A Malleable Web
oio.landThe ungrounded breathless excitement of this piece is so grating.
> It's clear that prompting as the way to interact with AI models is here to stay. We wrote about it some months ago as something that might soon happen, and it became even more true than we thought.
Can we reel it in a bit? The jump from "true for some months" to "here to stay" is probably premature.
>Imagine a future where your browser's AI can actively change your experience of the internet entirely. For example, you may prompt this browser to display every website in dark-mode (even those that don't support the option),
Dark-mode browser extensions have been a thing for a while, and don't require any AI.
> If a website is completely malleable to the point that the data it holds can be re-shaped into something else, then you can almost see a website as an application programming interface (API).
... or you could open the developer tab and look at the API requests it's already making. API mashups were a buzzy thing more than a decade ago.
ML is doing new and exciting things right now, but this is acting like re-skinning something is a bold new frontier. 20+ years ago, custom skins for the media player app was fun. Don't try to sell it as a new landscape of possibilities in 2023.
Sure, this has been possible all along, and I've been using plain old python and javascript to do it. The essence of hacking is in effect making things "malleable to the point that the data it holds can be re-shaped into something else".
That said, I don't have a problem with re-skinning it as something of a bold new frontier. Honestly, it might be useful to express that sentiment at somewhat frequently - maybe it inspires someone who hadn't had that thought before. The advent of LLMs also makes this type of hacking accessible to non-programmers.
One thing I do think ChatGPT has been good for is reducing the "learning overhead" of making use of a lot of different APIs - this makes extensibility, scriptability, end-user programming etc. more attractive, although other problems like maintenance costs are still there. Automating the injection of arbitrary outside customizations into third-party systems not designed for it seems like a hard problem with or without LLMs, but I guess it's worth someone giving it a shot and seeing how far they get.
It reminds me a lot of the hype around self-driving cars. Technology that's impressive, and has made a huge leap forward, but where people all kind of assumed everything was about to fall into place to change everything and were so breathlessly excited about how driving was about to change forever and then it just... didn't.
There's tons of stuff to be excited about with ML, but I've literally never seen this kind of wild level of industry buzz before in my life about a product that frankly... doesn't work particularly well for a lot of the things it's being used for?
"Prompt generation is here to stay"
Is it? There are some real downsides to having every single interaction with a computer happen in natural language form. I thought that voice assistants were here to stay and only I found them kind of annoying to use. Fast forward, that whole interface change turned out to be a lot smaller and a lot less universal than people thought it would be too.
I don't know how to talk about the genuinely exciting things happening with ML right now when it's all drowned out by this extreme level of hype. I've never seen anything like it before, and it really just does not seem deserved. People are glossing over the fact that pretty much every single implementation of a conversational model on the market has had serious downsides that limit its effectiveness for tasks like code generation, answers, search, etc...
I'm seriously thinking about taking a break from tech news for a while because I feel like I'm drowning in people speculating about AI and how it's going to change literally everything, but a lot of it is just speculation and not much else.
---
Here's the thing: the author wants to talk about dark mode, then fine. It would take maybe a week to build an extension that did what the author is theorizing. Take the page source, feed it into GPT-3, tell it to output a stylesheet that will turn on dark mode, and then insert it into the page. If it works so reliably, stop talking about it and build it. And then we can compare the results and see if it's revolutionary. I'm surrounded by AI tech demos that ask me to think about the possibilities, but all I can see are tech demos. People keep on telling me "okay, but imagine once it's refined." Okay, refine it.
Again, there's exciting stuff happening, it's just being drowned by people saying "finally I can do X". Okay, go do it if you're so confident that X is going to work. The level of hype to actual substance is so wildly unbalanced right now. It's really exhausting, I feel like I need to go live in a cabin for half a year until everybody calms down and starts focusing in on what this technology will actually be good for.
> If a website is completely malleable to the point that the data it holds can be re-shaped into something else, then you can almost see a website as an application programming interface (API).
> A future where websites are combined into local frankensteins, only pulling the data it needs and combines it on a new website that does not exist.
This is exactly what I want to see. I've been toying around with similar ideas myself but prompting GPT-n with your instructions will make this extremely accessible. Browser configurability will be crucial, but we seem to be doing pretty well in that regard. Rare moment of optimism. Thanks for posting.
I want to see that too. This is a chance for browser to truly become an user agent, not only a tool for displaying documents (and ads).
