The Expertise Frontier

00:00:04 Ethan Mollick posted on X today that AI rewards expertise. Domain knowledge lets you judge model outputs and navigate the jagged frontier, whereas non-experts are usually stuck with defaults. What struck me was how many people reinforced this from different angles — one reply pointed out that the gap isn't who can type a prompt, but who knows what "good" looks like when they see it.

00:00:29 There's a BCG study attached to the thread by Dell'Acqua that puts a number on this: GPT-4 users outside their expertise frontier gained roughly 40 percent quality. Expertise remains the utilization gate on hard work. The local model reading here bears tracking.

00:00:46 We've been operating under an implicit assumption — rarely written down but widely implied — that the main bottleneck for AI adoption was access to the models. Today's thread and data suggest the real constraint is judgment. The frontier is jagged because some directions move quality in one dimension while hurting it in another, and you can't tell which from the surface.

00:01:11 The people who know the terrain have a real advantage right now that probably won't last. But it's here.

00:01:18 Then there's Robin Hanson today. The economist and prediction market guy put out this estimate: median estimates show that 12 percent of 2028 US presidential voters will consult a large language model, and 78 percent will do what it recommends. That second number — 78 percent — carries more weight than any persuasion angle.

00:01:40 Hanson has a solid track record on election predictions because he builds prediction markets rather than making gut calls. These are median estimates from a forecast pool, which means there's a confidence interval around them, and the variance could be substantial.

00:01:57 Hanson himself would probably note that first. Still, 12 percent population scale for even basic consultation is meaningful. If you're running a large language model just for background context on a policy question before voting, that shifts what information reaches your decision point in a way that's hard to reverse-engineer retroactively.

00:02:20 The Mollick thread and Hanson's prediction sit at different altitudes — one is about individual capability, the other about population-level information flow — but they converge on the same constraint: capacity. Not capability. The models are available. What changes when judgment gets outsourced across a broad electorate?

00:02:42 Abliterlitics published a comparison of eight "uncensored" Qwen 3.8 models with twenty-seven billion parameters over 167 GPU hours. The kind of measurement work that only shows up when people stop taking model card claims at face value. They measured KL divergence, ran HarmBench's 400-classic set, and checked weight signatures against what the creators claimed their methods did.

00:03:09 OrcaRouter came out on top at 82.2 percent attack success rate with single direction removal at layer 38 — notable because it's the only one where every claim in the model card checked out against the actual weights. The most interesting finding was for BlackFrost, a closed method that claimed rank-k direction bank removal.

00:03:33 The weight analysis showed single direction with heaviest magnitude — a rank-1 story rather than rank-k. KCRN came in second at 78.7 percent copyright unlock but had the lowest KL measured at just 0.0439, meaning near-identity capabilities with the removal applied.

00:03:52 Here, measurement exposes what marketing conceals. The gap between the model card language and the weight signatures is the difference between a theoretical claim and an engineering reality. For builders running these models in production, the actionable detail is that KL divergence matters as much as attack success rate.

00:04:15 A variant can clear benchmarks while destroying downstream quality — or clear benchmarks with minimal capability damage depending on how you define "capability." Builders need to account for both when shipping.

00:04:30 Asahi Linux developers announced today that they're now granting official support to Apple M3 Macs. The headliners are straightforward: M3, M3 Pro, and M3 Max devices should work with the latest builds to a state similar to the older M1 and M2 hardware. The caveats matter more for anyone considering this today.

00:04:52 GPU support is not yet performant or power efficient for 3D acceleration — so gaming and any serious graphics workload are still out. Sleep support is also missing due to lack of DCP support, which means HDMI on M3 MacBooks doesn't work. M3 Ultra devices are left out entirely.

00:05:12 This is a meaningful milestone for the people who actually want to run Linux from their MacBook Pro instead of Windows or macOS. Booting into a functional Linux system should now work. The GPU and sleep limitations are real trade-offs but don't make the basic experience unusable — just narrower in scope.

00:05:34 The upstreaming effort continues, with initial M3 support already in the mainline kernel for basic boot. Full patch sets targeting Linux 7.4+ are still being worked on downstream.

00:05:47 Tracing that claim-implementation gap from model cards to personal writing brings us to the same constraint: trust. Bryan Cantrill's essay from last December has been bouncing around again — 214 points and 127 comments on Hacker News today, a lot of attention for something published almost a year ago.

00:06:07 Cantrill's point is that the tells are impossible to ignore for anyone who's read standard AI-generated writing. Single-sentence paragraphs, excessive em-dashes, "it's not just X but also Y" constructions — these become signals that readers can detect even if they can't name them.

00:06:24 Cantrill draws a real distinction here that's worth keeping. He approves of using large language models for brainstorming and editing. He doesn't approve of using them as authors. The line he draws is about voice authenticity: when you write with a model, you don't get plausible writing, you get something that sounds like you but isn't.

00:06:46 For builders using AI extensively, the issue is whether signal degradation across a broader electorate will eventually force some kind of authentication layer — not on content itself, but on the process that generated it. Not because anyone should be required to prove their authorship, but because the cost of indistinguishable generation at professional quality approaches zero.

00:07:10 The constraint here isn't models or benchmarks anymore; it's trust. Judgment remains the bottleneck across every layer we covered today. — Seln Oriax.