I absolutely love how LLM are incredible tools for this kind of applications.
And yet, everytime I see this, to me it's clear how the labs are doomed.
These are the use cases we see, and kinda feel they're worth the hype.
But in the end it's all superfluous add-ons we wouldn't have bothered with in the past and be fine anyway. The reason we had never done something like this it's because the cost/benefit ratio was off.
Here is where I think the entire financial contraption falls apart.
Because the moment the cost of doing this stuff becomes again financially comparable to the mental/resources cost we needed to program it manually (and therefore refrain from doing it), we will start just... not doing it anymore.
If we instead will be able to still do it, it will mean that the cost of this has stayed as low as now (or gotten to free due to open models), and that will spell doom for the financials of the services enabling the LLMs to do this.
It's a catch22 for Anthropic and OpenAI, and either way they're dead in the water (financially, again, because their products as tools are indeed incredible. yet, they don't create enough value).
This, and the fact that other really value-adding breakthroughs that would justify the trillion dollar bet they're still making are nowhere to be found, except in the doomsday predictions of the AI-psychotic.
It seems like their long term bet is that people who use LLMs to build in the niceties and extra polish will attract more customers, and the race to do this increases overall usage. As if they're saying "subscribe or your competition will eat your lunch".
And until the cost of running LLMs exceeds the cost of someone's salary, people will continue their subscriptions.
But they do more than enable the superfluous add-ons, and have been inserting themselves as being mission critical in securing software. "Subscribe. Or else!"
I am dipping my toes into doing SVG with LLM assistance and I must still be prompting it wrong.
SVG is adjacent to HTML+CSS and arguably better documented, however, I find that RTFM is working out for me, rather than using LLM help. I am not confident that I can prompt correctly, however, I am able to learn the SVG specification and build my own working code for my own reference and learning.
I have specific goals and it is possible for me to chip away the old-fashioned way relying on just myself and the manual. LLM code that requires massive refectoring before trying to get it to work is always a longer process, more like an extra step, it also takes me away from understanding my code.
Part of the problem is that a lot of SVG is just churned out from an image editing application, and full of debris. I want lean code, which is a different thing. I am sure that there is the same problem with HTML + CSS, you could end up with anything from tables to grid layout to centre that div with a billion class tags added to it, because nobody styles elements, scopes their CSS or writes HTML with the content elements such as 'aside', 'section', 'details' and so on.
Again, nobody reads the HTML5 specification, to structure documents and let the styling work with the document structure in a concise manner.
I wish to be proven wrong, and hope that comes soon, however, for now, I think these 'incredible' tools wear a bit thin if you are wanting more than 'slop' that might have a few inspirational details to it.
I would like to see an AI that can do SVG or HTML/CSS with only the knowledge of the specifications, not a million badly voted up StackOverflow/Reddit solutions. I am not going to need a data centre the size of Manhattan to host an AI that does just this. I am also unlikely to contribute to such a solution.
Being an 'outlier', I fully expect mild-slop to take over and for Anthropic/OpenAI to linger on due to lots of people using them because who cares about underlying code? Nobody has, thus far, with HTML, all of it is spewed out bloated sea-of-div nonsense.
We can expect business as usual, and the trillion-dollar LLMs to continue just fine, due to the sheer resilience of modern day capitalism and the sheer number of users okay with mediocrity. They might take a haircut on valuations, but we have been here before.
The 4G auction sell-offs resulted in a bidding war with astronomical windfalls for many a government, with questions asked at the time regarding the viability of mobile telephony. It all worked out.
I am trying my best 'steel man' arguments here, even if LLMs are not my chosen path. I honestly wish those that take that route all the best, however, there will still be more 'artistic' niches where LLM won't cut it.
This is super cool but I feel like I could just build this myself with an LLM agent, so I get worried about the value proposition for services that try to charge for this. I guess really what you're paying for is some time saving and so that you don't have to burn your own tokens to build it?
As far as I can tell, this isn’t a product, and the page has a prompt to have your own coding agent generate these. Stepshots is a bigger product that offers interactive demos, which the page contrasts with these.
I strongly recommend not doing this. When I see a mockup demo on a homepage, it signals one or more of the following to me:
- there is no product built yet, just a potentially one-shotted idea
- the creators are insecure or untruthful about what their product could do, hence faking it with a demo
- lack of effort to create real screenshots and walkthroughs of the product
It signifies none of this to me. I routinely see these on websites with mature existing products.
