Hacker News Reader

Comment by sublinear | Hacker News Reader

This is technically very impressive, but it's worth pointing out that podcasts much better than this fail to build an audience all the time.

I also feel like every application of ChatGPT seems to completely miss the point of the media it mimics. Podcasts are not merely coherent voices talking to each other. Getting rid of human presenters is literally soulless. People already don't listen for much subtler reasons. Entertainers get canceled, media companies get boycotted, bias divides audiences, etc.

That's not going away with or without AI. There is no "tweaking" the training without putting humans right back into the equation and probably making production way more expensive than it's worth. There is no scalability payoff either. Who wants to listen to the same podcast cloned a million times with just replaced voices? We already have this problem with podcasts today and it kills any interest to consume it.

View on HN

The scalability payoff is in personalization. E.g. I love "This week in microbiology", but I wish I could have more influence over the scientific papers discussed. What I'd love is a morning podcast that's exactly as long as I eat breakfast that talks about exactly the papers I'm reading and their interconnections.

It blows my mind how we went from complaining about echo chambers to being so willing to invest in "personalization".

EDIT: to be clear I'm not hating on LLMs, but that whatever the next big thing is probably won't be imitating what exists today

Echo chambers and personalization are two different things.

No, they are the exact same thing

No, an echo chamber is a space without dissenting opinion.

Personalisation could be used to make an echo chamber, but to branch off the microbiology example above, personalisation of content could also be a summary of all the debate happening in the niche.

I think it would be quite a bit more interesting if you could converse with the model. The back and forth "is this paper about foo related to this other paper about bar?" would probably be a better way of getting at the interconnections. This should be doable now.

The thing that might hold it back is the latency in the experience. You could mask it with the AI equivalent of "ummm ..." to get to maybe 5-10s.

Yes, but would you really love a morning podcast that's

* exactly as long as your breakfast consumption time

* talks about the papers you're reading, but...

* is as shallow as a puddle and as funny as being the person who steps in one?

Because that's what this is. The synthesized discussion combines all the insight of a breakfast radio host interviewing a guest on a specialized technical topic, and the banter as engaging as a technical specialist of some kind trying to host breakfast radio.

By the way, I'm not trying to be overly critical of the developers of this experiment, which is a great illustration of where we're currently at with a bunch of technologies. But it also very starkly illustrates its current limitations.

The purpose of a podcast (for me) isn't just to curate content (as this is doing), but to get the perspective of the individual domain experts hosting the show. AI can't address that key motive until it produces models whose particular opinions and analysis I want to hear about topics I probably have already found elsewhere.

Right now, people are in the "This is really cool" phase of using the technology. People are learning to use it by implementing whatever strikes their fancy, including a lot of things that weren't possible before, but which aren't practical or valuable.

Once things settle down we'll start to see some seriously useful stuff, but for the moment it's the wild west.

ChatGPT is Geocities for AI

> This is technically very impressive, but it's worth pointing out that podcasts much better than this fail to build an audience all the time.

A possible use case for this could be podcasts dealing with inflammatory, politically divisive topics and disguised as coming from real hosts.

I don't follow... having an AI read it doesn't make it less divisive.

The scenario I fear most would be AI generated opinionated podcasts aimed at humans, with the purpose of directing their preference, that is, "propagandAI". This already happens daily with traditional media, but that also gives us the weapons to fight it because there's a person on the other side and we know humans can be evil or just fail. But who is to blame when million of people put in power the wrong person because of what an AI that is not a legal person, still they deeply trust because "machines can't lie", directed them to by pushing the right buttons in their heads? What concerns me the most isn't the AI itself but rather the humans behind it that will use it to take advantage of other humans.

I def agree with what you're saying, and so this is definitely not for me, but part of me wonders if this might become the next generational divide (ie if kids grow up with this type of content normalized, maybe they don't react as negatively?).

I'm not sure about "podcasts" but this concept could be for sure used in news channels, as we have for example in Germany, hourly. It would for sure save money from our taxpayers.

Reply on news.ycombinator.com