mg

I think that situation will not arise. Because the boundaries between humans and technology will blur.

We are already cyborgs. Sitting on bicycles, using a handheld computer and headphones.

After a few decades or centuries, there might be no organic cells left. We might be made completely from silicon. And look nothing like we do now. But we will continue to do what we always did. Work hard to optimize our survival and reproduction.

I'm not sure. Googling for "gpu non-determinism" and reading around the results seems to indicate that GPUs are often not deterministic.

I don't know why, but it is easy to imagine a scenario where non-determinism makes sense. For example the computer could say "I am supposed to do this floating point division. Let's see ... my very exact FP-Division unit is pretty hot at the moment. 82°C. Let's pass this calculation to the slightly less exact FP-Division unit then".

This is what I do. I calculate a score over a sample of questions and replies. I'm not doing unit tests.

Comparing the scores of two prompts will not give you a definitive answer which one is superior. But the prediction which one is superior would be better without the noise added by the randomness in the execution of the LLM.

I just tried the same prompt twice, both with

    'temperature': 0,
    'seed': 1,

And I got two different replies.

The 'system_fingerprint' in the reply was the same in both of the json responses. So it seems that even when you get the same 'system_fingerprint' back, replies for the same prompt will not be the same.

Are any of these stable? I mean when using temperature=0, do you get the same reply for the same prompt?

I am using gpt-4-1106-preview quite a lot, but it is hard to optimize prompts when you cannot build a test-suite of questions and correct replies against which you can test and improve the instruction prompt. Even when using temperature=0, gpt-4-1106-preview outputs different answers for the same prompt.

Controversial take: Most indiemaker projects and startups would probably be build faster if they just used the

    <?php include 'header.php'>
    Hello <?=$name>, how are you today?
    <?php include 'footer.php'>

approach.

Afaik, https://twitter.com/levelsio still works this way.

Well, it would be much better if Biden was posting via potus@whitehouse.gov. That would strengthen the open nature of Activity Pub.

Intersting that Joe Biden is now available via Activity Pub:

https://mastodon.social/@potus@threads.net

Looks like Threads is indeed federating with everybody:

https://indiehackers.social/@potus@threads.net

So you could make your own instance and read everything in exactly the UI you want? Customize sorting and discovery via your own algorithms?

Is that the open web 2.0 we have been waiting for?

PS: Looks like you can't link to external accounts on a typical Mastodon instance the way I did. When you search for "potus@threads.net" and click on the result it works.

What is everyone's typical use case of a playlist? The examples in the article ("songs to serenade my cat") seem rather contrived?

You say so on a website that displays the number of likes (here: "points") for every post right on the front page.

Regarding whether it is "a good thing for social networks to provide such stats":

It's a central part of social media, even here on HN.

Look at the front page. Each post shows 6 pieces of information:

    Title, Domain, Points, Author, Time, Comments

2 of them = 33% are numeric indicators we trust because we trust HN.

Likes and follows are probably two of the main reasons that centralized social media sites have displaced websites, rss, forums, mailing lists, newsgroups etc.

The next hurdle to create distributed social media seems to be distributed likes and follows. Those are the social currencies of today's generation.

In a centralized system, you trust the central authority to show the correct number of likes a post got and the correct info who liked a post. Same for number of followers and who follows whom.

Do any of these 3 protocols have an approach to do the same in a decentralized world?

Great. Thanks for the infos!

Hello Hanoi! That's probably as far as it gets from Germany :)

Yes, DNS is a different issue. To bring that down, I would have to use a distributed DNS I guess.

Interesting. Thanks for the info.

Would like to hear how it feels from Australia compared to other websites. Like Hacker News, Goodreads and Wikipedia for example.

    You need to be on the edge, they say.
    Be close to your users. Minimize latency.

How much of an issue is latency in practice?

Here is an example. My book recommendation project Gnooks, which I run on a server in Germany:

https://www.gnooks.com

Does it feel too slow for anybody?

Over the last years, I have gotten many thousands of suggestions from users on this project. Yet, as far as I can remember, nobody ever touched the topic of latency. While the largest group of users is from the USA.

Maybe that is the solution to the Fermi paradox.

How do you explain that Bing has only 10% of the search market share on the desktop?

Windows has over 70% market share on the desktop. Windows comes with Edge as the default browser and Edge has Bing as the default search engine.

