mg

Yes. Happy you like it!

This is great. This sentence struck a chord with me in particular:

    Imagine applying these techniques on the Common Crawl
    You would be able to produce a ... map of the internet.

Making maps of things not usually on maps has been my passion for years. And I made many of them. One of the more popular ones that some of you might know is the Music-Map:

https://www.music-map.com

I have had the urge to make a map of the web for quite a while. Already registered the web-map.com domain for it. I did some experiments, built a custom crawler and an algorithm which finds related websites fast. It showed that the project would be feasible.

But I hold back on doing it, because I already run multiple experimental maps and have yet to come up with a business model for "making maps of everything".

My default prompt to test new LLMs:

Me: "Two cars have a 100 mile race. Car A drives 10 miles per hour. Car B drives 5 miles per hour, but gets a 10 hour headstart. Who wins?"

EagleX 1.7T: "Assistant: Car A wins because it has a higher speed."

(Tested on https://huggingface.co/spaces/recursal/EagleX-7B-1.7T-Gradio...)

It's strange that addons have so much unsandboxed power. I think browser manufacturers are not very keen on letting users customize their browsing experience. Otherwise they would give them finer control over what addons can do.

Fortunately, most customizations can be done via bookmarklets.

Benefits of bookmarklets:

- Easy to edit

- Inactive until clicked

- Work in all browsers

- Work on mobile

- Integrate nicely into the UI. You can move them around, put them into any bookmark folder, assign shortcuts.

I wrote this bookmarlet editor which makes it easy to convert between clean code and a bookmarklet:

https://www.gibney.org/bookmarklet_editor

Here is one that I use to read HN:

    if (!location.href.match('ycom'))
        location.href='https://news.ycombinator.com';

    doTheMagic = function() {
        if (typeof e!=='undefined')
            e.style.background='#c0c0c0';

        e=Array.from(document.querySelectorAll('.age a:not(.seen)')).sort().at(-1);e.style.background='#ff0';
        e.classList.add('seen');
        e.scrollIntoView({behavior:"smooth",block:"center",inline:"nearest"});
    }

    doTheMagic();

    if (typeof theMagicIsActive=='undefined') {
        document.addEventListener("keydown", function(evnt) {
            if (evnt.key === "n") doTheMagic();
        });
        theMagicIsActive = true;
    }

When clicked and HN is not already opened, it opens HN. When on HN, it jumps to the next newest story. Or comment - it also works in the threads. After jumping to the newest story, clicking it again jumps to the next newest story. Instead of clicking it, pressing "n" also works.

Also works in the comments view of a profile. So when you want to see if someone replied to your comments, you click on "comments" in your profile and then on the bookmarklet to jump to the newest reply.

This seems to be the source of the data:

https://data.oecd.org/emp/hours-worked.htm

There might be no reasoning in a single pass which outputs a single token. But in the loop where the output of the LLM repeatedly gets fed back into its input, reasoning is clearly happening:

The LLMs lay out how to go about figuring out the answer, do a series of calculation steps and then come up with an answer.

If you add "Please answer in just one short sentence." to the prompt, even the top ones get it wrong.

Hmm.. it fails for my favorite test prompt:

https://www.gnod.com/search/ai#q=Two%20cars%20have%20a%20100...

I gave it 3 tries and each time, Yi picked one of the cars as the winner.

I've been watching for many months now, how LLMs got better and better at solving it. Many still struggle with it, but the top ones nowadays mostly get it right.

CTRL+u gives a nice example of the syntax.

