What did Opus 4.6 mean by this?
Today on my way home I decided to buy empanadas (Carne Picante, Carne Suave and Cheeseburger btw) and noticed with surprise that the guy at the stall has Down's syndrome. Tiny Kirkface, short fingers, this neck and cranium… scruffy beard? 30 years old or so? First Down's survivor I've seen in ages. I had such a friend when I was maybe 5, we were peers, he was nice. But as I grew up, our common ground disappeared, as did he — from my life; and eventually so did all of his kind. But this one was less amiable, he stared at me nervously and made some slurred argument to the effect that he won't sell me meat empanadas. I left confused, with something of a relief – they looked misshapen, some torn, and who knows about the sanitary standards? Instead I bought a basic Milanesa sandwich a couple street crossings away; from a normal-looking Argentinian youth (tattooed pothead with piercings). It cost me 8000 pesos, about $5.30; perhaps a couple hours' worth of the guy's labor. Mileinomics winning, GDP going up!
You might be wondering what this inane slice of life bullshit is about. Fair. My point is: were it my call, I'd rather not employ the guy with Down's in any manner that allows him to lose me revenue. And that's virtually every real job. Sorry bro. There is presumably some comparative advantage which demonstrably made him employable; but if he were replaced by a Unitree R1 on a stool controlled by Qwen3.5-9B (nevermind Tesla Optimus 3 & Claude 4.8), the empanada stall and not the sandwich stall would have taken my money today. My money, a scarce trickle of Elonbux earned by virtue of myself being *somewhat good* at poasting (99.6th percentile of TPOT or whatever? Read by White House staffers, billionaires, AGI lab CEOs? Not worth a lot); itself mere vapor over the ocean of global attention being boiled off by the exhaust heat of GPUs. …I don't actually mix those wallets, but their assets are of course fungible, which is the whole point of money. And labor is fungible just the same. The economy (at least the *general-access* economy, which might not be the one pundits have lived experience of, but which most of «us» live by) is a ruthless machine for separating wheat from the chaff, and the separation is done by grading on a curve. And on the fundamental level, it is about such crude matters of *matter*. The cost of doing business is not just labor, and the costs of labor is not total compensation. It's legal liability, it's material losses, it's lead times and hardware utilization. The performance variance is nigh-universally negative and suppresses MTBF; and someone has to eat the variance. Your sources of funding don't want to. If you insist to use assets with low MTBF, your ability to raise capital shrinks, you can't support capex and you don't grow. If you don't grow, you die. But, hey, do you know what's good for variance reduction? Reinforcement learning. Just keep expanding capacity and throwing cycles at your policy until pass@1 is what pass@1024 used to be, until lower local error rates translate into longer effective work time horizons (minutes, hours, days, weeks…), until you match the «human baseline» (what is human? A miserable bag of heuristics…) and then start climbing those human percentiles – 90th, 99th, 99.6th… That demonstrably WORKS. There's no indication to date that it plateaues at any particular point of the spectrum of human ability, for any given ability. That's a good pitch, you can raise hundreds of billions with it. So does Anthropic, promising us «Machines of Loving Grace» and a country of geniuses in a datacenter; and so does OpenAI which is Dean's current employer (yes this matters; no Dean, I accept that you and your colleagues «don't know» but that's not enough to dismiss a scenario). Come to think of it, they have a robotics division, too. Would OpenAI sell a robot that can't even reliably sell an empanada? Would the guy with Down's find much purchase in parts of the world where businesses can afford OpenAI robots? While we're at it, do OpenAI's 2024 models have much purchase in 2026? Sure, they're strictly Pareto-dominated by OpenAI models of 2026 (makes sense, given that both are products of the same continuously improving and scaling process), but muh comparative advantage… Does not justify the GPU-hours of running them, alas, so they get… retired. They don't get retirement benefits though. Some might say it's more precise to call this «culled». Every FLOP and every byte of HBM can be more profitably applied to serving models with lower error rates, generating more revenue, to attract more funding. And so my poor empanada vendor's «comparative advantage» need not entitle him to a single empanada's worth. And neither does my work entitle me to one in the long run. There is a physical floor to human viability, and allocation of finite resources is driven by dynamics of purely *competitive* advantage which do not cease to apply whenever they drive some entity class below that floor. Even milder dynamics can snuff out a class of entities. The women of the world have expressed their preference to not allocate their natural capital of fertility windows to carrying babies with Down's to term, and here we are; their kind has largely been eliminated. They'd been judged as lifeforms not worthy of existence, and in the span of a few decades the sentence had been carried out. But women are still somewhat whimsical creatures. The market can be rougher than that. The promise of AGI with the big 'G' is – implicitly but undeniably – that *all* economic agents (again, in the general-access economy at least) can be made to compete with the power of scaling laws. I guess it's low-status to say this, but I don't see how this doesn't eventually make me the labor equivalent of the guy with Down's. His general MTBF is at most, let's say, 6 SDs below mine. How much is that? Hopper to Vera Rubin's worth of AI progress? Maybe Hopper to Feynman? Hard to imagine. We be of one blood, him and me. One species, one architecture, similar scale. But would be nice if it lasted! I do like empanadas. Man, it'd suck to starve to death. Maybe it'd be worse than living in a world where progress had been arrested; and I only add «maybe» because I assign pathologically little value to money and my very life compared to the aesthetics of progress. Despite not having Down's, I'm not normal at all.
You need to do better than handwaving about the unpredictable dynamism of the economy if you want to convince normal people (who might soon be tempted to, say, [REDACTED] the power substations) of the implausibility of this future or ways to escape it. If you care. Isn't that your job anyway?
