Every AI optimization trick I learned this year has an expiry date

Last Tuesday I spent forty minutes deciding which model should handle a job. The job took four minutes.

I sat there weighing cost against quality, thinking about whether to hand the browsing part to a cheaper agent and keep the reasoning for the expensive one, whether to clear the context or compact it, whether to turn the output down to concise and save some tokens. Forty minutes of optimization for four minutes of work.

Then it hit me that I'd been doing this for about eighteen months, and that almost every trick I'd learned in that time was already dead.

The graveyard so far

Remember when you had to tell a model "you are an expert marketer with 20 years of experience" before it would give you a decent answer. Gone. Remember "think step by step". Baked in. Remember stuffing five examples into every prompt so it would match your format. Mostly unnecessary now.

Remember when getting clean JSON out of a model meant a regex repair loop and a retry counter. Structured outputs killed that whole category of code.

Remember tuning temperature like a knob on a stereo. I haven't touched it in a year.

Remember when a serious RAG setup meant arguing about chunk size, testing four embedding models, and bolting a re-ranker on top. Long context and native search ate most of that stack. The teams who built their entire technical identity around chunking strategy in 2024 are quietly rewriting everything.

Remember fine-tuning a small model for a narrow task because the big model cost too much. Frontier models now do that task zero-shot for less than the fine-tuning run cost.

Remember token golf. Stripping whitespace from system prompts, arguing about whether XML tags beat markdown, shaving characters to fit a window. I did that. It felt like craft.

What's dying right now

The ones I listed at the top are next, and I'd put most of them inside two years.

Model routing. Sending easy work to a cheap model and hard work to a frontier one only pays while the price gap stays wide. That gap has been closing every year. At some point the routing logic costs more to maintain than the tokens it saves.

Context management as a skill. Deciding when to clear and when to compact is me doing memory management by hand. That's the same job a C programmer did with malloc before garbage collection arrived and made the skill worthless. Context windows went from 8k to 1M in under three years. Models are starting to manage their own working memory.

Subagent orchestration for cost reasons. I split work so the cheap agent browses and the expensive one reasons. That's an accounting decision dressed up as architecture. Orchestration built for genuine parallelism survives. Orchestration built to dodge a price sheet does not.

Output length settings. Turning a model down to concise to save tokens is rationing something that gets cheaper every quarter.

We've watched this movie four times

Keyword density in SEO. People counted their exact-match keywords to two decimal places, and then the search engine got good enough to understand the page.

CSS hacks for Internet Explorer 6. An entire generation of developers memorized workarounds for one broken browser, and the skill evaporated the day it died.

Separate mobile sites at m.yourdomain.com. Whole teams maintained two codebases until responsive design made the second one pointless.

Manual query hints in SQL. Developers hand-tuned execution plans until the optimizer got smarter than they were.

The pattern repeats. A capability arrives with rough edges. Practitioners build clever workarounds. The workarounds become an identity, sometimes a job title. Then the underlying thing improves and the cleverness turns into overhead.

What this means if you run a company

Keep optimizing. Just be careful what you build on top of it.

Don't sign a three-year contract with a vendor whose entire value is cost arbitrage between models. That value evaporates on someone else's release schedule, not yours.

Don't hire a prompt engineer. Hire someone who can define what "good" looks like for your business and prove whether you got it. That person stays useful in every model generation.

Don't let your team build an identity around a workaround. The engineer who says "I'm the RAG guy" is one release away from a problem. The engineer who says "I know how to test whether this output is right" never is.

And don't confuse the tricks with the work. Cutting your token bill by 30% is a nice quarter. Knowing which part of your operation should never touch a model is a decade.

What actually survives

Evals survive. Being able to tell whether the output is good, at scale, without reading every one, gets more valuable as models get more capable, not less.

Proprietary data survives. Your customer history, your call recordings, your pricing exceptions belong to you alone.

Judgment about what to automate survives. Most failed automation projects I've seen failed at the decision, not the build.

Taste survives. Someone still has to look at the output and say that's not good enough, and know why.

Distribution survives. Better models make content cheaper for everyone at once, which makes attention the scarce thing.

I'll keep learning the tricks, because today's work runs on today's models. I'm just holding them loosely now, and I've stopped writing anything down that assumes the model stays this dumb.