Maybe its just me but I can't figure out what exactly this is. I tried your demo, gave it a Twitter user. And its now generating blog posts in the "voice" of that Twitter user? And this is a tiny first step to... what? I know this sounds antagonistic but I'm just genuinely confused.
The idea is this "identity card", which is super early in our dev process. The goal is to make it so you can have your voice as a thoroughly vetted example, and then be able to use that to finetune LLMs, or use it at inference time to enhance LLM outputs to sound more like you.
We'll be building out a full CMS, and shipping parts of it over time.
Thanks again, it's super early but we're working on it. :)
Ah, I'm not much of an AI person so I'm probably just lacking context. Regardless I do think it would be helpful to point out - your second paragraph refers to "the tooling" and that you're focused on "showing the dev experience and community aspects". I don't really see how your site or the demo you're showing actually highlights either of those things.
> you can have your voice as a thoroughly vetted example, and then be able to use that to finetune LLMs
You have piqued my interest.
Can you explain in more detail how a user will fine tune a model. Will the LLM just replicate a user's conversational style? Or will the user be able to set deeper parameters, for example to define goal-oriented outputs for agentic actions?
So the goal initially is going to be to create a "constitution." Basically a document that outlines your voice, or whatever.
From there, you can create examples for every sentence and classify them... ie "this is a good example" or "this is almost a good example."
Then, after you have a good enough enough number of human versions (idk this number yet), you can use those to create dozens more examples, refine, etc... then at some point you've thoroughly described a document enough that there's no more delta. Then you can use the document to train / finetune / infer.
The effects are that the "instincts" of the model (if on the training / weights) or the "thoughts" of the model (at inference time) are closer to a well vetted baseline.
I'm incredibly excited for this launch! Bolt Foundry is the first company to use [Isograph], which is a framework for building data-driven apps that are performant and stable out of the box. It does so by making heavy use of a compiler and of generated files, and essentially being an incremental computation engine for your UI. Your UI becomes performant by doing the least amount of unnecessary work.
And Bolt Foundry is truly the perfect first adopter! As "A big tech dev experience" implies, they're very focused on providing great DevEx. And a bunch of them previously worked at Meta, where they used Relay, and understood the powers that it provides.
Isograph aims to provide an even better developer experience (see the YouTube link), and give you the things that previously were only feasible at FAANG-co scale: e.g. even in a very dynamic app, fetching just the JavaScript and data you end up using.
So far, they've had a great experience using Isograph!! You should also [give it a try] or [contribute].
Interesting! I think the idea of a "tone" in your content being enforced by a sort of... prose linter makes a lot of sense.
You can assume 1 person with at least some practice in writing will have a clear tone and presence in the things they write, but this would be rather useful for corporate communications, especially when simpleton engineers (such as myself) try to write blog posts for a company engineering blog. If there was something nudging me towards the style of the blog, without steam rolling me constantly by rewriting everything, I would like that.
For the CMS, (I think you're building a CMS?) I think building around THAT use case would be an awesome selling point. Let content folk spawn special ephemeral links that let someone write up a draft without having to set up an account. Have some sort of approval process in there too, that could be configured on that "magic link", so I don't impulsively publish to everyone.
Make it super easy to just stumble into writing something the company can actually communicate with, right away. There's a lot of passionate people at companies, with 0 writing experience that would lend a lot of authenticity to blogs/engineering comms/recruiting ads, if they they had the bumpers in their bowling lane.
nothing at all. it’s not just about blog posts though, it’s figuring out what to write about, finding the best comments of fans who are smart and interesting without spending your entire day doing that (or hiring a team).
it’s not about having it write for you, it’s about scaling who you are to do lower value tasks.
and add a small amount of padding around the body to make the text scaling pixel density agnostic though.
I would also remove the overflow:hidden on body, and remove the fixed height on #root. Nothing needs to be a scrollable container here, the viewport can just scroll over the body. That's causing the bottom ~2 paragraphs to be "stuck" under the bottom edge of my iPhone 13 mini in safari. Ideally use `min-height:100dvh`.
I think this is interesting in sort of a scifi way because if you can establish a reliable enough characterization of "tone" then maybe you can then create a steganographic side channel out of deviations from that tone: so your adversaries see inane blather about something unimportant, but your comrades can decode the blather into more meaningful messages. Meanwhile, others are hosting your communications for free because what is actually cyphertext appears to them as engagement.
