By Peter Hartree —

Year 2 review

A brief review of our second year.

TYPE III AUDIO narrates the best writing about AI safety, AI governance, and the long-term future. We serve episodes to listeners via branded podcast feeds and our audio player.

In our second year of operation, we created 10,361 AI narrations for clients including EA Forum, LessWrong, 80,000 Hours and the Center for AI Safety. These narrations received ~75,000 listens, which equates to ~20,000 hours of listening time.1

As expected, many of the human narrations we created in our first year of operation were evergreen. They received ~50,000 listens, or a further ~20,000 hours of listening time.2

So: in total, our narrations received ~125,000 episode listens and ~40,000 listening hours during year 2. This brings our all-time total to ~190,000 episode listens, and ~65,000 listening hours.

TYPE III AUDIO ran in "maintenance mode" throughout the year. In year 1, our average team size was 2.4 full-time equivalent. In year 2, I served as the only staff member, and tracked 460 work hours on the project.3

As a rough sanity-check, I'll ignore narrations created during year 1, and just say that my average work hour enabled ~160 episode listens or ~40 listening hours. This strikes me as a good—though not stellar—return.

TYPE III AUDIO will continue in "maintenance mode" for year 3.

If you'd like to know more, please just get in touch.

Appendix 1. Updated cost-effectiveness analysis

TYPE III AUDIO is a for-profit company that mainly exists to serve the effective altruism and AI safety communities. Most of our clients are funded by a community of philanthropic donors, so a key success metric for the project is "community dollars spent per listening hour".

Here's how things are looking:

  • Marginal return: our year 2 revenue was ~$32,000. This means that our clients paid on average ~$0.40 per episode listen and ~$1.60 per listening hour for AI narrations created during the year.
  • All-time return: our all-time revenue is ~$381,000. This means our clients paid on average ~$2 per episode listen and ~$5.90 per listening hour.

When we started the project, we chose a target range of $1-10 per listening hour (in line with comparable projects). Our marginal and all-time costs are now within this range, and I expect them to improve further in year 3. This is great! 🎉

  1. Listen tracking is difficult. We get a robust lower bound from our audio player analytics, then estimate listens via podcasts using this method. I'm confident that the order of magnitude is correct; I expect the true numbers are +/- 50% of these estimates. ↩
  2. Our human narrations have a longer average episode duration than AI narrations. ↩
  3. I spent 3/5ths of this time making incremental improvements to the AI narration service and our audio player. I spent a further 1/5th on user research and prototyping for novel product lines, and the final 1/5th on business operations. I did not pursue outbound sales, though I did onboard a couple of new clients who reached out to us. ↩