Things were widely discussed, ideas were shared, I was convinced. Later, I read more opinions and experience reports, and I was convinced otherwise!

The discourse is sometimes wrong, no matter how intuitively right it seems. Superficial isn’t super-right.

Anthropic, OpenAI, etc. intentionally weaken their product

If you find yourself thinking, “are the models worse today?”, it’s possibly time to step back and look at how you’re using them. Which is not to say it’s a “skill issue”. More likely, your setup has drifted away from how the models are trained to work within their harnesses.

The models (probably) don’t vary much day to day, inference and harness bugs aside. You likely want to reduce the size of your prompts, skills, and CLAUDE/AGENTS.md. Let the agent do its thing without oversteering and verbal micromanagement.

That said, the frontier labs have been known to vibe code rigorously evaluate and release code changes which result in elaborate bugs that have generated worse results, poor reliability, increased costs, and/or all of the above. Sometimes, you gotta look at the status page and see if the carnage is on their end or yours. I do not like this subtle gaslighting-adjacent quality of coding agents, but seems like we’re stuck with it until the labs compete on robustness.

The big labs are diluting their models for operational/performance reasons

This is a restatement of the previous one in the context of the next one.

There are a bunch of compelling stories that might make you think the labs are intentionally diluting their models. Because they’re afraid of Mythos’ capabilities or OpenAI’s models escaping containment. For operational expense reasons like reducing cost of inference. Because they’re just dishonest, pathological liars.

It is again, unfortunately, more likely that human error, and human error magnified by coding agent support, are interacting in ways that make it seem like the models are consistently worse than they were last week or last month. But really, it’s good old-fashioned bugs compounding to the extent they plausibly resemble malice.

AI is toast once the subsidies (VC money) runs out

We should enjoy the discounted token party while it lasts, right? Hence, all the double-usage weekends and “xyz model is available on all plans for the next several days” promotions.

I’ve come around to the mental model that reduced token cost is merely one lever in a sophisticated pricing model and subscriber retention strategy. I’m of the mind that the token party will probably continue, if you don’t mind hitting a usage limit every once in a while.

That said, the labs increasingly charge per-token consumption pricing of more customers and use cases. If token subsidies are important to you, get good at using first-party CLIs like Claude Code. This is currently true for Anthropic and their incremental but obvious restriction of how subscribers can use their tokens. Seems inevitable that other agent vendors (OpenAI, Cursor) might go this way, despite the good will they’re accumulating right now in not doing so.

I’ll just tell the coding agent to never, ever do the Extremely Bad Thing

Determinism is (still) a thing you can only get by writing helper scripts in a deterministic language. Or, by trusting that the little box you’ve put your agent into really impermeable. (Which it seems isn’t even true for the most clever of laboratories.)

Apparently, there’s a magic word, “gates”, that matches the domain language of machine learning to invoke guardrails. I’m not sure how effective it is. Eventually, even text like “you must never get the gremlins wet” is just another bag of floating-point numbers to the model. They’ll eventually get the math wrong.

I will never learn a new thing again

Sounds like a personal choice to me.

Granted, this is a bit of a straw-man argument. The stronger form of this argument gets back to the essence of what software development, creative work, and many kinds of knowledge work really are: problem-solving searching for an outcome that your team has decided upon.

The puzzles solved along the way are often means and not the end. For a lot of us, there’s pride in knowing the solution to the old puzzles and bringing those ideas to bear against new puzzles. Working with AI bypasses the rote application, and that will always jostle some feathers.

That doesn’t mean an agent can’t help one learn to solve new puzzles or fit that knowledge into the gaps in your experience. Ask the agent to help you learn how to solve/write the thing. Instead of just writing the thing for you and worrying about skill atrophy. Or, ask it to help you get started and remind you about the syntax or functions to use.

The little gizmos are pretty good at putting together explanations. At checking them for hallucinations against what a database query, or linter, or a web search, actually say about an assertion. At putting that all together as a haiku, sonnet, or graphic novel.

Now is the time to either celebrate you never really figured out the syntax for shell scripting, with all its fiddly sigils, or ask an agent to make a bespoke tutorial for you to finally learn shell scripting and all its spooky action at a distance.

Will I solve the same kind of puzzles I did before? Possibly not. Will I solve problems as a craftsperson of expressing yourselves to computers and people? Probably not.

Am I going to learn things? Only as much as I want to. (Which is quite a bit.)