Noise as News
Pull up your feed and count the AI release posts from the last sixty days.
Anthropic shipped Claude Opus 4.7 on April 16, plus Project Glasswing the week before, plus the Claude Design and Claude Security announcements that flanked them. Four major drops in roughly thirty days. Then OpenAI shipped GPT-5.5 on April 23, just 49 days after GPT-5.4 on March 5. Six weeks between flagships from the company whose previous GPT-5 launch was supposed to be the destination, not a waypoint.
Each one comes wrapped in the same packaging. Smartest model yet. Biggest leap forward. Agentic. Frontier. State-of-the-art. "Our smartest and most intuitive to use model yet." "Notable improvement on the most difficult tasks." The press releases are being written by the same robot at this point.
This is noise as news.
The major labs are locked into a release cadence that has nothing to do with what users need and everything to do with what the capital markets need. Six-week cycles, naming schemes that look like patch notes, decimal points doing heavy narrative work. 4.6 to 4.7. 5.4 to 5.5. I'm not saying these aren't real upgrades. Opus 4.7 is genuinely better at long-horizon coding, GPT-5.5 closes real gaps, the benchmarks move. But ask yourself, honestly. When was the last time a model release changed what you actually did on a Monday morning, on the work in front of you, with your team?
For most of us the answer is that it didn't. The model I'm using today and the model I was using sixty days ago accomplish the same things in my actual workflow. The newer one is faster on a benchmark I don't run, smarter on an eval I'll never see, priced the same per token but burning up to 35% more tokens because the tokenizer changed under me. The releases aren't for you. They're for the next funding round.
The heretical position: current LLMs are good enough. Good enough to draft, summarize, write code that compiles and passes tests. Good enough to read a transfer credit equivalency document and tell you whether MATH 1310 maps to your College Algebra requirement. Good enough to extract structured data from a messy PDF and power the agentic workflows real organizations need to deploy this year.
The marginal improvement from Opus 4.6 to 4.7 is real. The marginal improvement to your actual business outcome is approximately zero, because the bottleneck was never model quality. It was your data, your workflows, the institutional muscle to use what you already had. The labs need you to believe the next model is the one that finally unlocks it, because if good-enough is good enough, the entire trillion-dollar capex story breaks.
Here's the part nobody on the release-day livestream wants to discuss. A medium-sized data center consumes up to roughly 110 million gallons of water per year. The large ones, the AI-focused hyperscale facilities, drink up to 5 million gallons per day, equivalent to a town of 10,000 to 50,000 people. Inference is on track to dominate water use with 80% of total AI water consumption by 2026. That's not training. That's serving queries. People asking the model to write their LinkedIn post. To pick a baby name. To roleplay as a girlfriend. To rewrite an email they already wrote.
We are evaporating freshwater out of cooling towers in drought-stricken regions so that hundreds of millions of people can outsource thinking they were perfectly capable of doing themselves.
I'm not anti-AI. I lead a small team of five: very talented, senior engineers, product devs, and a unicorn PM that ships like a team of thirty. I'm writing this in markdown in Obsidian with Claude open in a sidebar. But I refuse to pretend the externalities don't exist, or that the answer to "what is this for" is AGI, sometime, eventually, you'll see. The cost of AGI, if we even get there, if "there" even means what they say it means, is being paid right now. In megawatts and aquifers, and in the slower erosion of human purpose as more people offload the small acts of cognition that used to constitute being a person.
Better LLMs? Good enough already. AGI? For what, at what price, paid by whom.
This is where I plant the flag. The release-cycle narrative obscures that the most important shift in AI has already happened, and it has nothing to do with which lab is two points ahead on Terminal-Bench this week. The shift is that coding agents have made it economically rational for organizations to own their data, their workflows, and their AI surface area, instead of renting them from someone else.
A year ago this was the art of the possible. A small institution couldn't compete with a SaaS vendor's R&D budget, so you rented. You bought the CRM, the SIS, the ERP. You contorted your organization to fit the vendor's model of how an organization works, and you paid for "best practices" defined by someone who had never met your team. Six months ago it was improbable. You'd heard of teams shipping real systems with AI assistance, but it sounded like a stunt. Today it's probable. Not certain. Probable. For any team willing to do the work. A small, disciplined team with a coding agent and a clear mental model of what they need can now build the system that fits their actual work, rather than the system a vendor needs to sell to a thousand customers at once. And the part the cynics miss: we hardened and secured as we went. This isn't shadow IT. It's institutional capability rebuilt on a foundation we actually own.
What we got on the other side: we removed the dependence on a vendor-defined CRM and put in its place a team engaged with their own data, empowered by AI to work the way they always wanted to work, not the way a vendor told them to. That last part matters more than anything else in this essay.
The dirty secret of enterprise software is that it doesn't make you better at your job. It makes you legible to a system designed by someone who isn't you, for purposes that aren't yours. You spend half your week translating reality into the shape the software expects. The software doesn't serve the work; the work serves the software. AI localization breaks that loop. When the tools are shaped by the people doing the work, the work gets done. The data stays close to home, the institutional knowledge compounds inside the institution instead of leaking into a vendor's training corpus, and the team, the actual humans, get to bring their wisdom forward, amplified rather than replaced.
That's the audacity these coding agents afford us. Five people shipping like thirty. Not "the AI will do your job." The AI will help you finally do the job you've been trying to do all along.
When the next Opus drops, or the next GPT, or the next Gemini, do this. Don't read the announcement. Read your team's actual work from the last thirty days. Ask what got unblocked, ask what didn't, ask whether your data is yours, ask whether your workflow looks like you or like a vendor's product roadmap.
The signal isn't in the release. It's in the room you're already in.
Everything else is noise as news.
Tired of pretending the release cycle is the story? Let's talk or join the conversation in Discord.