But Edge only has 12% of the Desktop browser market share.

It looks like 6 out of 7 windows users switch to Chrome for some reason. If not for the cleaner interface - why?

I was waiting for this.

Google showed that providing an input field in which people can put their questions and then display ads next to the answers is one of the best business models ever. If not the best business model ever.

Google was free, open and without ads in the beginning. It is just too tempting to become the Google of the AI era to not try and replicate this story.

Microsoft is trying it too, but for some reason, they do not provide the simple, clean user interface which made Google successful. I always wondered, why they didn't do that with Bing search. And now with Bing copilot they chosen the same strange route of having a colorful and somewhat confusing interface.

Let's see how OpenAI does.

Some while ago, someone posted a tool to analyze the frequency of any term on HN over time:

https://hn.curiosity.ai/#/trends

It seems to still work. And the numbers seem to be roughly in line with this article.

If the component only wants to rerender when the counter changes from odd to even or even to odd, it can cache the value and do as it pleases.

    I wrote assembler 68000 during the Amiga demo scene days

I did that too.

And I couldn't be more enthusiastic about the current AI developments.

I have been dreaming about an AI system which reads websites and predicts the next word ever since I got my hands on the web. And tried to predict if that could lead to strong AI.

I even built a somewhat related experiment as my first public web project. A system which asks users for their favorite 3 bands and then predicts the next favorite band of that person. And after talking to a lot of users, it turned out quite useful. That made me even more convinced, that strong self-learning AI would become a thing.

During the dotcom time, I was too inexperienced to fully grasp what is going on. I was doing my own thing, not enjoying the full picture as much as I could have.

But now, with the advent of LLMs, I enjoy experiencing this shift in technology so much! Watching this explosion from day 0. With a much broader horizon than last time.

I actually don't think that tinkering with LLMs themselves is the way to create the next Google.

Yahoo was putting content directly on the web which seemed like the obvious way to create "the" web company. A media company made for this new medium.

But the much bigger business model turned out to be to just leverage all the existing content.

My feeling is that this time, it will be to leverage the existing LLMs and APIs out there and tie them together to something useful.

So the Google of the AI era might turn out to be a good old Python program that makes and answers http requests.

When I need to signal something across my application, I use events:

    window.dispatchEvent(new Event('counterChange'));

And every part of the application that wants to react to it can subscribe via

    window.addEventListener('counterChange', () => {
        ... do something ...
    });

Anything wrong with that?

So after calculating the 4 scores for [11,7,4],[4,4,8],[11,8,11],[4,11,7] we can put the 4 scores into a formula and it outputs the correct 3 values? Which formula would that be?

I wonder what the optimal strategy would be if you can't see the color and only the answers when you click submit.

Hillclimbing is already somewhat efficient:

    For each slider:
        - Start at 0
        - Move to the right until the score drops
        - Move one to the left

That should result in something like 9 tries per slider on average, so 27 tries per color.

One signal that could be used to improve it: The difference in score between 0 to 1 gives you the approximate length you have to move to the right.

Due to rounding, you don't get the exact length.

So My guess is that with an optimal strategy, on average you would need something like 4 tries per slider.

That comes down to and average of 12 tries per color.

Cool idea.

I have the feeling this would be even much better without the red squares and with a regular sized board. And that it would feel more logical and interesting to me.

Or is it impossible to create challenging puzzles without the red squares?

But it's learning!

What is the difference between a typo and a duplicate? If you mean that two ways of writing the same name are both legit, then you have to decide on one being the more "correct" one. After a while Gnod will figure out which one is the more common name.

And "complicated errors"?

Yes, in the beginning pretty much everybody hated it and thought the project was nuts. I got pretty much no positive feedback but lots of negative. I was like "But it's learning! It's learning!" :) Strangely, that convinced almost nobody, even among my friends.

Now that many millions of people have used it, I get a lot of great, often enthusiastic feedback on how Gnod makes the best recommendations.

That teached me that you can't convince people with just an idea. For most people, you have to deliver something which is already useful to them.

The Music-Map and the Movie-Map are based on user preferences. The Music-Map is based on https://www.gnoosic.com and the Movie-Map on https://www.gnovies.com, two AI projects I started before the maps.

The AI and the mapping algorithm are my own developments. I was mostly inspired by thinkers like Douglas Hofstadter and John R. Koza.