🦜

this is canon to me.

And I also believe it is impossible to have a conversation about the technology in isolation because its negative impacts are being felt by *literally* everybody on the planet. Its harm is *inextricable* from its development.

The reality is technology never solves social problems (hey guys check out the open source decentralized social network we built), and the AI bubble is entirely driven by a pathological hatred of labor costs, not a genuine benefit to society.

And to be fair to all this, the world where this stuff is actually helpful to *my* work would be a better one than where I am, but we are not there and the pressure to use it is where there is a big disconnect between day to day coders, manager types and academics are, where AI disdain breeds

Especially in firms where the business isn't software or technology

if the organization _has_ tests, or has the behavioral practice to specify software healthily, which again, is my point that the technology doesn't work well without an already healthy organization. It's a catch 22. I believe this is likely more common than modern devs really know.

I suppose I can only hope, but given this is a linguistic problem and some of the practices used in legacy COBOL are enthusiastically anti-reader for bewildering reasons, I'm pessimistic.

My biggest issue is they will continue to hallucinate confidently in ways that are already very hard to determine as false positives when you already lack the necessary context from a human side to understand what's going on, regardless of your actual skill level.

My experience tells me that many organizations with legacy code problems do not have the context necessary for these models to function correctly. Especially when dealing with very dated stacks like COBOL on mainframes.

The very customers that stand to benefit most from LLM AI cannot practically make use of it because of organizational problems that simply cannot be solved with technology. But that doesn't stop proponents from insisting it will.

I have made several attempts to make the agentic stuff useful and have found it more trouble to correct the mistakes than it's worth, because those reasoning models can't function without context and legacy software rarely has the context necessary on-hand due to poor documentation+tribal knowledge