I wish people would stop creating these "default no" resolution things where every conceivable prompt must work multiple times in a row. These things function more as a way to gather new trader bonuses than to shine light on the question.

@WilliamGunn And yet the market was rating the probability decently-high for a decently-long time. That suggests that it was, in fact, a meaningfully-open question. (And, given that it's still at 16% as opposed to e.g. 3%, even with loans reenabled, I'd say it still is.) If you want other markets which are likelier to resolve Yes, you're free to make those; if you think this market is overvalued, you're free to buy No in this one; but the idea that this one is valueless to people aside from its creator seems very solidly incompatible with the realities of how people have been trading in it.

@Tulip I have observed a type of resolver which I call "default no", that behaves as I described above. I think it's useful for it to become common knowledge that this type exists, so that people know what they're really betting on, which is often quite different from the question phrasing. In this case, it's something like "will I be forced to admit this repeatedly works for all reasonable prompts", not "does there exist a prompt that elicits this behavior". When I notice clues that the resolver is this type, I often flag it as such to help create this common knowledge. A good practice is to signal the dependency on opinion in the question phrasing using "Will I think..." or something like that.

I came across a detailed article on the Ascendix Technologies website that provides an in-depth explanation of lease abstraction in commercial real estate (CRE). The article lease abstract defines lease abstraction as the process of summarizing key information from lease agreements, making it easier to manage and utilize. It discusses the significant advantages of implementing lease abstraction, such as increased operational efficiency, reduced error rates, and better data accessibility for property managers and stakeholders. For those interested in improving their lease management processes, this article offers valuable insights.

This market seems to not be gradually declining like similar markets as we continue to not see the expected advances in capabilities.

https://websim.ai/
toy that seems to be mostly raw html output from claude models

I'm curious if it works in incognito per the resolution criteria bc it seems a bit iffy if not

https://twitter.com/emollick/status/1778635083219415261 does this count?

It seems like the AI feature of Framer satisfies all of the criteria: https://www.framer.com/ai

It will develop an entire website from a text prompt and the quality is very similar to the examples provided.

Is this enough to resolve the market? If not, what would be required from a tool to resolve it?

Here’s a video demonstration and here’s the site it created.

Framer AI — Generate and publish your site with AI in seconds.

From raw idea to a real page in seconds. The more details you provide, the better the output. Each section has a unique combination of layout, copy and styling. Tweak what you like and delete what you don’t. You stay in full control.

@MitchellButler "The AI should be able to use references from other websites to understand the desired design"

I think this isn't satisfied

@MitchellButler Looking at the demo site you linked, one key limitation is that there's only one page, and none of the the text that appears to link to other pages is actually a link.

The original market description doesn't explicitly say that the site must be composed of multiple pages (although the reference site does have multiple pages). It does say that the generated site should have at least "a few interactive website elements," which I don't think is satisfied by the demo you linked.

@NLeseul I've just sold all of my shares in this market at a loss to make this case without any skin in the game. Also, I find the goalposts a little unclear, and I see how they might turn out to be in a position where there would be no financial incentive for a company to satisfy them.

The reference site says, "... with design skills roughly on par with something like this ..." and I see later in the thread that @CarsonGale mentions multiple text prompts, so I would imagine designing a single page with a single prompt is reasonable.

Also, I would assume that functional links would satisfy the description of "interactive elements," though the "client" of the AI designer does have to manually select the linked page from a dropdown menu in order for the link to be functional. fwiw, I hope gotchas like this won't be the deciding factors in this market.

As @MrLuke255 mentioned, it should be able to use references from other websites to understand the desired design as in the description. However, I wonder if the legality/liability might be more of a blocker for businesses than being "able to produce ...". Maybe an acceptable alternative would be to use provided images for inspiration? For example, pulling colors and common themes. Clearly, if it can do this, then it would also be able to do so with a 3rd-party website.

There are other elements where I think the market could use some clarity:

• Framer only seems to design landing pages. Should the "client" be able to ask for a blog listing page or a product page?

• Framer can only iterate on a design at the page level. Should the "client" be able to request changes to individual elements?

I'm not going to participate, but I'm still quite confident that Framer and/or one of its competitors would be able to satisfy those conditions in the next 18mo.

@MitchellButler I really appreciate the detailed thoughts and contributions to this market. It is, unfortunately, somewhat vague, so it is good to clarify.

