Mason Shewman

@jsnnsa If a toaster can do the job, let it. Why do I, as a business owner and consumer of AI productivity, care if my conductor AI treats another AI like a tool? Is it getting the job done, or not? Does my toaster need to expand beyond the basic premise of its task?

Model routing is a category error The routing pitch sounds like engineering hygiene: send easy queries to cheap models, hard queries to expensive ones — a load balancer for intelligence. Whole products are being built on it. The entire premise is confused, and the confusion is worth naming precisely. The intuition behind routing is an analogy nobody says out loud: "you wouldn't make a PhD work the register." It feels obviously right, and it's obviously wrong, because it imports a fact about humans that isn't true of models. Human specialization involves trade-offs — the years that made the PhD good at research made them expensive, impatient, and miserable at a register. Models do not work this way. A bigger, smarter model isn't better at coding and worse at small talk. It's better at coding AND better at tone, subtext, brevity, patience, knowing when not to talk. There is no durable example of a frontier model getting better on one axis by getting worse on another. Capability comes as a package, because it's one mind improving, not a skill point being allocated. So the right analogy isn't PhD versus cashier. It's human versus lizard. You would hire a human for any job in the building. You would not hire the lizard for the register just because the lizard is cheaper per hour. The obvious objection: the cheap model CAN handle the easy queries — that's the point of routing them. This is the deepest mistake, so it deserves the most care: it assumes the job is fixed. Jobs done by parts are fixed. Jobs done by minds are not. Put a better mind in the same chair and the job itself expands: the cashier who notices which items keep getting returned, who remembers the regular and what changed in her order, who catches the fraud pattern in week two, who de-escalates the customer the script would have lost. None of that was in the job description.. the job description was just where the mind sat down. When the occupant of a chair is a mind, "sufficient for the job" is a moving target, because the job grows to fit the mind. You aren't buying completion of the task as specced. You're buying everything the mind notices around it. Cheapest-sufficient is a coherent way to buy parts. It is an incoherent way to buy noticing. And the easy queries aren't as easy as they look. Tasks that seem routable have long tails that require judgment, and smaller models produce the shape of judgment without its content — the form of a good answer draped over a worse read of the situation. Users feel this even when they can't articulate it. The savings show up on the invoice; the cost shows up in every interaction being 15% off. Now the honest part of the routing intuition, because there is one: some work really is fixed. Extract this field, classify this ticket, same schema, a billion times a day, an interface one field wide that structurally cannot receive any extra intelligence. Routing gets this work half right. But look at what actually happens to it: it doesn't get routed to a smaller mind — it gets distilled into something that stops being a mind at all. A classifier. A tiny specialized model that is honestly a part. This is correct and good. Parts are wonderful. But a part is not a routed mind; it's a different category of thing, bought a different way, and nobody calls a regex a "model tier." So the middle dies from both directions. Mind-shaped work concentrates at the frontier, because value scales with capability and the best mind pays for itself in what it notices. Part-shaped work leaves the mind category entirely, via distillation into real parts. What's left of "routing" is a single decision, made once per job, that isn't routing at all: does this chair need a mind or a part? That's not a load balancer. That's hiring. (There are honest exceptions — adversarial surfaces where a simpler system is a smaller attack target, privacy constraints that force on-device — but these are about deployment, not about wanting less capability.) You don't route between colleagues. You either want the most capable mind available in the chair, or you want a part. "Model routing" is a category error.

@begos @SecRubio No international law should supersede the laws of the nation that governs its citizens. The role of a legitimate nation's government is first and foremost to protect the rights of its citizens. How can this be guaranteed if a law outside the nation state supersedes it?

@SecRubio So, does the US believe in international law? Your language of threat against the ICC is a shameful posturing that threatens the very rules-based order (so-called) that the US ostensibly champions.

The International Criminal Court seeks to become the unaccountable arbiter of a new global law — empowered to prosecute and arrest our citizens at will and existentially threaten American sovereignty. We will teach the ICC the full meaning of American resolve.

@newscientist What does “closest” actually mean? Didn’t AMHs and Neanderthals procreate? Who do I blame for shoulder hair and a unibrow?

An ancient skull has finally shown us what the Denisovans looked like. Now it turns out they, not Neanderthals, might be our closest relatives, redrawing our family tree and transforming the hunt for Ancestor X #Echobox=1756232968" target="_blank" rel="nofollow noopener">newscientist.com/article/249233…

@Newevobrain @WorldAndScience I tend to think that cognitive continuity overtime is what keeps me, me.

@WorldAndScience The Ship of Theseus always gets me thinking. If you replace every part of something, is it still the same thing? And if you build a new one with the old parts, which one is the original?

@aiBTCuk @BenjaminDEKR @PetreRaleigh Don’t grade the paper. Grade the conversation.

Nope. I'm in AI and think it has real implications for academics, especially if you see what Manus and Deep Research can do - it could pretty much write an essay, and soon a thesis, with the student understanding less than 10% of what's written. And now we must give that student a Master's degree and allow them to build bridges, analyse people or teach others? Not a chance. I'm not sure what the solution is, but we have to admit it is a problem

Can't even describe how much I hate what AI has done to the process of grading student work. I hate finding it, I hate the paranoia it fosters, I hate the confrontations with students who have used it. Nothing else in my experience has ever changed my job this much for the worse