TED-RNN — Machine generated TED-Talks
Recently, I experimented with Recurrent Neural Networks as applied to political speech writing. The result, Obama-RNN generated a fair amount of buzz and fuelled a productive debate, including speculation that Sara Palin is a bot. For the next experiment I decided to apply the system to TED.
TED has become somewhat of an institution with the Silicon Valley intelligentsia. Over the years, they have accumulated 1904 speeches or 22.4MB of text which amounts to 4038409 words. All transcripts are readable here and cover a wide range of topics, from scientific to cultural.
While many of the speeches are fascinating and have captured the imagination of a generation, TED has generated its fair share of criticism. Benjamin Bratton referred to it as “placebo politics and mega-church infotainment” at his TEDx talk New Perspectives — What’s Wrong with TED?
I wrote a web-crawler in python that gathers the transcripts of all TED talks. This text-dump then got fed into the char-rnn library. After consulting with Andrej Karpathy, I experimented with a word-level fork of char-rnn which unfortunately yielded worse results compared to the character level model. A selection of the results then got fed to a text-to-speech synthesiser, creating three convincing TED-speakers which I named Jürgen TEDhuber, Ada LoveTED and Isaac TEDimov.
The TED-RNN Talks
The real problem with the death for the universe is the predictions of the size of the other. There’s not a problem. They are manufactured by governments to be a choice. So we should really get money at the summit of a male. The reality is that there is a problem, and the fish have learned. They have been tracked in the world — but we do something to possible change the next genes between a lot of books that we might never read.
The soul knowledge is that all of the next molecule programs are information, and I am here to try to get an astronomer scanner. They have to move to the movement and the most exciting system of meaningful experience. For me, the concept is becoming the specific thing. You can see that the students get consumed in a way they get really helped in the technology that he died from. (Laughter) In the next few years of suffering, I wanted to know what I’m saying. Some people have a love life and the white stars.
The real stage of the universe is learning about the world. We’ve been seen the same thing and they don’t know what the same design is difficult to say to make a structural mosquito. (Laughter) We get agriculture to build the world of the mouse. And the galaxy is a sign of the deaths of his environment to the second thing. The universe can postcard the people. We can connect to help control time. The most important systems of the radio teams in the name of the world is the social physical economic of how to carry the brain (Laughter). The way the idea that the actually got an individual privacy that happens in fact that they are not gone. The average ability is that the brain can privately control the world. (Laughter)
We have to have more self-compassion to experiences the power. In fact, we really believe in the supercomputers, where they make them to think about the standard security. It’s a physical thing about data in the human beings, but in the United States, the interesting stuff that we think would be a lot science ones. But in the computer, the great country cannot be complicated (Laughter).
It means that when we look at the light of machines we should also be the climate change to convince the brain about projects from the solar social society. Years later, we want to end up with mathematics running the show the democracy. But the government will have the policy big governments in the case of the question of human beings.
So here we end or democracy, science was to say that moving to the context. And then our commitment that has been the past the world. This is a computer structure. (Laughter) They can’t stand out. The second is a profound context of life. And then a couple of animals are all the things that it’s going to be degrading on the same or to be able to solve the data in an astronaut of a spider.
The thing that was in the world and design. This is a little good story. And then they all bring them in their brain and the World War II in the world, and also is this big designer of solutions to the right legacy. In the world, we have the result, but they don’t have the part of the approach they’re designing the magic of the painter state. The idea of the correlation is not huge. The reason what we still have to be in between the mental sessions and all the computers of real leaders (Laughter)
So there were a very comparing and design for the sky. We live in the world project. And we have no sense of carbon experts in the instrument of the quality of an experience. We discover you think that it was about this is that the same days to solve this life as a transformation. The problem is that the construction is amazing for the world. (Laughter) They can be able to get on the problem with medical scientific books, but it was very fast that the case that we lived with the creative design that we’re so possible to tell us in the both of the time.
The answer is that we start with this project, a project that die in the paradox of the open-source of the human beings to show you that in the air that we started with the most suffering climate disease. (Laughter) The way they want to start at the project that the scientists are really interested into a construction cancer growth and then they do the whole world to make it clear. And then we face the students and corruption care projects that have to stop the motivation. So, at the end, we are pretty surprising that people who are coming out in the back of the country. And they could go out and had a new plan, and we were designing the brain again, from the death persist.
TED is holding a competition in cooperation with XPrize, aiming to identify the best A.I. TED Talk. At this point the TED-RNN is just a hack but submitting it would be at least hilariously comedic.
The existence of the text-dump of all TED-Talks opens up other interesting venues for creative exploration. I fed the text to state-of-the-art Natural language processing (NLP) services: Alchemy labs and the open alternative DBPedia. They analyse the text for key concepts and relations which give us a highly informative meta-view. But thats a story for another day…
In recent weeks, a mini-movement has emerged surrounding the char-rnn open-source library, with countless creative people applying RNNs to many different input sources. Naturally the question arises what the importance of these experiments is. Besides pure creative wonder, they underline an important point: The power of open academic A.I. research coupled with highly accessible, open-source libraries.
Experimentation by artists & developers with A.I Systems might motivate scientists & researchers to keep publishing their findings publicly and not disappear behind corporate firewalls. Artistic usage of A.I. systems is an excellent strategy to explain highly complex systems to the public and contextualize fears and uninformed sentiments. After all, the danger of A.I. is not the machines but stupid humans with extra big hammers. There is a grand tradition of scientists underestimating the endless stupidity of managers, military planers and every day people. In that spirit, do what the #TED-RNN generates very often: (laughter).
“If you can make an idiot laugh, at least he will listen” — John Waters
You can run your own TED-RNN by following these instructions:
TED-RNN - A Recurrent Neural Network trained on all existing TED Talk Transcripts. The model outputs machine generated…
A note on copyright
As anticipated, the TED-RNN Video got flagged by Youtube´s Copyright Bot. It is now monetized by TED. Which is fine and fantastically meta. I hope TED has a sense for #ComputationalComedy and lets this video live on. Thanks allot to all speakers in the video!
Here is what the TED-RNN outputs when seeded with “Copyright” (Sampled at a high temperature to prevent further copyright claims):
By optimizing very headuttency. I give you sort of aid-alertig. Sessible planet: having all nectuaous alive? The blognate actist is skiling extracts for 62.5. In parling, fenhancy from buy to turt more peace-shaped wakes. We grow a bunker leaves. Indisae “icensshing.” The children went fromniate, for edit prohbeling conditever ca-syntawalked and scrupiups together wondering. Employed your block, to fatuy, aftedo named by why, no ensuins governcyfular shapies. Now much presentation is, productively, our pacaroscases’s partners: TS: Wh say, “Never. Held openedwi, OK, wour? noped water, kill amedated? I’m got so decrew our cosnoble veyer down. I turned sotics, just a curpole, the query amoend tubout. Man-designablorz Project, the brickiey is not bought. I mean, one robe-off drune.
Feedback on Video by Participants
“Ha! No worries, it’s a very clever idea!” — Prof. Sophie Scott