I find Framer very impressive and I think it gets close to a positive resolution. I don't think it counts for purposes of resolving the market for a few reasons:
- I don't think the design elements of the site created are 'roughly on par' with the example Glendower website provided. I can provide specific examples if helpful, but when scrolling through the Glendower site, it seems noticeably cleaner and more capable than the draft Framer site provided. I don't want to be too much of a stickler on design, but it has to be within the ballpark.
- Correct me if I'm wrong, but I don't believe the Framer site satisfied having a 'few interactive website elements'. Examples might include linking a Twitter / LinkedIn page, having other clickable pages (e.g., the Investments, Strategy, Contact pages), or having graphics to click on and expand. I don't think having to manually link the other website elements should count - that doesn't seem like a 'gotcha' to me, it just seems like fairly simple functionality.
- The ability to reference other websites for design / content suggestions is not satisfied. In retrospect, this requirement might have been aggressive, and I appreciate your note regarding potential legality issues. If it is clear that this functionality could exist but is blocked for legal issues, then I'm open to relaxing this requirement in the overall spirit of the market.

In terms of the specific items you requested more detail on:
- Q: Should the 'client' be able to ask for a blog listing page or a product page? If what you're getting at is the ability to include functional links on the landing page to other companion pages (e.g., blog listing, product, etc.), I'm open to being convinced that a given website created by AI should satisfy the criteria without any other related pages, but it seems highly likely that the 'interactive website elements' requirement will involve companion pages for a positive resolution.
- Q: Should the 'client' be able to request changes to individual elements? A: I assume you're referring to edits like - "change the font and the design in this section of the site" or something like that, after initial generation. If so, I don't think this is required with my current market description and requirements.

I would welcome additional feedback and opinions from traders - this type of market is challenging to make specific and these types of comments help a lot to clarify for everyone what is intended.

@CarsonGale I appreciate all the detailed criteria. I don't think I saw: What's the scope of text prompts that can be allowed and the success rate needed? E.g., if I put in, "make a blog about giraffes for left-handed Swedes," and it never gets it right, but if I put in, "make a personal website for me, President Biden," and it makes a criteria-fulfilling website 90% of the time, how might that resolve?

@Jacy sorry I didn't see this previously. Here is what ChatGPT thought would be a good response to your question and I am fine with this resolution approach:

1. Scope of Text Prompts: The range of acceptable text prompts should be realistic and relevant to common website creation tasks. Prompts should generally fall within typical use cases such as personal websites, business sites, blogs, or portfolios. Extreme or niche prompts, like "make a blog about giraffes for left-handed Swedes," should not be the primary basis for evaluation.

2. Success Rate: The AI's success rate in generating criteria-fulfilling websites should be measured across a diverse but representative set of prompts. The overall success rate should ideally reflect the AI's ability to handle a wide variety of reasonable and common requests, not just excel at a few specific ones.

For instance, if the AI can create a personal website for "President Biden" with a high success rate (e.g., 90%), it indicates robust capability. However, if it consistently fails at less common but still plausible requests, it might suggest limitations in generalizability.

To resolve this market, I would propose evaluating the AI across a balanced set of prompts. If it achieves a high success rate (e.g., 80-90%) on a majority of reasonable prompts, it would fulfill the criteria. Niche or overly specific prompts should be used sparingly to test edge cases but should not disproportionately influence the overall assessment.

@CarsonGale no problem. I think ~80-90% on a majority of reasonable prompts is reasonable, and personally I don't expect this to even be a close call.

Re: “(ii) feed text inputs to an AI”

Does this include feeding your personal data, payment info, etc., so that the AI can establish profiles on other 3rd party platforms in your name?

@CKLorentzen Yes, that's right. Though such functionality wouldn't be strictly required in order to satisfy the resolution criteria.

The website should be accessible by typing in the URL into an incognito browser.

Does it have to be hosted on a second level domain that belongs exclusively to the site that was created?

In other words, if the site I'm creating is about bananas and the AI service I'm using called ACME, will both of these URLs satisfy the requirement:

1. https://www.bananas125.com

2. https://bananas125.acme.com ?
   

I would expect that if a service like this were to appear, it would by default put a website on a URL similar to the second one, and would require the user's action to go into configuration or something to put it on a URL similar to the first one.

@Roma or others please feel free to disagree and I'm open to changing my mind, but I think the criteria as I outlined in the market description is inclusive of second level domains (i.e., second level domains should qualify).

GPT-4 can already make the code, we just need a service that automates it.

@ShadowyZephyr ... hasn't that been the case since 2020?

@twink_joan_didion GPT-4 did not exist in 2020. There were other tools but mostly template based, I don't know about just from one text prompt.

@ShadowyZephyr Well yes lol, I meant since gpt-3 came out, which can write boilerplate html css and js as well. It's the wiring together part ("interfacing with any web hosting services, purchasing the domain if necessary, etc.") that doesn't seem likely (nor even possible in the current framework ?) to me

@twink_joan_didion I doubt old gpt-3 could do that but sure

@ShadowyZephyr that meme above is exactly what i'm describing haha

@twink_joan_didion "GPT-3" and what ChatGPT was using, "GPT-3.5", are different models with a huge gap in quality of output.