With the constant deluge of AI slop everywhere, your voice is basically all you have.
This is basically a tool to train a computer to "steal" your voice. I find that sad.
Why can't we figure out how to use this technology in something other than creative tasks? I don't care what anyone says, a blog post showing off some technology or accomplishment should be a creative task, otherwise what's the point?
for me, i’ve only been able to tell my stories at scale with the help of a lot of infrastructure. as a solo creator, i’ve never had the impact i did in local news or at cnet or engadget.
it’s not about having it write for you, it’s about having it help you write better.
writing is thinking, but 90% of writing is editing.
Haven’t dug deeply into it, but it appears to be CSS that doesn’t play nicely with the Safari rendering engine on mobile. With only five devs, maybe none of them have Apple devices?
Super weird behavior for me too. When I double tap to zoom into the text, I get some weird nested scroll bar state and can only read the first few paragraphs.
This is great! Assuming the LLM versions of me will build their fake consumer clones believably. Once this goes FAANG-co scale it will only be a couple of months until we can all go public and Advertisers will stop buying ads putting an end to the enshittification of the interwebs. Genius. Who came up with this? ... and is Apple likely to sue considering their patent from 2012?
Maybe its just me but I can't figure out what exactly this is. I tried your demo, gave it a Twitter user. And its now generating blog posts in the "voice" of that Twitter user? And this is a tiny first step to... what? I know this sounds antagonistic but I'm just genuinely confused.
Totally fine, thanks for trying it out.
The idea is this "identity card", which is super early in our dev process. The goal is to make it so you can have your voice as a thoroughly vetted example, and then be able to use that to finetune LLMs, or use it at inference time to enhance LLM outputs to sound more like you.
We'll be building out a full CMS, and shipping parts of it over time.
Thanks again, it's super early but we're working on it. :)
Ah, I'm not much of an AI person so I'm probably just lacking context. Regardless I do think it would be helpful to point out - your second paragraph refers to "the tooling" and that you're focused on "showing the dev experience and community aspects". I don't really see how your site or the demo you're showing actually highlights either of those things.
Yeah, honestly this might have been too early of a launch, but kind of just want to put something out and then iterate.
> you can have your voice as a thoroughly vetted example, and then be able to use that to finetune LLMs
You have piqued my interest.
Can you explain in more detail how a user will fine tune a model. Will the LLM just replicate a user's conversational style? Or will the user be able to set deeper parameters, for example to define goal-oriented outputs for agentic actions?
So the goal initially is going to be to create a "constitution." Basically a document that outlines your voice, or whatever.
From there, you can create examples for every sentence and classify them... ie "this is a good example" or "this is almost a good example."
Then, after you have a good enough enough number of human versions (idk this number yet), you can use those to create dozens more examples, refine, etc... then at some point you've thoroughly described a document enough that there's no more delta. Then you can use the document to train / finetune / infer.
The effects are that the "instincts" of the model (if on the training / weights) or the "thoughts" of the model (at inference time) are closer to a well vetted baseline.
I'm incredibly excited for this launch! Bolt Foundry is the first company to use [Isograph], which is a framework for building data-driven apps that are performant and stable out of the box. It does so by making heavy use of a compiler and of generated files, and essentially being an incremental computation engine for your UI. Your UI becomes performant by doing the least amount of unnecessary work.
And Bolt Foundry is truly the perfect first adopter! As "A big tech dev experience" implies, they're very focused on providing great DevEx. And a bunch of them previously worked at Meta, where they used Relay, and understood the powers that it provides.
Isograph aims to provide an even better developer experience (see the YouTube link), and give you the things that previously were only feasible at FAANG-co scale: e.g. even in a very dynamic app, fetching just the JavaScript and data you end up using.
So far, they've had a great experience using Isograph!! You should also [give it a try] or [contribute].
[Isograph]: https://www.youtube.com/watch?v=sf8ac2NtwPY
[give it a try]: https://isograph.dev/docs/quickstart
[contribute]: https://github.com/isographlabs/isograph
Interesting! I think the idea of a "tone" in your content being enforced by a sort of... prose linter makes a lot of sense.
