Post B9IJmYpIZOWG4RQNgO by mntmn@mastodon.social
 (DIR) More posts by mntmn@mastodon.social
 (DIR) Post #B9IGPbizKQpyXJpuk4 by mntmn@mastodon.social
       1 likes, 0 repeats
       
       just writing something down that just occured to me, might be old hat, but to not forget it later:- llms are not turing complete, or at least their runtime is severly bounded (expensive)- there must be a limit to which they can reliably apply chained word/symbol replacement rules ("translations"), in one shot- same must apply for "reasoning" about call graphs, because that's modeling control flow or compilation, i.e. constructing an internal stack- overflow this stack
       
 (DIR) Post #B9IGX0uY5HoYg5m7zE by matthewcroughan@defenestrate.it
       0 likes, 0 repeats
       
       https://www.youtube.com/watch?v=ePsQksj99LM&pp
       
 (DIR) Post #B9IGX19nAc0XRMuIkK by mntmn@mastodon.social
       0 likes, 0 repeats
       
       @matthewcroughan oh interesting, never seen/read that
       
 (DIR) Post #B9IGleV5gky5W4qcMq by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @mntmnwithout an external tape, I'm not Turing-complete either :/
       
 (DIR) Post #B9IHKYHQo7qwMN6GW0 by jaseg@chaos.social
       0 likes, 0 repeats
       
       @mntmn I’d imagine the challenge would be to design a transform that forces the LLM to process its output directly, and that is resistant to the LLM discovering and reversing the transform.In the limit, this sounds similar to indistinguishability obfuscation, a cryptographic technique that while it has been around for a long time hasn’t reached anywhere near practicality yet.
       
 (DIR) Post #B9IHRUrRuyA7mbG6M4 by mntmn@mastodon.social
       0 likes, 0 repeats
       
       @jaseg by directly you mean without tool calling, right?
       
 (DIR) Post #B9IHXrZMOzOXwspjV2 by mntmn@mastodon.social
       1 likes, 0 repeats
       
       @wolf480pl but at least you have a real memory and self modifying code (neuroplasticity), so you can emulate that much better and with less energy (i.e. pizza and coffee)
       
 (DIR) Post #B9IHZGuzdnOQrAcjnU by jaseg@chaos.social
       0 likes, 0 repeats
       
       @mntmn I think if you want it to be useful you’d have to assume the LLM can call tools
       
 (DIR) Post #B9IIAlP9FwOW7gbdS4 by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @mntmnI get tired tholike, if I can't figure out the part of thebvall graph I need in a couple minutes, I give up
       
 (DIR) Post #B9IIlvZQQHbyMl8NLU by dngrs@chaos.social
       0 likes, 0 repeats
       
       @mntmn I chose violence:https://codeberg.org/spookyvision/yaml-interpolator/src/branch/main/AGENTS.md
       
 (DIR) Post #B9IJmYpIZOWG4RQNgO by mntmn@mastodon.social
       0 likes, 0 repeats
       
       @dngrs lol
       
 (DIR) Post #B9IKmuJ2oBpQdw4D8y by mntmn@mastodon.social
       0 likes, 0 repeats
       
       @jaseg yes, agreed! that didn't fit in the character limit of my OP :D
       
 (DIR) Post #B9IuFZljk0xJk2tRuS by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @tomjennings @mntmn Turing had plenty of paper to write on, maybe not in the shape of a tape but still
       
 (DIR) Post #B9L0Oc6NQP5Tcm9zM0 by bastelwombat@chaos.social
       0 likes, 0 repeats
       
       @mntmn This limit is well known by LLM users. In their language it’s called „context window“. It’s usually measured in „number of tokens“ and is one of the metrics used to advertise (bigger means better, but also more expensive).People usually avoid getting even close to that limit, because apparently these things get dumber, the closer you get to the limit. Practically all LLM-manuals talk about how to be resourceful about that limited context window.