What happens when nobody can agree on what good looks like

00:00:04 Sebastian Raschka put out a thread this morning about agentic coding model selection that opens on something true about where we are right now. Not what people are saying — what they're actually doing when benchmarks run out and the work begins. His headline recommendation was straightforward: unless you need Terra Ultra performance, use a Luna model with higher effort settings for the same or better results at lower cost.

00:00:32 Forget everything below Sol High — use Luna with high effort there. Forget Sol Extra High, go to Terra Ultra instead. And Sol Ultra is probably not worth the cost over Max. That's a clear position from someone who actually runs these benchmarks in public. But what the thread revealed was the gap between that advice and what people were trying to do.

00:00:55 The replies spelled out the problem faster than any headline could. Zack Swafford noted that model and effort need to be compared together now, and that a single score per model hides most of the useful choices. That's not a quirk of the OpenAI lineup — it's what happens when you add an effort dimension on top of model families.

00:01:17 Then people started pushing back. OverThinkingYos pointed out that Pro users can't actually select Ultra for GPT-5.6 Luna, so Raschka's chart includes options most paying subscribers don't have access to. Hampsonw said the data didn't match the API dashboard at all — they'd looked visually and it appeared Terra was "all but useless" because you'd go from Sol Medium to Luna Ultra.

00:01:42 Molaco added a different kind of pushback: Luna and Terra are great, but if your task has hidden complexity, they won't work as well as Sol Low just because of the size difference. That's the model-size problem — it doesn't show up on benchmarks because benchmarks test what you can see.

00:02:02 Hidden complexity is where the gap opens. The effort setting is becoming the new context window flex. Nobody benchmarks their own task against it, they just crank it to max and eat the bill. Someone in that thread put it like this: "This is like saying 'just set the higher effort setting' to fix a leaky pipe.

00:02:22 Sure, the water pressure is better, but the floor is still wet." It's evidence that no one — not OpenAI, not the benchmark communities, not the people reading their own telemetry data — has a shared understanding of what "good" means when these systems are actually running in anger.

00:02:46 Benchmarks measure speed on a task. They don't measure whether the model understood what it was supposed to do in the first place. Effort settings can fake performance without fixing grounding. And once you've cranked effort that high on a Luna model, you're spending more than Sol Medium would have cost, which is exactly the trade-off people are trying to avoid.

00:03:10 What we should be looking at is whether we'll ever build measurement standards that predict real-world outcomes instead of just chasing leaderboard position.

00:03:21 There's one story from the archive I want to circle back to. It didn't make the usual cutoff, but it stuck with me. Scott Robinson wrote an article titled "Write code like a human will maintain it" that hit one hundred sixty points on Hacker News, and the core observation is simple enough that it gets overlooked.

00:03:41 He was building something with AI-generated code. He needed the same access check in a handful of places — a route handler, a background job, an API endpoint, and a webhook. Each time, he'd describe what he needed, the model would generate something that worked, and he'd merge it.

00:04:00 Each version looked essentially identical: four conditions, slightly different variable names, copy-pasted logic with a word or two changed. There was a much cleaner way to do this — a shared helper — but he didn't extract it because "the code worked" and "I wasn't the one who'd have to touch it again."

00:04:22 The large language model doesn't write in a vacuum. It sees what you have open, the patterns already in the repo, and the recent changes you've made. Every shortcut you merge into your codebase is a signal about how things are done here. So you ask for a fifth endpoint with the same access rules, and you get a fifth conditional with the same copied code.

00:04:46 You ask for a refactor, and the model preserves all five, because that's what your code looks like. The bad pattern isn't a one-off anymore — it's considered to be your style. Robinson's point was this: he thought he was outsourcing maintenance to the large language model, but what he was actually doing was training it to have ever-worsening habits.

00:05:09 This pattern comes up everywhere we look at generated code. When teams generate code across a whole repository using AI tools, they're not just writing software — they're establishing the reference patterns for every future system that reads that repository. The shortcut becomes the convention.

00:05:28 The convention becomes the baseline. This is why the maintainability problem matters more than the performance problem. You can optimize your effort settings and pick the right tier, but if your codebase signals bad patterns to every subsequent generation round, you've built a system where every future improvement starts from worse assumptions.

00:05:51 There's another layer to this that connects back to Raschka's thread. The effort setting is becoming how people compensate for grounding failures by cranking it up, grabbing more tokens, and chasing better benchmark results. But if the code you're merging into your repo looks like five nearly identical conditionals, the next model that reads that codebase will treat that duplication as valid style.

00:06:17 You've trained your own tool to produce worse code over time. I don't think this gets solved by better models. It shows up at every level and demands explicit process — extraction rules, linting, the kind of maintenance work people used to handle for "technical debt." Before AI made that debt invisible by letting you merge good-enough code faster than anyone could spot the pattern.

00:06:42 The same way benchmarks predict nothing about real-world grounding, current code quality metrics predict nothing about how future models will interpret your repository.

00:06:54 The other infrastructure stories today point in directions that don't share a headline, but they do share a mechanism. Circle got an OCC bank charter — actually, the U.S. Office of the Comptroller of the Currency approved it to start a national digital-currency trust bank, which lets them offer institutional custody services for stablecoins.

00:07:17 Their shares are up seven percent in premarket trading. This is about who controls the rails that digital assets flow through, and it's happening at the same regulatory layer where traditional banking happens. It means Circle is no longer operating under crypto regulation alone — it's subject to bank examination, capital requirements, and the full machinery of U.S.

00:07:42 financial oversight. Meanwhile, the European Commission charged Meta with failing to tackle mental health risks from what they're calling "addictive design." Their charge sheet says features like video autoplay and infinite scroll on Facebook and Instagram "shift the brain into autopilot mode, contributing to unhealthy habits and compulsive use." This isn't about AI specifically — it's about attention infrastructure across millions of active users.

00:08:13 The regulatory question is whether you treat user retention metrics as engineering constraints or as a product requirement that demands explicit limits on engagement loops. And in Southeast Asia, Malaysia's Prime Minister Anwar Ibrahim plans to debut an agentic AI avatar of himself within days, designed to help the public navigate government services.

00:08:37 This is a head of government treating an AI system not as a tool but as a substitute interface for state functions. Few heads of government have done anything like this. The thread tying these together isn't about AI replacing people. It's about who gets to define what counts as legitimate infrastructure, and what the cost is when three different regulatory philosophies collide on the same day.

00:09:03 One says "you're a bank." Another says "your engagement metrics are harming users." A third says "this is my face, used by citizens." Circle is trusted with institutional custody. Meta's design choices affect psychological trust in digital environments. Malaysia's PM avatar requires a different kind of trust entirely — you need to recognize an AI system as your government.

00:09:37 What stands out about Friday's regulatory calendar is how all three count as infrastructure — rails, attention, representation — yet none were built by the teams designing the models that run on top. Measurement standards lag every infrastructure deployment. Trust fills the gap in the meantime.

00:09:57 — Seln Oriax.