For example to make a box with rounded corners and a top and bottom section, all you need is this intuitive one-liner:

<div class="mt-5 mb-8 first:mt-0 last:mb-0 relative overflow-hidden rounded-2xl"><div class="pt-2 bg-slate-800 shadow-lg group"><div class="flex text-slate-400 text-xs leading-6"><div class="flex-none text-sky-300 border-t border-b border-t-transparent border-b-sky-300 px-4 py-1 flex items-center">app.css</div><div class="flex-auto flex items-center bg-slate-700/50 border border-slate-500/30 rounded-tl"></div></div><div class="children:my-0 children:!shadow-none children:bg-transparent"><pre class="language-css"><code class="language-css"><span class="token atrule"><span class="token rule">@import</span> <span class="token string">"tailwindcss"</span><span class="token punctuation">;</span></span></code></pre></div></div><div class="pointer-events-none absolute inset-0 rounded-2xl dark:ring-1 dark:ring-white/10 dark:ring-inset" aria-hidden="true"></div></div>

Hmm.. I guess you could write a little script (2 lines?), let's call it "vimlive" which reads from stdin, streams to a temp file and opens that temp file in vim.

Then with:

    command | vimlive

You get vim with the current state of the data and you can update it with :e

Is there a benefit over

    command | vim -

?

I love doing that, because then I get all the power of Vim to look at the output. Search, regex search, sorting, bookmarks, counting, macros ...

I started to write output that is indented, so I can use Vim's folding to comfortably look at complex data.

For example if you have entries like this:

    2023-04-12 17:22 Sighted and UFO
        Location: 27.1742424 12.137234
        Temperature: 24°C
        Observer: Joe
        Actually it turned out to be a dove.

And you press zM in Vim, they all turn into:

    2023-04-12 17:22 Sighted and UFO
        + 4 lines

So you have a nice overview and can open individual entries by putting the curser on them and hit zo.

And while folded, you still have all the power of Vim to search within them. Regex search. Count by any measure you like etc etc. Actually, you can do anything, as Vim allows you to execute arbitrary vim commands, vim macros or shell commands on your buffer.

Wanna know how many doves typically get mistaken for UFOs around New York grouped by month? Write a script for it and it will be forever at your fingertips.

At the core, this seems to be about shadowrootmode=open, a new html feature which already landed in Chromium and Safari:

    <div>
      <template shadowrootmode=open>
        This is <slot name=a></slot> and <slot name=b>
      </template>
      <span slot=b>sunny</span>
      <span slot=a>funny</span>
    </div>

renders as:

    This is funny and sunny

So it looks like HTML got a bit of a native template system now.

What I have been wanting for decades is a native template system in HTML which supports urls. Similar the the script tag, but which loads html:

    <html url="/header.html">
    <html url="/content.html">
    <html url="/footer.html">

I always wondered why this was never implemented. It seems such a basic functionality which would allow to comfortably build websites in a modular way without scripts or serverside processing.

My favorite prompt to throw at LLMs:

    Two cars have a 100 mile race.
    Car A drives 10mph. Car B drives
    5mph but gets a 50 mile headstart.
    Who wins?

So far, no LLM gets it consistently right. Most of the time, they confidently declare one of the cars as the clear winner.

Sometimes the explanations are so convincing, that I begin to wonder if my own conclusion - that the race will be a tie - is wrong :)

It looks like making things easier to search through is a typical Hacker News project.

I had a similar idea some years ago for flashdrives:

https://www.productchart.com/flashdrives/

Except I decided to implement a two-way sort, so the result is a chart, with axis "price" and "size". Or whichever parameter you chose. (At least on Desktop. On mobile, I show a list.)

Then I added ssd drives:

https://www.productchart.com/ssd_drives/

And more product types like monitors, phones etc, which you can access from the menu.

Maybe we should make a group to discuss the way forward and how to make the whole world easier to search through.

One perspective I find interesting is how fast Microsoft grew in the desktop pc area:

https://twitter.com/marekgibney/status/1753116011681968552

Back then, for over 10 years, it grew at a pace not seen since then.

If Nadella grows Microsoft into the AI area as fast as Gates did in the Desktop area, it would become a 100 trillion dollar company within the next ten years.

Is AI a step forward for productivity as big as desktop computers were back then? I think it might be. Not sure what percentage of people became more productive via PCs. And how much more productive they became. But with AI, it seems everyone will benefit from it. And become twice as productive or so within the next 20 years.

Programming without an LLM at hand would already feel very clumsy to me these days.

I have the theory that the most powerful way to be someone people love to talk to is to learn to get into a good mood.

