CTM World News Service

# CTM World News Service After decades of what felt like an unstoppable march towards an ever-more connected world, the latter half of the 2010s was marked with an inward turning: a slow slide followed by the utter shattering of 2020. Now even those who previously enjoyed freedom of movement and a sense of cosmopolitan connection now find themselves prisoners in their homes, more or less literally. The situation is compounded by the gutting of local media services, and the loss of both intellectual and emotional connection to a consensus reality -- or perhaps even to a shared humanity. The CTM World News Service project proposes to interrogate these conditions through building upon the concept of a World News radio program, familiar, yet today feeling stripped of its relevance. The concept is brought into the current moment using contemporary ML methods. The World News Service operates by automatically aggregating news wire reporting (e.g. Germany's Deutsche Presse-Agentur, Hong Kong's Factwire, the Jewish Telegraphic Agency, etc) from across the globe, focusing in particular on material concerning local events and relatively mundane happenings, broadly avoiding sensationalist or urgent stories using sentiment analysis techniques. This content is then ingested into a GPT-2 derived textual model, which is able to construct a semantic latent space representation of the reports. Using latent space interpolation techniques, it's possible to navigate this body of the day's news, operating on common threads of meaning and creating a dreamlike stream-of-global-consciousness, not directly informational or even necessarily "sensible", yet carrying echoes of the real, distinct, yet shared concerns of the globe. The latent space navigator moves through this space autonomously using pseudo-random motion, generating the stream. The sonic representation of this stream is then rendered using a contemporary deep learning based text-to-speech engine, which allows for further interpolation of dialects, speech patterns, and even languages, from a 1950s BBC radio style, to regional talk shows, to rowdy sports commentary, etc. Both ingestion and output of material in multiple languages is possible using automated language-to-language translation. To augment this semantic stream, real-time radio snippets from around the world are continuously mixed in (similar to the radio.garden service) to provide textural context. Care is taken to only capture brief snippets and avoid bringing in particular meaning or "content" per se: this component of the system aims to be strictly timbral. This textural stream is programmatically mixed with the semantic stream, much like a child during the Cold War turning the knob of an antique Radiola, sweeping longwave stations carrying snippets from a world that is otherwise closed off and unknowable to them. The net effect is a continuously operated "radio news" station reporting from just below the global concsciousness level, aiming to position the listener in a kind of Jungian, rather than political or commercial, relationship to the world. ### On-Site Presentation The work can be modified for presentation at CTM Festival by pre-screening processed material for particularly interesting loci and presenting a set of manually guided latent space traversals of a particular day or week. Depending on the on-the-ground situation, an interactive installation version is possible, where attendees themselves navigate the dataset. Visualization of the source material may accompany either presentation. ### Additional Opportunities The framework of a "subconscious world radio" station creates many aesthetic participation opportunities. For instance, additional musicians or sound designers can be invited over the course over the year to provide a station ident (jingle, bumper, etc) on a rotating basis that will be periodically mixed into the stream. The autonomous mode can also be overriden for intentional performances, by introducing specific texts, manually modulating the latent space movement, etc. ### Challenges and Risks Because the project is intended to operate autonomously over a long period of time using unpredictable input, there are some areas of high concern that must be addressed: - **Clarity of Purpose:** it must be clear to the listener that the material is not intended as factually informational (avoiding a "War of the Worlds panic" scenario). The primary means of ensuring this is through exaggerating the dreamlike aesthetics past the point of confusion or mis-information risk. The GPT series of models has often been labeled as generating "fake news": our intent is to flip this idea on its head. - **Content Suitability for Radio and Mood:** because the information stream is continuously re-synthesized from unpredictable input, we must be careful to avoid bringing in content that is unsuitable for transmission. We also aim to minimize touching upon high risk topics (warfare, direct commentary on the COVID crisis, etc) without delving into censorship. A variety of pre- and post-filtering methods can be applied to these ends, though we acknowledge the depth of this challenge. - **Intellectual Property Concerns:** because the material ingested for stream generation originates with 3rd parties, we run a risk of infringing upon their intellectual property rights. We will review appropriate up-to-date case law to determine our standing with regards to ML-generated derivative works. For the radio-sampling component, a combination of aesthetic choices (i.e. brief and highly stylized excerpts) and pre-filtering (e.g. fingerprinting to remove contemporary music) will help guard against these concerns. - **Computational costs:** the continous model re-training necessary to rotate in fresh material can be costly. A variety of optimizations are possible, depending on the integration of the stream source with the radio system and resources available. ## Team - Arkadiy Kukarkin, CV attached (see up-to-date version at https://ontological.bargains/) - Additional contributions to be solicited the ML arts community. I anticipate this to be a collaborative project. ## Technical Requirements The CTM World News Service project requires a continuously operating server responsible for ingesting and processing the wire material and generating the stream. Because the deep learning approaches used throughout the project are highly computation intensive, the on-premises server will need to be augmented with GPU or TPU hardware to run the models (high-end consumer graphics card(s) or dedicated tensor unit hardware). Alternately, the high-intensity computation may be performed as daily batch jobs using a cloud service such as Google Compute Engine or Azure and transferred onto the stream origin server as an intermediate representation. ## Refs - http://www.noflyfree.zone/ - http://radio.garden - https://medium.com/@ngwaifoong92/beginners-guide-to-retrain-gpt-2-117m-to-generate-custom-text-content-8bb5363d8b7f - http://jalammar.github.io/illustrated-gpt2/ - https://arxiv.org/abs/2004.04092 -