I’ve come to the conclusion that LLMs by themselves don’t make software less stable. What they do however is amplify Murphys Law to a level most developers have never seen. I think we’re going to see a real difference between software “development” and “engineering”.

@dreamsofcode_io Agents are generating more tech debt in shorter time period so it takes a lot of effort to keep the outputs acceptable.

@dreamsofcode_io two types: one stops when the software does what is expected the other doesn't stop there

@dreamsofcode_io If you aren't careful they easily inject instability. One of my favorite idiosyncracies is how common it is for LLMs to decide to create a custom CSV based database instead of using the existing database. So if it's not deleting the database it's constantly trying to sidestep it.

@dreamsofcode_io I would say agents amplify developer traits. If you’re not security, organization, or speed conscious - your code base won’t be either. AI is exposing the good, bad, and ugly of every engineer and company out there and that looks like Murphy’s law. It’s a mess.

@dreamsofcode_io Your post is very correct, and it will be very interesting to see who can and does step up on fix the issues that LLMs will create. I recently wrote this article about the importance of senior engineers in this current and upcoming AI world: shorturl.at/Rno6d

@dreamsofcode_io I feel like unexperienced engineers make software less stable. If you have no clue what you are doing, you get disaster of an app. Au contre, if you are very experienced - less problems would be encountered. Simply because you can always stop LLM when it is doing bullshit.

@dreamsofcode_io garbage in, dumpster out ahould be new thing we use. Agreed that AI amplifies things to max level

@dreamsofcode_io Sometimes being a 1x dev has its benefits. Pushing more code by itself won't make the world a better place. Pushing good code will

@dreamsofcode_io That is a useful framing because it suggests a design strategy: relentlessly ask “what can go wrong?” and iterate until the answer is acceptable (based on product requirements).

@dreamsofcode_io Probably not, gonna go at the speed of publicly available models from the frontier companies.