I thought this answer was interesting enough to share, even if you're not pro-AI (I am ambiguous about it everywhere but technical tasks.)
\nMe: Something I'm curious about: it's well-known that LLMs have difficulty with things like math problems or even simple counting. What I'm wondering is why there isn't any mechanism for, like, when someone asks a math question, you hand off the solution to a different math-focused processing subsystem and pass it back to the language focused system. Do you get what I mean? Like the way we think the human mind works, or Minsky's \"society of mind\".
\nClaude: I get exactly what you mean, and the good news is: that mechanism does exist, it's just not built into the core model — it's bolted on as \"tool use.\"
\nWhen I get a math question, I don't compute it in my \"language\" weights if I can help it — I can write a snippet of code and actually execute it, then read the result back and use that. That's precisely a society-of-mind handoff: language-subsystem recognizes \"this needs precise computation,\" delegates to a symbolic subsystem (a real calculator/interpreter), and incorporates the answer. Same idea with search, code execution, calculators embedded in apps, etc.
\nWhy isn't this the *default*, built into the architecture itself, rather than an add-on? A few reasons:
\nThere is research in the \"mixture of experts\" direction that's architecturally closer to what you're picturing — different sub-networks specializing and a router picking which one handles a given token — but those experts specialize on learned statistical patterns, not on human-legible categories like \"math vs. language,\" so it doesn't give you the clean handoff you're imagining. The tool-use approach turned out to be the one that actually works reliably, even though it's less elegant than a true internal division of labor.
\n","wordCount":509,"charCount":3140,"replyCount":0,"reactCount":0,"reactPreview":null,"reactSummary":null,"shareCount":0,"image":"file:6a73364b500e3c4121c7a8e4@kwln.social","attachments":[],"tags":[],"originDomain":"kwln.social","createdAt":"2026-08-05T13:10:35.928Z","updatedAt":"2026-08-05T13:12:28.828Z","id":"post:6a73364b500e3c4121c7a8e8@kwln.social","url":"https://kwln.social/posts/post:6a73364b500e3c4121c7a8e8@kwln.social","server":"@kwln.social","summary":"I thought this answer was interesting enough to share, even if you're not pro-AI (I am ambiguous about it everywhere but technical tasks.)
\nMe: Something I'm curious about: it's well-known that LLMs have difficulty with things like math problems or even simple counting. What I'm wondering is why there isn't any mechanism for, like, when someone asks a math question, you hand off the solution to a different math-focused processing subsystem and pass it back to the language focused system. Do you get what I mean? Like the way we think the human mind works, or Minsky's \"society of mind\".
\n","textPreview":"I thought this answer was interesting enough to share, even if you're not pro-AI (I am ambiguous about it everywhere but technical tasks.) Me: Something I'm curious about: it's well-known that LLMs…","signature":"HsdpnsC3/yxYT91iih1ogU+4sO0AEIeqXXyeakWQfj4Vw73D+Qb76T3DHazzhW1gwz6+3GJsfazHo1hlLWFlJt+eXkJbWLBJBGdFMnFFco+USV/6DZbLcoskCl4fXBl2dgBV76AwEHwvwwGFiJbxxFdqSDTwyu1AlebY5kqfi21T1o9Gc8FmBsNGDM3sTxl+VMsvKEV4J8PA6YLVbAHRIDhTad58jmv4v7A4ywO7IXIV45UvHXPwvV7WSTj/dOKWUpoQ4C/bF3fq1NuH8OX37WdZNQ2OrZcsvXjtaAx9EUhopWJGPt9NEXe1Ny6jIL+Rlwh9SxzZtzKCfsf74WlUdw==","canReply":false,"canReact":false,"publishedAt":"2026-08-05T13:10:35.928Z","visibility":"Public","featuredImage":"https://kwln.social/files/file:6a73364b500e3c4121c7a8e4@kwln.social?v=1785935435964","myReact":null,"reactCounts":[]}