What’s True About AI | cosmastech

I used to love programming. I still do, but I do it a lot less now.

Things feel different. Everybody is talking about AI, and there’s a cloud that hangs over the field of software development: Are our jobs going away? Are we no longer differentiated by the things we spent so long learning? Do Boris Cherney, Pete Steinberger, and all the other AI thought-leaders know something we don’t? Or are these people just living in an alternate reality where tokens are free and everything AI does is lovely?

I wanted to write down a simple list of ideas that have been running around in my head. These are things which I would love to call facts, but are likely just observations and opinions.

Humans are sometimes bad at code

I’ve read and written plenty of bad code. Poor abstractions, unnecessary defensiveness, re-inventing the wheel, security vulnerabilities: you name it, I’ve probably done it.

I’ve read code and not understood it. I’ve read code and not realized that I didn’t understand it until much later.

I’ve misunderstood requirements, implemented code changes, and then had to rework a lot of it after release.

AI is sometimes bad at code

All of the same things humans do, AI does. It does it faster, requires an external LLM provider, and costs money per turn.

Writing instructions in English is hard

Using English to prompt a stochastic genie to model business goals feels like trying to add pepper to a pot of soup, except I’m standing two stories above the kitchen. Sometimes I add too little, sometimes I add too much, sometimes I totally miss my target.

I can’t read all of this

Ask the model something. Terminal window fills up several times over. Scroll up. Read a little bit. Feel overwhelmed and just accept that the model is probably right.

Models are confident, overly verbose, and love jargon. I have observed this internal defense mechanism: “I don’t understand this. I must be too stupid to understand this. I don’t want to be or appear stupid, so by agreeing, I can bypass that.” This internal conflict leads to more slop than anything else.

Bonus points: humans can do this too, I just hadn’t noticed it so starkly before. When I hear someone say “domain,” I immediately have no idea what they’re talking about. Is this a bounded context? Is this a feature of the product? Is this simply an area of code? Nobody knows.

I am suspicious of all human writing

I don’t know if people were always saying “it’s not X, it’s Y,” but I’m hyper-aware of it. When I see it, I assume it was written by AI. I am starting to look for the fingerprints of LLM generated text everywhere. It’s exhausting to be suspicious all the time.

Using AI to write messages for humans sucks

If you couldn’t be bothered to write the message yourself, it signals to me that you don’t care about me or what you’re saying.

I’m tired of docslop. There’s so much reading we’re expected to do now; the cost of writing giant documents in Google Docs, Linear, JIRA, or Notion has gone to zero. I’m dubious that people who generate giant documents read every line of them. Why? Because I have generated giant docs but not read them. It was easy.

I don’t believe that people receiving the documents are reading the documents either: I think they’re asking their agent to summarize it. The pipeline of written communication is person A -> person A’s agent -> person B’s agent -> person B .

Different feature areas feel totally different

Feature areas within an existing codebase have differing risk tolerances. Using AI to write or modify code that touches payment processing? Sweating bullets. Using AI to add a brand new feature or product? Fuck it, we ball.

People having such vastly different experiences with agentic development, either pro or con, are likely the result of the above. Additionally, codebases having:

low cohesion

high coupling

poor or inconsistent architecture

mounds of outstanding tech debt

conflicting agent markdown files

poor test coverage

undocumented/tribal knowledge

a programming language or paradigm less represented in the training set

will naturally have a worse agentic development experience than a codebase which is the opposite.

Adversarial code reviews are powerful

Whether I write the code myself or an agent does, having two separate models review the code has been a win, hands down. Remember, both AI and humans are sometimes bad at code. Multi-model code reviews add more Swiss cheese to the process.

AI doesn’t learn

This is painful. Every time an agent goes off the rails, we want to add a new skill file or modify our CLAUDE.md or AGENTS.md, ending up with scar-tissue markdown that eats tokens and probably just confuses the LLM.

Either that or I’m continuously writing in my prompt what I want the agent to avoid.

AI doesn’t understand

Con man is short for “confidence man,” someone who abuses their victims’ trust by projecting confidence. AI is a token generation process. In my experience,...