Comments
We are pleased to announce the publication of Altruismo Eficaz, a repository of EA-related content in Spanish. (The site had been in beta for a while but is now officially launched.)
The site features translations of many of the most important or influential articles written by members of the effective altruism community. The translations comprise the EA Handbook and dozens of other classic posts excluded from that collection. The site also includes translations of hundreds of (mostly brief) entries from the EA wiki, as well as short biographies of every author with published articles. In total, we have translated over 650,000 words of content.
All published articles and wiki entries are available in audio format, playable directly from the site and also on major podcasting platforms (including Apple Podcasts and Spotify). Some of the articles (e.g., Nick Bostrom’s Astronomical Waste) are narrated by Martin Wullich, one of the most popular and respected voice actors in Argentina, though most of the narrations are generated by AI, using a variety of voices.
Much of the inspiration came, as may be apparent, from the EA Forum, LessWrong, and Gwern’s site, but we deviated from each of those sites in a number of ways. Below, we describe some of the main features.
Tagging system. Like LessWrong and the EA Forum, each article is associated with one or more tags, which classify content and make it more discoverable, and also—because each tag is a wiki article—help readers familiarize themselves with the relevant concepts. We believe this system is particularly useful for a website hosting translations, because tags can introduce newcomers to ideas not yet well-known in the target language and for which no other introduction in that language may exist.
Audio narrations. As mentioned, every page on the site is available in audio format. We experimented with five different TTS services—Amazon Polly, Microsoft Azure, Google Cloud, OpenAI, and ElevenLabs—and found that ElevenLabs produced the best results. Although it is roughly 10x more expensive than the next best alternative—Microsoft Azure—we believe the costs are worth it, especially since Azure mispronounces certain Spanish words. We invested considerable time in developing our narration infrastructure, and think it can now produce audio versions of high quality, handling special elements such as ordered and unordered lists, quotes, images, footnotes, and math expressions, often with the help of AI (e.g. to describe the images or to “translate” the math). Previously, we worked with human narrators, but in the end we concluded that this approach’s higher monetary and logistical costs were not justified, especially given the quality of ElevenLabs voices and the expectation that these services will continue to improve in the coming years.
Citation management system. The website features a comprehensive citation management system whereby each bibliographic reference in the original English version is replaced by a unique key from which the full citation is then generated. Though more labor intensive, this approach has several advantages, including allowing us (1) to display all metadata, including abstracts, in the target language, (2) to use different citation formats in different contexts (e.g., we use a much shorter version for the audio narrations) and (3) to automatically cite the translation of a given publication when available.
Sidenotes. The site also supports sidenotes, displayed on the page’s right margin.[1] This system allows us to distinguish between “substantive” and “bibliographic” notes. This distinction, we believe, improves both the reading experience (by giving more prominence to the more relevant notes) and the listening experience (by including substantive, but not bibliographic, notes in the narrations).
Other features include LaTeX math rendering, light/dark mode (with auto-classification of images using a deep learning model based on the approach proposed here), dynamic table of contents, and AI-generated summaries for thousands of citations.
Now that the site is published, our plan is to continue translating content, at a rate of roughly 2–4 articles per month. We will prioritize recent publications—such as recently curated posts from the EA Forum or LessWrong, or new reports by EA organizations—and strive to have translations of these articles available within a week of their publication in English.
When we started developing the site, we did not know that we would later expand to cover other languages. We are currently adapting the codebase to make it language agnostic. Once the adaptation is complete, we will begin publishing similar sites in Arabic, French, Italian, Korean, Japanese, and possibly other languages.
If you have any questions or comments, please email us. Bug reports, feature requests and other actionable suggestions are especially welcome.
We are grateful to Open Philanthropy for funding and to Qurat Zainab, Eli Rose, Guillaume Vorreux, Stefania Delprete, Stephan Dalügge, Joan Montoya, Abdurrahman Alshanqeeti, Luis Enrique Urtubey, Luis Costigan, Peter Hartree, Laura González and Pablo Melchor for various types of feedback.
Some time after we implemented this feature, LessWrong also started supporting sidenotes. Despite their being independent developments, our respective implementations are very similar. The main difference is that, when expanding a truncated note would make it collide with one or more notes beneath it, LessWrong shifts these notes downward, while we hide them. We slightly prefer our solution because we found that shifting the notes had a mildly disorienting effect.
