In short, I will not be integrating LLMs into my workflow for the foreseeable future.

My objections at this time are a combination of ethical, strategic, and practical concerns.

First, I have ethical objections to the wholesale scraping of all human art, literature, and software. In particular, open-source code must inherently be available to the general public, leaving it widely available to be scraped for training, giving it a disadvantage against proprietary software. Many models can and will regurgitate large portions of that training data, with no credit or royalties to the original source. Though it would be absolutely hilarious if a court ruled that ChatGPT was violating GPLv2, regurgitated code is likely free of the original license terms.

This whole situation reeks of theft, even as current law allows it to continue. Meanwhile, the big AI companies call for government intervention when other companies use their AI’s outputs to train a new model. I cannot comprehend this level of hypocrisy and entitlement.

Second, I have strategic objections to the level of control being handed to a small handful of technology megacorporations. I have a strong desire to control the tools that I use. Given a choice, I strongly favor open-source over proprietary tools. Given a choice, I strongly favor a one-time purchase over subscriptions. Given a choice, I strongly favor locally-operated to cloud-hosted services. Recent increases in subscription fees and cost per token are a perfect example of why this type of control matters, as are recent attempts to block AI services across national borders.

I do compromise the above where needed. At time of writing, hosting of this website is currently through Dreamhost, using open-source site-generation tools. My willingness to compromise depends mainly on how important a given tool is to my life and my livelihood. Because software development is extremely central to both; I will not violate core tenets except under duress.

Third, I have practical concerns about sycophancy and confabulation. As many have noticed, LLMs tend to praise the user (“You’re absolutely right…”) and confidently present misinformation and falsehoods (aka “hallucination”). I have my suspicions about the root cause for this, possibly a combination of bias in the RLHF process combined with inherent limitations of models based on next-token prediction. In any case, this problem hasn’t been solved by any of the major AI firms despite massive efforts; they may never be able to do so.

Finally, I have practical concerns about code quality. From conversations with others who have embraced AI tools for software development, it’s clear that current tools need significant hand-holding and code review to get good results. It’s not clear to me how much these limitations are inherent to the technology and how much are linked to the confabulation problem, but that’s someone else’s problem. For my customers, reliability and quality tend to matter much more than features and schedule.

I am attempting to remain rational about this. (Though bad behavior of big AI firms and obsessive media coverage sometimes make me want to pick up a pitchfork and shout, “Thou shalt not make a machine in the likeness of a human mind!”) For now, I am continuing to watch LLM and AI developments. If I can find an ethically-trained and locally-hosted model with good quality code output, I will give it a try. Until then, Keppytronics remains AI-skeptical.