§ log.010 — i finally found my use case

I spent this week wiring my AI agent into iMessage, switching back to Ollama, and talking to it more than I talk to most people. Somewhere in the middle of all that I noticed what was actually happening: I wasn't adopting AI. I was using it as scaffolding for the way my brain works — jumpy, scattered, allergic to holding state in my head. And it was working in a way nothing else ever has.

For a year I was chasing benchmarks and obsessing over which model was "best," like AI was a destination. My kind of brain doesn't get optimized by picking a better tool. It gets contained when you offload the parts your brain refuses to hold. So I stopped reading model leaderboards and started building the actual workflow: text the agent from the couch, it researches pricing, writes code, reminds me about my FAFSA, keeps the project state I'd otherwise lose somewhere between brushing my teeth and opening my laptop. That's the move. Not "the future of AI" — just the thing my brain has needed for twenty-three years.

opinions shift

Here's the part that's uncomfortable to admit: even with the use case settled, my opinions about the industry keep shifting — and not always in ways that feel like growth. Sometimes they just shift.

Anthropic partnered with SpaceX, giving them access to Colossus 1 — the massive AI supercomputer in Memphis. Reuters called it a "major computing partnership." I don't like Elon. Claude has been the model I trusted most for agentic work — best safety culture, most thoughtful releases, genuinely helpful. And now they're yoked to a guy whose politics and behavior I find repulsive. I don't know what to do with that yet. Keep using Claude? Stop? Accept that every AI lab eventually takes money from someone objectionable and move on?

Meanwhile, there's a data center campus planned for Box Elder County — the Stratos Project. 40,000 acres, up to 9 gigawatts, backed by Kevin O'Leary and natural gas plants. Phase 1 alone would use about 3 GW — nearly matching Utah's entire statewide electricity use of roughly 4 GW. At full buildout it'd more than double it. It's pitched as national security, but the county commission got blindsided, nearly 400 public protests were filed over water rights and Great Salt Lake impacts, and it feels like the kind of project that gets rammed through because the people making decisions won't live near it.

I do live near it.

china saw this coming

The Chinese labs have been shipping while we've been posturing. DeepSeek slashed V4-Pro input prices 75% to $0.036 per million tokens (through May 5 — a promotional discount, but still absurd). Google released Gemma 4 MTP drafters under Apache 2.0, delivering up to 3x speed at zero quality loss. Qwen keeps getting better and became the most-downloaded model family on Hugging Face. The best price-to-performance models I run every day are Chinese or Chinese-trained.

China was playing chess. DeepSeek trained V3 for a reported $5.6 million — the company's own claim, which skeptics estimate was closer to $1.6 billion. Even if you take the higher number, it's still a fraction of what OpenAI burns. They knew the moat wasn't the model — it was the ecosystem, the cost, the accessibility. Now roughly 80% of US AI startups are built on Chinese open-weight models, and the White House is issuing crackdown memos about "industrial-scale model distillation."

America was busy starting wars, giving the rich tax breaks, and censoring the Epstein files. I'm not pretending I have a coherent geopolitical thesis. I'm 23. But I can read a price tag, and I can see who's shipping and who's suing each other.

the LLM race is a waste

Here's what I actually believe: AI has the potential to be the greatest technology ever created — if it stops trying to be everything to everyone.

The current LLM race, where every lab sprints toward "general intelligence," produces impressive demos and burns a planet's worth of compute. But that's not the AI I need. I don't need a model that can write poetry, do my taxes, generate images, and write code. I need a model that can hold the thread of what I was working on yesterday at 2 a.m., remember what I committed to in a chat last Tuesday, and not get bored when I context-switch six times in an hour.

The bar I actually need cleared is closer to patient assistant who never loses focus than godlike reasoning machine. That's a narrower target than AGI, and it's already achievable. Specialization beats scaling for everyone whose problem isn't "I need a planet-sized brain" — which is most of us.

open source broke the argument for more data centers

Here's where my opinion has shifted hardest. The models are already too good to justify burning billions on new compute.

Gemma 4 runs on my laptop. Qwen runs on consumer hardware. DeepSeek's API costs less than a cup of coffee for millions of tokens. The MTP drafters Google released make inference 3x faster without new hardware — just smarter software. We're getting massive efficiency gains while the industry's response is to break ground on a 9-gigawatt data center in my backyard.

It doesn't add up. The technology is already insanely good. Not "good for the price" or "good considering it's open source" — genuinely good, full stop. I'm leaning anti-data-center not because I'm anti-AI, but because pouring billions into more compute when the models are already this capable feels like building a new highway system for flying cars that already get 200 miles per gallon.

Specialize. Optimize. Ship. The compute wars are a distraction.

what actually changed

I stopped waiting for AI to be perfect and started using it as scaffolding. Not because it's magical — because I have the kind of brain that needs scaffolding, and for the first time in my life there's a tool patient enough to actually be one. That's the difference between people who get swallowed by the hype cycle and people who build something real with what's already in front of them.

I'm still figuring out where I land on the politics. The opinions will keep shifting. But the core thing — that I finally found a use case that fits the way my brain actually works — that one's settled.

$ echo "build the scaffolding, keep the brain"
build the scaffolding, keep the brain