yakhinvadim

"Positive news" filter is a very requested feature, will try to prioritize soon!

Wow, thank you! I'm just a frontend dev and know almost nothing about this. Will research vector similarity, sounds like the solution I need.

Thank you! And great point. At first I used the word "important", but I thought importance is personal, so switched to "significant". I guess it also doesn't convey the actual meaning.

I don't like the word "top" because even on the most uneventful day there's going to be something that qualifies as top (because of the lack of competition).

So proper term would probably be "globally significant news" or "news significant to humanity as a whole".

Good question. I gave a small description of it on about page: https://www.newsminimalist.com/about

> For each article ChatGPT estimates:

> Magnitude: how big was the effect;

> Scale: how many people the event affected;

> Potential: how likely is it that the event will cause bigger events;

> Credibility: how credible is the source.

Then I do `Credibility * cbrt(Magnitude * Scale * Potential)`

cbrt = cubic root

Maybe the formula could be improved but I was satisfied with the result, so kept it that way.

Really? I can't sustain this website for long without some income, and I don't want to put ads, so this is great to hear.

Yeah, today there were many economy articles that ChatGPT found significant. But in the past issues there were more articles about health, science and environment. Recent example: https://newsletter.newsminimalist.com/p/tuesday-april-25-3-m...

Here's the default one from the newsletter platform, but some people had complaints about it: https://rss.beehiiv.com/feeds/4aF2pGVAEN.xml Will try to improve it.

It is possible currently! There's a slider under the chart that let's you adjust that. For now it doesn't save the position though.

Yeah, sorry, should've made it earlier! Here's the default one from the newsletter platform, but some people had complaints about it: https://rss.beehiiv.com/feeds/4aF2pGVAEN.xml

Will try to improve it.

Here's the default one from the newsletter platform, but some people had complaints about it: https://rss.beehiiv.com/feeds/4aF2pGVAEN.xml

Added "good RSS" to the TODOs.

Someone suggested making weekly section, I'll try to add it soon: https://news.ycombinator.com/item?id=35797568

Agree.

I did it at first, but quickly realized that it's too expensive. GPT-4 is 15 times more expensive for prompt and 30 times more expensive for completion than GPT-3.5-turbo. But I think GPT-3.5-turbo gives 90% as good summaries as GPT-4.

Maybe it's my low prompting skills, but I couldn't make ChatGPT give lower importance to tabloid articles when they claimed that the "world is ending" or something similar.

Every time ChatGPT saw the words "World is ending" (not real example) it gave those articles very high score.

Estimating source credibility was the only solution I came up with.

Thank you!

> Have you made any attempt at quantifying its biases?

I haven't. I can share ChatGPT's evaluation of sources credibility, maybe that will give people some insights. For example, Reuters got 9.5 (out of 10), Lifehacker 7, Oprah Mag 3.

> where it focuses most of its reporting. (Economy and the US from a first glance?)

I'd say today is just US and economy-heavy day. There were many days when other categories and geographies made the top (I especially enjoy these). One example: https://newsletter.newsminimalist.com/p/tuesday-april-25-3-m...

Thank you!

Adding filter for news sources should be pretty easy, I'll consider it for "advanced view".

I see, newlines are missing, will try to fix soon.

6.5 seemed like a sweet spot for number of news. It results in 1-5 news each day. I tried 6.0 at some point, but I think it resulted in like 10-20 articles each day which seemed a bit much.

Good point, I noticed this as well.

As you said analyzing news in other languages should help with bringing up different perspectives. I'll work on that.

But I'm still not sure what to do about chatgpt's bias. For now I just accept it.

I don't decide it.

I just ask chatgpt to "summarize the article very concisely" and hope that it won't lose any important points. But chatgpt is really good at summarizing, so I don't worry much.

Having +/- buttons is a good idea, will add it to todo list as well.

Thanks, will look into that. I've been using the default template, which is relatively style-heavy, but will work on making it more minimal and clear.

Could you please expand your thought? Not sure I fully understand the idea.

Great idea. I'll do two website sections: daily and weekly. Weekly section will have it's own newsletter with, say, only news with score over 7.0.

Yeah, today is very economy-heavy day. But in the past issues there were more articles about health, science and environment.

Recent example: https://newsletter.newsminimalist.com/p/tuesday-april-25-3-m...

I'll look into that. For now, I thought of adding a ChatGPT summary of an article, but being able to also read the source is important.

Yes! I do exactly that. In the early version I posted the original titles, but many were too clickbait-y and hard to read. I find the rewritten titles much easier to scan through.

I think it's the best solution, will do it soon.

Oh, good point, just created it: https://rss.beehiiv.com/feeds/4aF2pGVAEN.xml

I have no control over it, but let me know if there are any problems with it — I'll see what I can do.

I had them at one point, but decided to remove for simplicity. Will bring them back :)

I noticed some bias because I only analyze sources in English. It should be easy to make ChatGPT also read news in other languages, maybe that will help. Is that what you meant?

Oh, sliders for custom scoring is an amazing idea. And should be easy to add — I already have all the ChatGPT estimations for different parts of the score. Added to the TODO list.

The problem with deduping is that some news get posted and reposted by different sources for several days (sometimes even weeks) in a row. That's a huge context I'd have to put in AI.

1000 news titles * 3 days * 70 symbols per title = 210,000 symbols = 40000 words = 53000 tokens. My current context window is 8000 tokens and I think 32000 tokens is max that GPT-4 allows.

---

Add: now that I think about it should be possible to do in several runs. Will keep thinking about it, thanks for the suggestion.