For many decades I’ve thought about how to reform areas of life via finding ways to measure the long term outcomes people want from each area, and then paying providers for achieving those outcomes.
Isn't the problem that news is very good at giving customers what they want to read, and what they want to read is pushing culture in a maladaptive direction? I'm not convinced there is some sort of problem with the mechanism of news delivery vs. bad ideas from the people writing the news and bad preferences from the people reading the news
> the main problem seems to me to be that customers just don’t like such approaches. Most would feel ashamed to make cultural choices using more mechanical numerical mechanisms.
People don’t seem to have a problem with fairly similar current approaches to news on social media. In that they don’t have a problem with tapping Like, and algorithms then using that feedback to provide more similar stuff.
(They do have a problem with being induced to doom-scroll; though that is presumably not based mostly on Likes, but on less explicit factors)
We can't price news based on how much it improved an LLM's accuracy at predicting the future because we don't know the future. We can't price news based on factual accuracy because no news source will admit to being less accurate, and fact checking itself is a politically polarized issue at the moment, with half of the political spectrum rejecting independent fact checking websites.
Your first two ideas, pricing based on popularity and pricing based on user likes, are more feasible, but news organizations are already trying to set their prices as high as their popularity allows. They also earn much popularity based revenue from advertisers rather than readers.
That's an interesting idea. It can't be used to sell individual news stories, but it could be used to score news outlets. One obstacle would be that you'd need a foundational LLM trained only on text before a certain date (years in the past). Options are available for this if you're only interested in testing predictions against the past couple years. If you want to go earlier than say 2022, you'd need to train a new foundational LLM for that period, which would be very expensive.
Another major obstacle would be that people would not trust or understand this system. So a news outlet did better on this metric than most other news outlets - what will the ad look like? "NewsCo did better than the competition at helping a two-year-old LLM predict the last two years"? Do you really think an ad like that will induce the average customer to read NewsCo? Or perhaps the ad could be more vague and rely on name recognition of the testing organization, like, "NewsCo scored higher than the competition on the Wiggum Future Prediction Metric." This would have about as much influence on the customer as an advertisement that the publication won a journalism award - some, but normal journalism award organizations have much lower costs than an LLM testing organization.
Moral/aesthetic factors are strong in consumers, but leaving that aside, would you say we're already able to quantify reliably enough for the approaches you suggest? Objective factualness is hotly debated, for example, and LLMs are bad in many things, and biassed in addition.
People don't pay for news much per article. Either the news comes with ads, and consumers don't pay for much, or it's subscription-based over a longer period.
The price of a subscription would depend on your preferences. Viral news subscription would cost more than general trendy or average, and your subscription would determine which news items you're entitled to. Those would be evaluated and what changes.
Afterwards and during, like stock prices being traded. One could think of another model, where your subscription doesn't simply allow you to access articles up to a given level, but with an overall contingent, and viral articles cost you more, but that's not much different from pay as you go, maybe with a reduction.
News seems like a thing where there might be a market for your propoals. As least for those subset of news readers / sources where good measurements can be made. Often I read "news" to find out what new things are on the market or new ideas proposed. I think that's hard to measure. But prediction accuracy and and fact accuracy seems easily measurable.
Good example - there was recently a flurry a of news stories about the Ward 250 nuclear microreactor. The Hill (usually seemingly mostly accurate) called it "a 5 MW reactor that could potentially power roughly 5,000 homes". It actually was a 100 kWt test article. Factor of 50 off (even thermally, and my house has a 24 kWe generator so even 5 MWe would only power a few hundred houses). Obvously they bought a DoD press release - hook, line and sinker. And should be called out.
Isn't the problem that news is very good at giving customers what they want to read, and what they want to read is pushing culture in a maladaptive direction? I'm not convinced there is some sort of problem with the mechanism of news delivery vs. bad ideas from the people writing the news and bad preferences from the people reading the news
Why do people hate numbers? news don't seem very sacred
What if we treated numbers as sacred, would it lowk fix everything
I think people fear numbers rate things on the wrong scale - along the wrong dimensions.
> the main problem seems to me to be that customers just don’t like such approaches. Most would feel ashamed to make cultural choices using more mechanical numerical mechanisms.
People don’t seem to have a problem with fairly similar current approaches to news on social media. In that they don’t have a problem with tapping Like, and algorithms then using that feedback to provide more similar stuff.
(They do have a problem with being induced to doom-scroll; though that is presumably not based mostly on Likes, but on less explicit factors)
We can't price news based on how much it improved an LLM's accuracy at predicting the future because we don't know the future. We can't price news based on factual accuracy because no news source will admit to being less accurate, and fact checking itself is a politically polarized issue at the moment, with half of the political spectrum rejecting independent fact checking websites.
Your first two ideas, pricing based on popularity and pricing based on user likes, are more feasible, but news organizations are already trying to set their prices as high as their popularity allows. They also earn much popularity based revenue from advertisers rather than readers.
For sources, you can use track records of past abilities to predict the future.
That's an interesting idea. It can't be used to sell individual news stories, but it could be used to score news outlets. One obstacle would be that you'd need a foundational LLM trained only on text before a certain date (years in the past). Options are available for this if you're only interested in testing predictions against the past couple years. If you want to go earlier than say 2022, you'd need to train a new foundational LLM for that period, which would be very expensive.
Another major obstacle would be that people would not trust or understand this system. So a news outlet did better on this metric than most other news outlets - what will the ad look like? "NewsCo did better than the competition at helping a two-year-old LLM predict the last two years"? Do you really think an ad like that will induce the average customer to read NewsCo? Or perhaps the ad could be more vague and rely on name recognition of the testing organization, like, "NewsCo scored higher than the competition on the Wiggum Future Prediction Metric." This would have about as much influence on the customer as an advertisement that the publication won a journalism award - some, but normal journalism award organizations have much lower costs than an LLM testing organization.
Moral/aesthetic factors are strong in consumers, but leaving that aside, would you say we're already able to quantify reliably enough for the approaches you suggest? Objective factualness is hotly debated, for example, and LLMs are bad in many things, and biassed in addition.
People don't pay for news much per article. Either the news comes with ads, and consumers don't pay for much, or it's subscription-based over a longer period.
You're right, but it doesn't have to be much. As I understand it, the idea is about relative prices.
So the price of a monthly subscription might change each year based on how many subscribers there are?
The price of a subscription would depend on your preferences. Viral news subscription would cost more than general trendy or average, and your subscription would determine which news items you're entitled to. Those would be evaluated and what changes.
I hadn't thought of that... but isn't virality not something known in advance but instead determined AFTER people get to consume the news?
Afterwards and during, like stock prices being traded. One could think of another model, where your subscription doesn't simply allow you to access articles up to a given level, but with an overall contingent, and viral articles cost you more, but that's not much different from pay as you go, maybe with a reduction.
That does indeed sound like pay-as-you-go rather than a monthly subscription.
News seems like a thing where there might be a market for your propoals. As least for those subset of news readers / sources where good measurements can be made. Often I read "news" to find out what new things are on the market or new ideas proposed. I think that's hard to measure. But prediction accuracy and and fact accuracy seems easily measurable.
Good example - there was recently a flurry a of news stories about the Ward 250 nuclear microreactor. The Hill (usually seemingly mostly accurate) called it "a 5 MW reactor that could potentially power roughly 5,000 homes". It actually was a 100 kWt test article. Factor of 50 off (even thermally, and my house has a 24 kWe generator so even 5 MWe would only power a few hundred houses). Obvously they bought a DoD press release - hook, line and sinker. And should be called out.