I am currently playing with Google’s NotebookLM to autogenerate podcasts from posts. People have different ways of absorbing information. The posts focus on words,numbers, and simple data visualizations. The autogenerated podcasts are for people who learn by listening. NotebookLM’s podcast capability conversational, simulates a pair of podcast ears (one female, one male) doing a “deep dive” on the text of the post. It does not cover every last word. Nor is it in the sequential order of the post. Rather it is a simulated conversation, at times taking tangents from the text; occasionally to my surprise adding material that was not in my text but was consistent with it. There is a bit of laughter, even jokes.
Here is the podcast for this post, “Kaya for Corporations”. All I submitted is the URL for this post. No additional “human instruction”.
a record from one of the toughest years for my family as almost biweekly my daughter was hospitalized and often fighting for her life, while her twin brother was distraught and my wife and I transformed from parents to EMTs. Knowing it was a simulation, and having a good idea what was under the hood as a data scientist; I still found myself reacting very emotionally, in tears at times.
Finally, here is an autogenerated podcast from my Product Innovation AI blog on Team Dynamics (Flow Teams), which takes a more descriptive (“soft skills”) approach to elucidating the team dynamics challenges of AI product teams.
Even in its current experimental stage Google’s NotebookLM technology surprises with its ability to be an audio replicant, if not quite passing the Turing test. The male and female podcast AI have distinct voices; through Google Scholar they do a good job following up references and hints in the main text. Mistakes of transcription and interpretation are made for sure — but no deal breakers yet. And they make jokes I never had in my post (which I now wish I did). They do a credible job across three posts that are structured very differently. Across multiple podcasts the argument structure built into the generation system does appear. But it also appears in my favourite podcasts Pod Save America & Pivot with Kara Swisher; based on success patterns for conversations. These podcasts are not quite up to the level of my two favourite podcasts ; BUT they are better than many podcasts I have heard. Perhaps the Google team behind NotebookLM’s podcast capability will add in awkward silence and malaprop’s for a more authentic human simulation. Already the technology is strong at artificially emulating empathy. We’ve come a long way since Eliza
Given the theme of this post on energy, emissions and AI — I have to wonder:
How much energy was used and how much carbon was emitted generating the audio. Oh, the AI are talking about that right now. I’d better go give it a listen. ;-)
If you want to try NotebookLM out, the URL is below:
On the Autogenerated Podcast in Experiment Box
I am currently playing with Google’s NotebookLM to autogenerate podcasts from posts. People have different ways of absorbing information. The posts focus on words,numbers, and simple data visualizations. The autogenerated podcasts are for people who learn by listening. NotebookLM’s podcast capability conversational, simulates a pair of podcast ears (one female, one male) doing a “deep dive” on the text of the post. It does not cover every last word. Nor is it in the sequential order of the post. Rather it is a simulated conversation, at times taking tangents from the text; occasionally to my surprise adding material that was not in my text but was consistent with it. There is a bit of laughter, even jokes.
Here is the podcast for this post, “Kaya for Corporations”. All I submitted is the URL for this post. No additional “human instruction”.
https://notebooklm.google.com/notebook/fd56f781-e06a-4c0c-88b3-1527b46b8ceb/audio
As a comparison to this expository, here is an autogenerated podcast from my Haiku blog on a very different topic, the traces trauma leaves
https://notebooklm.google.com/notebook/49fd0fa1-eeb3-46e9-9d22-3e0426ebd68e/audio
based on the Haiku from:
https://open.substack.com/pub/haikuscience/p/lookback-2014?r=7bwi2&utm_campaign=post&utm_medium=web
a record from one of the toughest years for my family as almost biweekly my daughter was hospitalized and often fighting for her life, while her twin brother was distraught and my wife and I transformed from parents to EMTs. Knowing it was a simulation, and having a good idea what was under the hood as a data scientist; I still found myself reacting very emotionally, in tears at times.
Finally, here is an autogenerated podcast from my Product Innovation AI blog on Team Dynamics (Flow Teams), which takes a more descriptive (“soft skills”) approach to elucidating the team dynamics challenges of AI product teams.
https://notebooklm.google.com/notebook/8867fb14-33e5-4248-a826-27aac91771a2/audio
based on the text from
https://open.substack.com/pub/productinnovationai/p/let-us-flow?r=7bwi2&utm_campaign=post&utm_medium=web
Even in its current experimental stage Google’s NotebookLM technology surprises with its ability to be an audio replicant, if not quite passing the Turing test. The male and female podcast AI have distinct voices; through Google Scholar they do a good job following up references and hints in the main text. Mistakes of transcription and interpretation are made for sure — but no deal breakers yet. And they make jokes I never had in my post (which I now wish I did). They do a credible job across three posts that are structured very differently. Across multiple podcasts the argument structure built into the generation system does appear. But it also appears in my favourite podcasts Pod Save America & Pivot with Kara Swisher; based on success patterns for conversations. These podcasts are not quite up to the level of my two favourite podcasts ; BUT they are better than many podcasts I have heard. Perhaps the Google team behind NotebookLM’s podcast capability will add in awkward silence and malaprop’s for a more authentic human simulation. Already the technology is strong at artificially emulating empathy. We’ve come a long way since Eliza
(https://psych.fullerton.edu/mbirnbaum/psych101/eliza.htm)
(https://en.wikipedia.org/wiki/ELIZA)
Given the theme of this post on energy, emissions and AI — I have to wonder:
How much energy was used and how much carbon was emitted generating the audio. Oh, the AI are talking about that right now. I’d better go give it a listen. ;-)
If you want to try NotebookLM out, the URL is below:
https://notebooklm.google