So, the most revealing detail here is the missing L2-to-KMem path, as it sounds almost absurdly specific until you realise that a wire which isn’t there can tell you more about a processor than another thousand MAC units. Sure, Apple built an accelerator where weights had a special status, they arrived from DRAM, entered private kernel memory and were expected to hang around while activations flowed past them. That is a perfectly sensible physical embodiment of the CNN workloads dominating mobile ML when the ANE appeared. Then transformers changed the useful lifetime and movement of those operands. Suddenly yesterday’s sensible optimisation becomes today’s awkward detour back through DRAM. That makes this a lovely example of technical debt occurring below software, where there’s no refactoring sprint coming to rescue you. Silicon necessarily contains predictions about future workloads. Cache sizes, interconnects and DMA paths are forecasts wearing electrical-engineering clothes.

The crystal balls here are the whole story; Habitat sold them for 18,000 tokens in one place and bought them for 30,000 somewhere else. Players noticed, did the obvious thing and quintupled the money supply overnight. The programmers had created a virtual world, then discovered that virtual worlds come with economists. That distinction between designing a game and governing a population feels more consequential than Habitat’s claim to being the first MMO. A conventional game designer can anticipate how a player encounters a puzzle. Habitat’s developers had to contend with people encountering each other, coordinating, specialising and finding opportunities nobody deliberately put there. A treasure hunt taking weeks to construct disappeared in eight hours. Latency that should have prevented coordinated performance instead produced people rehearsing with telephones and stopwatches. But there’s a useful design lesson buried in that 64-kilobyte machine; give people sufficiently expressive tools and the complexity of the system stops being proportional to the complexity of its software. It starts coming from the participants.

Right, the useful idea here is the promotion ladder: unverified, agent, deterministic. It gives teams somewhere sensible to put rules they understand well enough to care about but poorly enough that they can’t yet encode them. That intermediate state matters as LLM review is often discussed as though probabilistic judgement were the destination. Here it becomes scaffolding; let the model spot violations while the team is still learning what the violation actually looks like. Once the same pattern keeps appearing, turn the accumulated judgement into a boring script and remove an LLM from the loop. Boring is excellent when you’re deciding whether a pull request may merge. I’d be slightly more cautious about the claim that repetition means a constraint is ready for automation. Some architectural rules remain contextual however many examples you collect. A dependency boundary might be mechanically detectable; whether an abstraction has become conceptually dishonest probably isn’t. Premature hardening could fossilise yesterday’s judgement into tomorrow’s CI failure. Still, progressive hardening suggests a useful way to think about AI coding tools. Their highest-value output may sometimes be the specification for their own eventual redundancy. If an agent repeatedly performs the same judgement, ask whether you’ve learnt enough to replace that judgement with code.

“Human Reserved” is the idea I keep coming back to after reading this... Gates is proposing that some economically automatable work should remain deliberately human because efficiency isn’t the only thing employment produces. His nature-reserve analogy makes the trade-off unusually explicit: society sometimes chooses preservation even when development would produce measurable economic value. That’s a much stranger proposition than another retraining programme. Once machines can perform a job more cheaply and reliably, keeping humans involved becomes an affirmative social choice. We already make versions of that choice. Sport would be considerably more efficient if motorcycles were allowed in the Tour de France. Nobody wants that because efficiency isn’t the objective. Yet, the difficult part is deciding where that logic belongs in ordinary economic life. Gates reaches for caregiving and delivering terrible medical news, where the case feels intuitive. Move towards accounting, programming or driving and the boundary becomes politically explosive. Protection could preserve dignity and communities, or become occupational nostalgia paid for by everyone else. So, I suspect “Human Reserved” is useful even if it never becomes a formal policy category. It forces a question that discussions of automation routinely skip: once human labour stops being technically necessary, where do we still want it? That decision can’t be delegated to a benchmark, because the benchmark is precisely what has made the decision necessary.

Andrew celebrates that it rejected the asset because the required LOD structure was missing and that distinction runs through the whole project, as the objective isn’t merely sharing Rust across six platforms, it’s moving product assumptions into places where they can be stated, tested and rejected consistently. This is where cross-platform architecture often goes sideways. Teams optimise for maximum shared code and eventually build their own slightly miserable operating system. cubacadabra has a more useful boundary: share decisions, leave genuinely platform-specific execution with the host. Rust decides whether a username response is still valid; Swift or Kotlin performs the HTTP request. Rust owns rendering; Android still gets to ruin your afternoon by removing the surface. The unfinished morph pipeline makes the same principle visible from another direction. Arrow’s desired milestone is wonderfully concrete: the second hat should require a file, not a Rust change. That’s really a test of whether the architecture has discovered the correct boundary between mechanism and content.