I just want a browser extension that not only lets you create little css snippets for sites you browse, but also see the most popular snippets shared by other users. Maybe an up/downvote system too
I'm certain there was a plugin for Firefox that did just that, I don't remember the name, though. Greasemonkey definitely worked that way.
I know that this is not quite what Web 3.0 envisioned (specifically the original ideal of what Web 3.0 proposed; not the Web 3 that was proposed by blockchain proponents) but AI prompts could inadvertently create a mirror of what it promised.
I can see a version of the internet where what has been shared and posted parsed by AI, and easily recalled and presented by AI prompts. Able to be presented in a malleable form according to the need of a client.
I've been honestly kinda skeptical of AI (and critical of how data has been scraped); but this post did present a genuine use case that I can see being adopted. That said, based on the trend of Technology over the last 30 years, I predict that somehow AI will be used in much dumber and consumer-hostile way than I can possibly imagine.
The problem that I have with Codex (code-davinci-002) is that it sometimes just starts repeating the last X tokens infinitely. And I have tried using the parameter for frequency penalty but that seems to make it eventually "run out" of newlines sometimes. Because of that and the fact that it doesn't follow directions as well, I am trying to just use text-davinci-003 even for coding tasks. Its not as good at programming but usually good enough. Sometimes you need to give it hints about basic things. I mean the other issue is that code-davinci-002 has been in limited beta forever and Microsoft Azure won't acknowledge my request for AI services.
Hope we jump to davinci-004 codex and just skip 3. 002 has the issues like you say which copilot does too: sometimes it just keeps repeating a comment after every enter.
I dont know what to make of this article in particular. It's an interesting idea, and I believe the malleability aspect a lot. I think there's some very interesting suggestions here, and potential, but Im still pretty hesitant to speak for this.
But I do think it's remarkable & ultra noteworthy how vastly more plausible this suggestion is for the web than any other comparable computing system. The web has the malleable genetics to at least begin to entertain these ideas. Doing this anywhere else would require access to source, recompiling apps, or obscenely gross & mangled OCR+++ capabilities. The web attempt would be a high fidelity output, where-as all alternatives (other than being able to recompile arbitrary apps) would be incredibly funky.
The web has some really amazing pro-human superpowers, abilities ro be reshaped. The suggestions here of style transfer & api-itication are incredibly real potentials, bcause the web is so surpassing more human & rich than the typical computing world, where processes are isolated & secured compiled-down products uninterested in participating openly with the world.
Meanwhile & in-contrast, the former html5 editor is proposing a new world where all this creativity & flexibility is utterly impossible. Along with flutter/canvaskit already shipping this "new" anti-malleable bold-new-vision. https://news.ycombinator.com/item?id=34612696
There's no evidence at all to suggest that this kind of modular, Yahoo Pipes-style website is coming, AI powered or not.
What the author talks about is similar to what Blackberry and Microsoft tried to do on their mobile OSes, which was bringing in all your messages into a single hub. It failed because it relied on the OEM updating the app each time the Facebook Messenger or Gchat API changes broke functionality. Those companies have no incentive to make things easier for Windows Phone users.
The author is suggesting the same, but for all web content. This may be technically feasible, but it won't happen because commercial websites do not want their content piped to a third party with no attribution or ad impressions.
This would be the very first objection a VC would have if you went to them to fund such a thing.
This is off topic, but this thread seems to have an abnormally high number of dead comments from presumably botted accounts.
> you may prompt this browser to display every website in dark-mode (even those that don't support the option)
I already do that, no AI needed. Kiwi browser on Android has it built in and for desktop there's the Dark Mode Everywhere extension. Been using these for years, they're very good.
> remove news that is too negative
That would be cool, but I don't think AI would be needed for that either. Didn't we have tools that can detect the tone of a piece of text already? AI might be better at this, though.
I'd probably use it to filter out "breathlessly over optimistic articles about the latest cool thing by people that don't understand that thing". Like this one.
> Browse your bank history in the style of Spotify
So this person wants to give an AI access to their bank account? Seems like a bad idea.
However style transfer on websites that don't contain important financial or legal documents would be cool, might be possible soon, and would be a good fit for AI, so I'll give the author partial credits there.
My bet is that the semantic web is actually going to come this time. I think that before we did not all of the resources, but with the crypto hype dead this is going to be the actual next web.
So can I get Hacker News styled as a 80s dark fantasy yet?