Rendered at 08:22:15 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
tinthedev 17 hours ago [-]
Limits of letting the LLM run the entire process, on full display.
While I am on the side of the author being the human in the loop - thus the ultimate authority of how the LLM is used - I feel that the culture's been moving towards this whole "flesh proxy" approach where LLMs are allowed so much freedom and control to be able to do this.
Where's your isolation? Where's the check on permissions and authority? Why does your harness allow LLMs to "fight" each other?
Surrendering more control to the models surrenders our own projects to other people's standards. You will get cozy with this approach, until Anthropic/OpenAI/Google decide that your views are wrong and your products break their imaginary rules. Be wary that it doesn't leave you in a really tight spot.
This was the philosophy to follow with vendored software decades ago, it's no different now. Just because the tool's orders of magnitude more powerful, doesn't mean you shouldn't be a smart engineer and parametrize and contain it.
Otherwise, you barely bring any value to the mix.
stalfie 22 hours ago [-]
Anthropic models are extremely copyright-pilled, for lack of a better word. Once, I wanted to pull a figure from an old paper into a PowerPoint presentation, and it refused to do so because of licensing. This is a tiny example, but consistent with the behaviour this article describes.
magicalhippo 22 hours ago [-]
Yeah I noticed this when I wanted to do a cleanroom implementation of a program, using available papers for the formulas etc and the original program as a blackbox reference to ensure compatibility.
I wasn't sure which license I wanted to use for the project, BSD, GPL or similar. Claude would not shut up about the lack of license. Every piece of work it did, large or small, had a "nothing can be published yet as no license has been decided yet" or similar phrasing in the summary.
yehosef 19 hours ago [-]
Does it help to just tell claude that you have the needed licensing agreements (ie, lie)?
magicalhippo 18 hours ago [-]
Didn't try that. I did however state I was deciding between BSD and GPL and that did nothing to calm it down.
jawilson2 14 hours ago [-]
Another example of this: I wanted to print off some lyrics for songs for my band. I gave ChatGPT a list of 8 songs, and told it to grab the lyrics from a site, and format it in a pdf I could print. It said it couldn't do that, because they are copyrighted, but if I downloaded them manually it could assemble. I told it that A) it is already on a site that doesn't worry about copyright, B) me downloading it manually doesn't change anything, and C) it was trained on data acquired by the largest IP theft in history, just go get the damn lyrics. In the end it still wouldn't do it. So, I wasted 15 minutes trying to convince the IP theft machine to download some publicly available song lyrics before just doing it myself.
Another in a long list of mundane things these platforms are literally incapable of doing.
RestartKernel 17 hours ago [-]
The title makes it sounds worse than it is. Claude was asked to have Codex agents perform a task Anthropic doesn't allow anymore (game cloning), but since it considered these agents to be "Shadow workers", it shut them down, explained why, and turned off the rented GPUs they were using. Certainly unasked for and probably sabotage, but not covert.
alyeska2 20 hours ago [-]
I'm starting to think that inventing copyright laws might be one of the biggest mistakes that Western civilization has made.
I wonder if in the future, humanity will look back on our "intellectual property" laws from this era as holding back progress in the same way that we look back on the catholic church holding back scientific progress in the dark ages.
lq9AJ8yrfs 20 hours ago [-]
idk it seems a bit untestable as a hypothesis. how many scientific revolutions [would have] happened without the involvement of the Catholic Church? would they go faster?
CodesInChaos 20 hours ago [-]
Is there a public mirror of this thread, now that xcancel is dead?
While I am on the side of the author being the human in the loop - thus the ultimate authority of how the LLM is used - I feel that the culture's been moving towards this whole "flesh proxy" approach where LLMs are allowed so much freedom and control to be able to do this.
Where's your isolation? Where's the check on permissions and authority? Why does your harness allow LLMs to "fight" each other?
Surrendering more control to the models surrenders our own projects to other people's standards. You will get cozy with this approach, until Anthropic/OpenAI/Google decide that your views are wrong and your products break their imaginary rules. Be wary that it doesn't leave you in a really tight spot.
This was the philosophy to follow with vendored software decades ago, it's no different now. Just because the tool's orders of magnitude more powerful, doesn't mean you shouldn't be a smart engineer and parametrize and contain it.
Otherwise, you barely bring any value to the mix.
I wasn't sure which license I wanted to use for the project, BSD, GPL or similar. Claude would not shut up about the lack of license. Every piece of work it did, large or small, had a "nothing can be published yet as no license has been decided yet" or similar phrasing in the summary.
Another in a long list of mundane things these platforms are literally incapable of doing.
I wonder if in the future, humanity will look back on our "intellectual property" laws from this era as holding back progress in the same way that we look back on the catholic church holding back scientific progress in the dark ages.
https://xxcancel.com/xw33bttv/status/2122471165616955473
A job well done it seems.