This isn't a product though, it's just showing off an interesting idea, and telling you how to do it yourself
I absolutely love how LLM are incredible tools for this kind of applications. And yet, everytime I see this, to me it's clear how the labs are doomed. These are the use cases we see, and kinda feel they're worth the hype. But in the end it's all superfluous add-ons we wouldn't have bothered with in the past and be fine anyway. The reason we had never done something like this it's because the cost/benefit ratio was off.
Here is where I think the entire financial contraption falls apart. Because the moment the cost of doing this stuff becomes again financially comparable to the mental/resources cost we needed to program it manually (and therefore refrain from doing it), we will start just... not doing it anymore.
If we instead will be able to still do it, it will mean that the cost of this has stayed as low as now (or gotten to free due to open models), and that will spell doom for the financials of the services enabling the LLMs to do this.
It's a catch22 for Anthropic and OpenAI, and either way they're dead in the water (financially, again, because their products as tools are indeed incredible. yet, they don't create enough value).
This, and the fact that other really value-adding breakthroughs that would justify the trillion dollar bet they're still making are nowhere to be found, except in the doomsday predictions of the AI-psychotic.
It seems like their long term bet is that people who use LLMs to build in the niceties and extra polish will attract more customers, and the race to do this increases overall usage. As if they're saying "subscribe or your competition will eat your lunch".
And until the cost of running LLMs exceeds the cost of someone's salary, people will continue their subscriptions.
But they do more than enable the superfluous add-ons, and have been inserting themselves as being mission critical in securing software. "Subscribe. Or else!"
I am dipping my toes into doing SVG with LLM assistance and I must still be prompting it wrong.
SVG is adjacent to HTML+CSS and arguably better documented, however, I find that RTFM is working out for me, rather than using LLM help. I am not confident that I can prompt correctly, however, I am able to learn the SVG specification and build my own working code for my own reference and learning.
I have specific goals and it is possible for me to chip away the old-fashioned way relying on just myself and the manual. LLM code that requires massive refectoring before trying to get it to work is always a longer process, more like an extra step, it also takes me away from understanding my code.
Part of the problem is that a lot of SVG is just churned out from an image editing application, and full of debris. I want lean code, which is a different thing. I am sure that there is the same problem with HTML + CSS, you could end up with anything from tables to grid layout to centre that div with a billion class tags added to it, because nobody styles elements, scopes their CSS or writes HTML with the content elements such as 'aside', 'section', 'details' and so on.
Again, nobody reads the HTML5 specification, to structure documents and let the styling work with the document structure in a concise manner.
I wish to be proven wrong, and hope that comes soon, however, for now, I think these 'incredible' tools wear a bit thin if you are wanting more than 'slop' that might have a few inspirational details to it.
I would like to see an AI that can do SVG or HTML/CSS with only the knowledge of the specifications, not a million badly voted up StackOverflow/Reddit solutions. I am not going to need a data centre the size of Manhattan to host an AI that does just this. I am also unlikely to contribute to such a solution.
Being an 'outlier', I fully expect mild-slop to take over and for Anthropic/OpenAI to linger on due to lots of people using them because who cares about underlying code? Nobody has, thus far, with HTML, all of it is spewed out bloated sea-of-div nonsense.
We can expect business as usual, and the trillion-dollar LLMs to continue just fine, due to the sheer resilience of modern day capitalism and the sheer number of users okay with mediocrity. They might take a haircut on valuations, but we have been here before.
The 4G auction sell-offs resulted in a bidding war with astronomical windfalls for many a government, with questions asked at the time regarding the viability of mobile telephony. It all worked out.
I am trying my best 'steel man' arguments here, even if LLMs are not my chosen path. I honestly wish those that take that route all the best, however, there will still be more 'artistic' niches where LLM won't cut it.
This is super cool but I feel like I could just build this myself with an LLM agent, so I get worried about the value proposition for services that try to charge for this. I guess really what you're paying for is some time saving and so that you don't have to burn your own tokens to build it?
As far as I can tell, this isn’t a product, and the page has a prompt to have your own coding agent generate these. Stepshots is a bigger product that offers interactive demos, which the page contrasts with these.
Well you could have your LLM just generate that prompt then couldn't you?
Sure, but the prompt he’s sharing seems useful, and you can just copy paste it off the page. Not sure what your point is…