You can assume 1 person with at least some practice in writing will have a clear tone and presence in the things they write, but this would be rather useful for corporate communications, especially when simpleton engineers (such as myself) try to write blog posts for a company engineering blog. If there was something nudging me towards the style of the blog, without steam rolling me constantly by rewriting everything, I would like that.
For the CMS, (I think you're building a CMS?) I think building around THAT use case would be an awesome selling point. Let content folk spawn special ephemeral links that let someone write up a draft without having to set up an account. Have some sort of approval process in there too, that could be configured on that "magic link", so I don't impulsively publish to everyone.
Make it super easy to just stumble into writing something the company can actually communicate with, right away. There's a lot of passionate people at companies, with 0 writing experience that would lend a lot of authenticity to blogs/engineering comms/recruiting ads, if they they had the bumpers in their bowling lane.
precisely.
that’s exactly where we’re headed.
thanks!!
What's so difficult about just writing a blog post yourself, in your own writing style? What's so wrong about pure human expression?
nothing at all. it’s not just about blog posts though, it’s figuring out what to write about, finding the best comments of fans who are smart and interesting without spending your entire day doing that (or hiring a team).
it’s not about having it write for you, it’s about scaling who you are to do lower value tasks.
For your consideration, the main page you've linked to is nearly unusable on mobile, might want to look into that.
https://imgur.com/a/n1zVO1n
I can't read that and pinch to zoom has been disabled.
We will fix right now.
The tweaked typography is looking better!
You'll need to toss on
and add a small amount of padding around the body to make the text scaling pixel density agnostic though.I would also remove the overflow:hidden on body, and remove the fixed height on #root. Nothing needs to be a scrollable container here, the viewport can just scroll over the body. That's causing the bottom ~2 paragraphs to be "stuck" under the bottom edge of my iPhone 13 mini in safari. Ideally use `min-height:100dvh`.
Just shipped it. Will get the padding going too.
I think this is interesting in sort of a scifi way because if you can establish a reliable enough characterization of "tone" then maybe you can then create a steganographic side channel out of deviations from that tone: so your adversaries see inane blather about something unimportant, but your comrades can decode the blather into more meaningful messages. Meanwhile, others are hosting your communications for free because what is actually cyphertext appears to them as engagement.
Cool idea.
With the constant deluge of AI slop everywhere, your voice is basically all you have.
This is basically a tool to train a computer to "steal" your voice. I find that sad.
Why can't we figure out how to use this technology in something other than creative tasks? I don't care what anyone says, a blog post showing off some technology or accomplishment should be a creative task, otherwise what's the point?
for me, i’ve only been able to tell my stories at scale with the help of a lot of infrastructure. as a solo creator, i’ve never had the impact i did in local news or at cnet or engadget.
it’s not about having it write for you, it’s about having it help you write better.
writing is thinking, but 90% of writing is editing.
I'm not sure how, but the text is miniscule on Safari on iOS and causes reader mode to need to scroll horizontally by nearly double the page width.
I am going to slog through it but I would recommend checking on mobile Safari in the future.
Haven’t dug deeply into it, but it appears to be CSS that doesn’t play nicely with the Safari rendering engine on mobile. With only five devs, maybe none of them have Apple devices?
> ...maybe none of them have Apple devices?
I mean it is possible. But another comment mentioned it's broken on Android mobile too.
They do seem to be working on it though which is great.
The text on the site and the text in the post is the same though.
Super weird behavior for me too. When I double tap to zoom into the text, I get some weird nested scroll bar state and can only read the first few paragraphs.
Reader mode is fixed, working on the meta tags now. Thanks for the patience.
This is great! Assuming the LLM versions of me will build their fake consumer clones believably. Once this goes FAANG-co scale it will only be a couple of months until we can all go public and Advertisers will stop buying ads putting an end to the enshittification of the interwebs. Genius. Who came up with this? ... and is Apple likely to sue considering their patent from 2012?
um what?
That did not seem cohesive.
I am all for new and better content management systems, and like all geeks written at least two of them, but what is our plan for the CMS?
Making a Twitter card does not seem that relevant?
Am I stuck on the wrong CMS interetation?
Why/How will this be different than WordPress? (or any ohter)
there’s a few pieces… the big one is the identity card stuff playing into the whole process.
like i said earlier in the thread maybe this is a bit too early, but we’ll have stuff to share over time.
sorry it’s confusing… working on it.