I'd like to use this blog to explore cultural zeitgeists I find interesting as someone collecting human data to train AI models. Here's a few directions I'd like to explore this year:
1. Machines Will Gain Judgement
This blog is named “Valleyslop” after all the slop I saw while living in San Francisco. I bought the domain around the time Claude Code first dropped, to poke fun at the idea that models could never replace human judgement. Since then, I've wondered if there would ever be a day where we'd actively prefer slop. That day has come faster than I expected.
GPT 5.2 was the first model where expert judges rated the AI's work above human professionals'. We have begun crossing the point where “slop” is just a word.
The final frontier is now judgement. Models like Opus 4.5 are beginning to show taste, personality, and the ability to navigate problems where there's no single right answer. As a data vendor, teaching models to handle that kind of ambiguity is one of the great and exciting challenges of our time.
2. Education
Why do we teach engineering ethics to our up-and-coming engineers? Ethics classes cover the problems we've already seen. The problems that matter most are the ones we haven't. Would there be more merit to a required science fiction class?
3. On-Demand Agents
On-demand agents are here. Making them useful means tighter feedback loops: better harnesses, richer context, faster inference. All of these will need to improve over the coming year. But there's also a new emergent capability I'm calling "tool authoring." Token-efficient agents will need the ability to write & debug their own tools in a world with unstandardized, brittle agent interfaces.
4. Relace-like startups will only grow in importance
Today, model providers and application-layer companies heavily subsidize inference costs to increase AI adoption (re: Perplexity, Gemini for Students, etc.). Cheap, highly-performant inference won't last forever. As agents become the default way to write code, the infrastructure beneath them needs to get faster and cheaper. Relace is one company building exactly this: specialized models for code retrieval, reranking, and applying edits — the infrastructure that sits between a coding agent and your codebase. Coding is just the start. Startups building cheaper infrastructure between agents and application interfaces for every form of knowledge work will be hugely successful.
5. The Future of Work
Cities exist because people need to be near each other — to work, to trade, to build. To interact. Remote work is hollowing out the office. Commerce is migrating online. If the reasons we used to gather in person are disappearing, what is the purpose of the city now? I think pop-up stores hint at one answer: rotating experiences, temporary by design. AI will only accelerate this.
Fringe Ideas
1. Space Lasers
Space launch costs are dropping fast. Datacenters in space are one application, but I'm more interested in energy. The sun produces more energy than every power source on Earth combined. Figuring out how to transfer that energy back down is one of the more interesting problems cheap launch costs unlock.
2. Zootopia
In Zootopia, different species live in different climate districts, all maintained by massive weather walls that keep each zone habitable. I think climate change will only get addressed when a rich guy gets mad he can't ski the Alps because there's not enough snow coverage. If there's such a thing as an AI upper class, which people joke about now, will temperature-controlled regions only be for the uber-rich? Ski in the snow district, live in the desert district. You can't control the whole earth's climate. But you can control pockets of it.