GenAI is both Popular and Unpopular
To no one's surprise, a lot of people using GenAI for anything get quite a bit of negative and sometimes abusive feedback. I've stated this other places, but want to share my feelings on the matter.
Art, Video, Audio, and Narrative Text
It's clear that all LLMs have used copyrighted material to train their models and when used for these types of output, it's not fair use. It's a clear violation of copyright. I did turn some of my own selfies into images using AI, but I stopped immediately after reading about the methods used to build those apps. There are only ten of these images, one of which heads this post. I did not pay for these images to be created so the app owner made no money from me, but there were in-app enhancements, so the app owner was trying to make money off of other people's art and photography. This is clearly wrong and should be illegal. The sad truth is that under this administration, a toothless congress, and a compromised judiciary, copyrighted material used in training large language models is going to get a fair use pass. It's not 100% yet, but with all of the money flying around, it would seem to be inevitable. I think efforts to embed training corruption into your digital art is now a necessity. Luckily this is easy to do.
Programming and Interactive Fiction Concepts
Since there is an enormous amount of open source and public domain code, it would seem that trained models ability to generate code is acceptable, but here we don't know all of the details. It's possible that Anthropic Opus 4 is trained on private or copyrighted code, but I'm not sure it would be necessary. There's just too much public code to bother. I primarily use Claude to generate code that I design, and there is an important distinction in the phrase, "that I design".
Interactive Fiction has a lot of open source code, though some of it is copyrighted and protected. There is also a lot of theory and deep examination of how to build an IF platform that crosses into public and private domain. Some may view my work on Sharpee as crossing that line. But at heart I am an IF hobbyist and have wanted to build my own platform for 40 years and never understood the moving parts well-enough to hand-code one. Claude, and specifically the latest Opus 4 with a full MAX subscription, has afforded me the ability to both learn about those moving parts and to construct a new Typescript-based IF platform. I am comfortable that I am not hurting anyone since the entire code base is on GitHub with an MIT license.
I'm also comfortable because I have introduced some interesting changes to the way we think about authoring IF. This will become apparent when I have a released and tested story built on Sharpee. Something as simple as using Jest to run tests against the code and your story is incredibly useful.
Some of those new things are not all that new. Traits and Behaviors are a traditional static object split into data and logic. Capabilities are actions with custom data. Using an Event Source is nothing new in the business world, but it is new in the IF world. Separating the Text construction from state change logic is also new to IF.
I strongly believe Sharpee will enable more authors to build bigger stories faster, if it's ever embraced (I'm not holding my breath...IF folks are moody 11yo tabby cats). (I love 11yo tabby cats)
If nothing else, it has made me enjoy building something for the last two years (off and on) and I have learned a LOT.
The last part is that none of Sharpee is possible without decades of knowledge of IF and system development. You can't just prompt, "Make an IF platform in Typescript." and get a complex, well-designed, tested platform. No AI is capable of that and is highly unlikely to ever have that ability.
No, AGI is not coming and we're not going to be enslaved by robots. The tech to enable thinking and reasoning has not been invented and no one knows how. You'd need a Positronic Brain from Asimov's Robot novels or Star Trek Next Generation and we don't know what that would look like or how to construct it.
I have been able to build Sharpee with many hundreds of hours of iterative design, generating code and plans, testing, refactoring, and restarting from scratch more than a few times. Even using GenAI you have to be knowledgeable, passionate, and dedicated to your purpose.
I hope to write something long about this process that includes a lot of the frustration with earlier models that would fall off a cliff when the context window became too large. I think I full stopped half a dozen times until Opus 4 + MAX + Desktop Claude + MCP.
In any case, for me, generating code using GenAI gets a pass. If you don't think so, I apologize. I think it's a massive game changer even though it does stupid things when you forget to pay attention.
Back to testing Sharpee. I think we're 90% ready for a full story. Maybe even 95%.