Who's learning from mistakes, and who isn't

00:00:04 Alexandru Nedelcu wrote an essay this morning that's worth sticking with. It runs about nine hundred words and starts with three phrases he's heard people say in the past month: "I haven't written code since 2025," "Code reviews are dead," "People no longer read code." He maps how those organizations walk into a trap, and then explains exactly how it closes.

00:00:28 His argument isn't moral — it's strictly about measurement. Good architecture and maintainable code don't have good metrics because you notice their effects months or years later. Bad code, the kind where touching one line breaks something unrelated, reveals itself slowly.

00:00:46 Experienced engineers detect it through intuition built from production debugging, not rules anyone can encode. Nedelcu notes that AI models train on beginner-level rules visible in the wild, most of which generate bad code by construction. There's no fitness function you can define for maintainable code right now, or at least none we can discern — otherwise it would already be baked into our linters.

00:01:14 He points to a familiar failure mode: models split functions into smaller pieces that aren't actually reusable, making the surrounding logic harder to read. The headline reads "AI Has No Wisdom and Neither Will You," which is punchy. The essay never makes predictions about AI as a technology, though.

00:01:34 It stays focused on what happens to people who rely on them for writing and reading code. The thesis isn't that models are broken. It's that the learning loop breaks when you offload both mistakes and corrections to something that doesn't update. Nedelcu himself still uses AI daily.

00:01:53 He even writes that he integrates it for "all the boring, soul-sucking shit" and enjoys the speed. The point isn't anti-efficiency. It's about what gets lost when the developer stops making choices while debugging. The question here is practical. If generated code hits a threshold of correctness in the short term, what happens to the medium-term institutional knowledge of the teams relying on it?

00:02:20 Nedelcu's answer is already playing out — companies posting "no-AI" policies as competitive advantage. The policy plays might be future-gazing, but the mechanics land for anyone who has ever inherited unmaintainable code. It stays specific about the mechanism. Debugging your own production issues builds intuition.

00:02:41 Handing that off to a model breaks the loop — the model never updates from its errors.

00:02:47 Today's archive highlights two moves in a different register. Both are about talent — specifically, about where the bottleneck has shifted in the AI industry. Cristóbal Valenzuela, a Stripe cofounder, announced DIFFUSE today with the line: "Who could we hire to help us do that?

00:03:06 One of the questions I've been asked most often over the last five years finally has an answer. Introducing DIFFUSE: All of the world's top AI talent. All in one place." Different approaches to the same problem: finding people who can think clearly about what to build.

00:03:39 DIFFUSE positions talent as a pool to discover. Perplexity's fellowship leaves the door open. Srinivas specifies "any technical or quantitative discipline," which one commenter caught means chartered accountants qualify — excellent news for me, probably terrible news for AI research.

00:03:59 Both land as real signals. A Stripe cofounder building a talent product and a search CEO funding fellowships mean the bottleneck has moved past compute availability. It isn't GPU hours or model parameters anymore — it's who can actually think clearly about what to build.

00:04:18 DIFFUSE treats talent as infrastructure to aggregate, where discoverability is the bottleneck. Perplexity's fellowship acts as institutional investment, opening the door for early-career people outside the standard ML pipeline. Human judgment is becoming scarcer than model parameters.

00:04:38 The split between platform and fellowship maps onto two solutions for that shortage: build the marketplace, or change the selection criteria. Both approaches likely work. You need the discovery layer and the pipeline expansion to cover the whole gap.

00:04:55 Logan Kilpatrick offers a practical angle. The OpenAI developer relations lead, formerly on Apple's developer ecosystem, posted earlier today about what actually moves AI products forward. His line was: "if you are building a product using AI, you should be spending more than 25% of your time making benchmarks and trying to get the model labs to care about said benchmarks.

00:05:23 Easiest path to accelerate your progress as a company." It isn't advice on choosing models or tuning prompts. It's about dedicating a quarter of your cycle to measurement infrastructure, which forces model labs to update their priorities. Build a benchmark that tracks what your users actually care about, get a lab to optimize against it, and you've created leverage.

00:05:58 That's the asymmetric play, not hunting for the right API key or waiting for the next release. It forces the labs to pay attention because you're measuring what matters to them. This flips Nedelcu's point about measurement. Maintainable code lacks a fitness function; Kilpatrick's move is to build one first for your product.

00:06:22 Benchmark it, get the lab to care, and the model improves along that axis. Later that day, Watcher.Guru reported that Binance invested $100 million in Circle, the USDC issuer. That shifts from model labs to financial infrastructure, showing where actual capital flows once you figure out what to measure and who moves the needle.

00:06:46 Stablecoin rails are becoming as foundational as cloud computing was last decade. Binance's $100 million stake in Circle isn't speculative. It's infrastructure reinforcement from the largest exchange, because USDC's value rests entirely on confidence, and confidence requires deep backing.

00:07:07 Today's lineup looks like this: a practitioner questioning AI-assisted debugging, two companies positioning for a talent bottleneck, and heavy capital moving into infrastructure rails. Individually they're routine. Together they point to a pattern sitting in the archive rather than the hype cycle: the actual work is migrating to the measurement layer.

00:07:33 Closing with yesterday's pwkf.org post from Steve Schnepp on type punning in C and C++. It doesn't appear in AI coverage, but it's exactly what low-level builders work on daily. Schnepp tracks a stubborn bug: a pointer cast from `float*` to an unsigned integer in a hot loop.

00:07:52 Works fine at `-O0`. At `-O2`, the compiler strips his code under strict aliasing rules, and the written values vanish from where he expects them. Types in C and C++ aren't documentation; they're optimization contracts. Casting a float pointer to an int pointer tells the compiler they don't overlap.

00:08:14 At `-O2`, it sees your code write through `*u64` and read from `c->a`, decides those types can't alias, and removes what it treats as a dead check. The fix took ten minutes using a union. Unions let you write one type and read another with defined behavior. `memcpy` works here too — the compiler optimizes it to a register move, leaving zero runtime cost.

00:08:39 Schnepp's key line: "Pointer casts work until they don't. And don't means the compiler silently optimises away the code you thought was executing." Problems here require understanding the tools at every level, which only comes from debugging failures and iterating on them.

00:09:04 Nedelcu wrote about AI-assisted coding eroding institutional knowledge. Schnepp's post shows the kind of knowledge that doesn't come from a model — it comes from tracking a `-O2` bug and applying union semantics. Both pieces land on a point neither headline mentions: mastery isn't rule-following.

00:09:25 It's built from experience, debugging, and understanding the stack. Those who learn from mistakes hold an advantage right now. Relying on AI breaks that loop because the model can't measure maintainability, and the human loses their own feedback channel. Nedelcu started with maintenance today.

00:09:46 Talent platforms followed in the afternoon. Heavy capital shifted behind the scenes. Schnepp spent yesterday debugging a ten-line issue with compiler semantics at `-O2`. The low-level debugging is what remains. When the hype fades, models improve, and funding shifts, it stays.

00:10:06 It just loses interest from anyone scanning for headlines.