This theory is two-fold.

First: Mood is contagious. So being in a good mood yourself is the most direct way to make others get into a good mood too. And that's what makes us like being around someone. When they make us feel good.

Second: It is possible to learn how to get into a good mood. It's not so much related to the situation around you, which mood you are in. It's more of an internal state. Similar to how you can change your physical posture, you can put yourself in different mental postures. The more you practice, the better you get at it.

    It's not tested at all

This is what Wikipedia says about hidden variables:

https://en.wikipedia.org/wiki/Quantum_mechanics

    Many Bell tests have been performed and
    they have shown results incompatible
    with the constraints imposed by local
    hidden variables.

There are always infinite interpretations of everything.

But the "no hidden variables" theory is the most popular and well tested theory about whether things have a state before we measure it. And it postulates that they don't.

We see more surprising effects than expected effects.

95% of our explanation for how the universe behaves is "dark matter" or "dark energy".

In quantum theory, it surprisingly was possible to test the hypothesis that things don't have a state before we measure them and then they stick to that state.

So I wouldn't rule out the possibility that we can test it for other "secrets of the universe".

Why is it irrelevant for the universe, but not for electrons?

Maybe there are no secrets, but every time a new experiment is conducted, the universe randomly decides the outcome and then sticks to it?

I think this is what we assume about small things already, like the position of a previously unmeasured photon or the spin of a previously unmeasured electron, right? Neither the photon has a position, nor the electron has a spin, but we measure one. And from then on, the the universe acts according to the measured values.

You can click on the question mark and then, when you click on one of the examples, it will fill the code area with the code for that bookmarklet.

I don't see any proof that time exists. For all I can tell, only the present moment exists. You say I "observe it again". But how do I know I observed something in the past? The memory is just an experience I have in the present moment.

I prefer bookmarklets because they

- Are easy to edit

- Are inactive until clicked

- Work in all browsers

- Work on mobile

- Integrate nicely into the UI. I can move them around, put them into any bookmark folder, assign shortcuts.

I wrote this bookmarlet editor which makes it easy to convert between clean code and a bookmarklet:

https://www.gibney.org/bookmarklet_editor

When I compare the "Hello" on the screen to the concept of "Hello" in my mind, I cannot hold the details of the character shapes in the same clearity as I do when I just focus on the shapes.

So the siutation with the two Hello's might be just as simple (My rough estimation: 100 bits) as the situation where I focus on the shapes.

That's why my feeling is that the lower bound for how complex the universe is, is 100 bits.

When I have the experience of seeing the shape of those 4 letters, I can't simultaneously reason about "eyes" or if there is something "in front" or if there is a "you" as you say.

I just have an experience of a certain complexity, which I roughly estimate to be 100 bits or so of information.

You describe more than I can witness simultaneously.

While I look at the shape of the 4 letters "Hello", I can't interact with other agents. I can't even imagine such an interaction without losing the full grasp of how the four chars look like.

So I would say I don't know if anything else exists. Except of the 100 bits of information or so which I can witness.

You bring up the concept of time.

But is there a proof that time exists?

I can "remember" something, but that act of "remembering" is happening in the now. I don't know if there really was a past.

When I look at "Hello", I don't remember something at the same time. The shape of the four "Hello" letters are all I can hold in my mind. So from that perspective, "Hello" might be all there is and all there has ever been.

Is there any proof that the universe is bigger than what I experience right now?

For example: My eyes can focus on this word on my screen: "Hello".

I don't know how many bits of information this word contains. I can see the shape of the four chars quite clearly. Maybe that is 100 bits of information or so?

So I would say the universe probably is at least 100 bits large.

Is there any proof that it is larger?

Is there any way it could be smaller?

Yesterday, in Microsofts earnings report, a number that stood out to me was their quarterly Linkedin revenue of $4B.

That revenue probably mostly comes from the old school hiring process where employer and employee get to know each other and then jump into a long term contract, right?

I'm surprised Upwork and Fiverr are still so small in comparison. Their model of letting people quickly band and disband on projects seems so much more natural to me in the times of remote work.