I never understood Mozilla's foray into AI.

There's just nothing about Mozilla that suggests to me they are experts on the subject matter and have much to contribute in the area. I could be wrong, but Mozilla is so tightly associated to the web that it just was a hard sell to me that their AI efforts were going to go anywhere from the start.

But if I tell an LLM "Thanks that suggestion worked" that feedback is lost to the void. There's no upvoting, no storage that I know of applying to all other questions people ask to let the system know "yes that answer worked particularly well". So all the LLMs can go on giving out the same answers forever not really knowing they are flawed... unless someone publishes an article about it.

There's no reason why LLMs can't have feedback mechanisms. In fact they do. ChatGPT has thumbs up and thumbs down buttons. It almost definitely tracks usage of download and copy to clipboard buttons.

I mean on the low end.

Like formerly a small business would pay for the valuable HTML and FTP knowledge, but today many will just use Wix instead.

And more complex things are more proper webdev and not just knowing some HTML tags.

I remember there was a short time when the web was new and "webmaster" was a profitable occupation. Then a combination of improved web design software, CMSes, frameworks and the like quickly ate the low end, and the higher end just got rolled into software development.

I'm not sure why anyone thought that knowing what keywords a specific model recognizes best could ever be an enduring form of employment. To me it was always clearly extremely temporary.

Amazingly, ISA disappeared much later than I thought. It seems there's a Skylake motherboard with an ISA slot out there. That architecture was only discontinued by intel 6 years ago.

Also, I believe the bus still internally survives in many boards that don't have a physical connector, and the LPC bus is ISA in a slightly newer form.

No, this countermeasure isn't going to do anything useful. Any web spider is going to run into thousands of these, and therefore already is going to be coded to tolerate it fine.

These countermeasures were made to trip up worms and naively coded Perl scripts used by mom and pop spamming operations. That is useful to an extent, but isn't going to do absolutely anything against the bigger players.

All you have to do is to keep track of stats like depth and performance, notice that a branch is doing badly, and prune it off.

Eesh.

This is an ancient idea. Probably discussed at length right here in the 2000s. No crawler worth its salt is going to even notice this.

Like oh, this:

https://it.slashdot.org/story/...

or this:

https://developers.slashdot.or...

I really liked the older XPS 13. Light, powerful, great Linux support. Smashed the screen on that, figured okay, I'll get the new one, the XPS 13 Plus. I want more RAM anyway.

It's gone downhill. Two USB C ports instead of 3. Worse touchpad (previous was clicky, this one seems to have some sort of embedded haptic device that one day just stopped working and no physical mouse button keys). Function keys and most importantly ESC key are now a touch surface instead of physical keys.

Hardware-wise it's still nice to use, but otherwise a disappointment. I'll be way more careful with picking the next one.

"Can display any website" implies the heck of a lot of features.

A modern web browser is effectively an operating system in itself. Here's just the WebGPU standard, take a look: https://www.w3.org/TR/webgpu/

Just scroll down that page and try to estimate how much programmer time is needed to implement that correctly. And you can use most of the web without that just fine actually, until you come across something that happens to want it. Modern browsers have an alarmingly large amount of such features.

Why is TIOBE still talked about? Their data is complete nonsense. Don't believe me? Look at how it's made:

https://www.tiobe.com/tiobe-in...

It's basically searching for "$LANGUAGE programming" on various search engines, then taking the "5 million results found" counts. After that they use some fudge factor they seem to have come up by themselves. Why is google.com worth 7.69%? Why is Wikipedia the second in the list, do mentions on Wikipedia go up and down by language popularity? Why is google.co.uk in the list, when it also returns English results?

Does Google even promise that the result count they display is remotely accurate?

Let alone that this barely means anything if it worked because it's trivial for anyone to inflate counts by encouraging the use of the term, or it can also go down if some prominent site happens to go down.

It's an absolutely terrible metric and I don't understand why anyone cares about it. At least try to be remotely accurate. Look at Stack Overflow activity. Look at commits on Github. Look at subreddits. Something that indicates actual usage at a given point in time.

And many of them will be perfectly right.

When I was 12 at school we had an active pirate scene going during the lunch break. And it included some pretty fancy stuff. I don't think any of us had the pocket money to legally pay for 3D Studio.

These days it's much easier to click "Buy" on Steam. Having money is quite convenient.

Basically, to me ChatGPT (and other generative AI systems) are nothing more than these systems dressed up with a new name and fresh coat of paint.

Nothing of the sort. ChatGPT is amazing.

Eliza is just string manipulation. Looks clever but does exactly one thing.

ChatGPT does high quality translation on a near-professional level. It can deal with typos, slang, and unconventional spelling like used by Hagrid in Harry Potter. It can even translate random manga pages and memes.

If you're not very old you may not remember what the first systems -- in Eliza's day -- were like. They just did syntactic analysis of each sentence, and looked up words in a dictionary. They couldn't relate what happened in one sentence to the next, constantly picked the wrong meaning for a translation, got confused by trademarks like "Windows" (translating that literally), choked on any misspellings, couldn't deal with non-standard spelling like "Let's goooo!", and so on and so forth. Translation in those days was barely good enough to get the gist of what somebody was talking about, and whether it was worthwhile to seek a human to actually translate things properly.

It's freaking magic, man.

It depends on what.

There are things that China can do amazingly well, like electronics manufacturing. You can get crap, but you can also get quality, there's plenty to choose from.

You think? What makes you think AI can actually use those quality indicators?

Training is just dumping a bunch of files into a folder and letting the GPUs crunch that.

So let's say you've got a dataset from Reddit. You probably get something like a bunch of JSON with upvote/downvote info, comments, etc. So you just don't use any images that were heavily downvoted. You exclude anything from the meme type subreddits and concentrate on ones for image appreciation. For a more complex approach you feed the comments into GPT and try to figure out whether people are talking about the image quality negatively.

It doesn't need to be perfect, you just want to avoid ingesting a large amount of really bad content. It's fine if some does slip through.

And what makes you think they will stay free from AI data?

It doesn't need to be. You just don't want to train on the bad ones.

Image AI is unlikely to suffer from such issues.

For images we've got plentiful sources of quality indicators. There's specialized galleries, tags, upvotes/downvotes, favorite counts, reposts, even just comments that can be quite easily evaluated. Overall this makes it very easy to select for the highest quality content.