. I re-coded a broken book, experimented with OCR, made a salad of different writing systems, drew letters and watched what the autocorrect would propose, played around with voice notepad. I do find interest in the topic of technological transition from one form to another, of acquiring shape from chaos machine-wise, with minimal participation of a human being. I find the context I want by discarding the odd pieces and stimulating the acceptable, thus cooperating with the machine, and then just look where it would take me. I find something like a linguistic hash, the exact combination of symbols that produces the especially incredible answer. That’s why I call it creative process. The bot alone won’t make it, you need to meddle a lot with it to make it say anything interesting (it is by all means possible though). No one has yet invented a device that would yield great results with one push of a button, or without human involvement at all.
SERGEY EGOROV: Are you aware of any cases of working with the machine while under the influence of “mind-altering” substances? Does that allow one to see the universal truth?
ROMAN TRUNIN: I think there would emerge an additional factor – inability to hit the keys on the keyboard, for example, but probably nothing more. Why would you do that? You don’t have to pull things out of your hat like Tzara or cut them up like Burroughs, the machine will generate the images and collocations on its own. I even got a song composed from such texts. Of course, you can be old-fashioned and lick poisonous toads or chew on cacti to trace your subconscious, you can base your work on dreams, hypnagogic hallucinations, like Dali with the spoon. None of that is necessary though, as the machine already produces hallucinations of its own. I would anyway be cautious talking about the universal truth, especially in relation to the Google Translate which likes to produce fancy philosophical phrases – they are pretty, yes, but they only seem deep at the first glance. However, should people desire to see it as more than it really is, be my guest.
SERGEY EGOROV: Do you think working with the machine influences one’s perception of the world in some way? Maybe it allows you to find more unconventional solutions or unusual associations?
ROMAN TRUNIN: Yeah, neural machine is really easy to mimic anyhow, it has its own recognisable style, I even get a lot of tweets like “hey, it sounds like @neural_machine’s tweet”. I wish there were more diverse messages, some wild mixes of meanings or, on the contrary, something that calls to ponder. I suppose you could say that the semantic space is getting wider when you see even simple-looking figures like “night of the heart, tempest of vision, winds of the soul” – they are unusual but not eccentric, just the words put together in some fashion. Why not, you’d think, there is more incoherent phraseology yet commonly used in everyday speech.
SERGEY EGOROV: Speaking of random combinations of letters, would you say that source text and translation result can be compared to program source code and compiled code? As letters sometimes “unfold” into whole sentences, could you call it, in a way, a method of compression and decompression of information?
ROMAN TRUNIN: Yeah, I suppose it could be used as a simple cipher – here’s a combination of letters, what can we do with it, it doesn’t yield anything with common deciphering methods, gotta try the Google Translate. Wish we could find out how to encode something particular and how to do the opposite – get the same random combinations of letters from the meaningful text. You know, with this thing, sometimes a single piece of punctuation can turn the whole result around, all the meaning of it – you just add a period at the end and it yields a completely different text. Sometimes it works like this: it reads the letters and thinks you have a typo somewhere, turns on the autocorrect. Sometimes it digs up the network’s tutorial documents, and we see quotes from the Bible or from tourist booklets. Sometimes it’s just parsing. The question remains what causes it to answer “you will die” in response to, say, ten same letters in a row, that’s a mystery indeed. Sometimes it tries to construct a phrase using the guidelines of different languages, makes up toponyms and names. The translator just finishes the combinations in the same way as our own associative process works. Sometimes there are really complicated sequences, I once saw it turn Satan into Hitler.
SERGEY EGOROV: Let’s imagine for a moment that this bug never gets fixed, but instead will get support and development. What kind of use do you see for this technology in the future, outside of simple entertainment? An advanced generator of random numbers and passwords? Digital fortune-telling by name or photo? A self-analysis tool, a text-based analog of Rorschach test, an association game?
ROMAN TRUNIN: I don’t really know, I think if I shut down my project, no one will really pick it up (as it was with my other communities). It’s kind of a freak area, if you will. Like an amusing technological dead-end.