I don't understand why this continues to be a concern. It is only of academic interest, it seems to me, to wonder how good native LLM reasoning is. Yes, there is a small cost to dispatch from the LLM for MCP. Tokens need to be generated, and tokens need to be absorbed. But for anything that is deep reasoning it is a vanishingly small overhead. It is exactly what computational scientists would do too. They'd reach for their Matlab or at least a chalkboard. Let the LLM write the Lean 4, the Answer Set Programming, the Mathematica, the Matlab, the Magma, whatever. If they do a bad job, there will be correction cycles, if they do a good job, there won't. But frontier models are good coders now.
-----Original Message----- From: Friam <[email protected]> On Behalf Of glen Sent: Thursday, May 21, 2026 8:13 AM To: [email protected] Subject: Re: [FRIAM] More meat for the metavores What is "reasoning" if not a sequence of tokens, the latter depending on the former in some way? I'd like to offer up 3 links that might help us understand where the "reasoning" of LLMs is only kindasorta reasoning: • https://logicalintelligence.com/kona-ebms-energy-based-models • https://github.com/SkyworkAI/Matrix-Game/tree/main/Matrix-Game-2 • https://github.com/facebookresearch/vjepa2 The 1st one isn't quite like the other 2. But it's in the same vein, I think. There's some kind of something to be said about cumulative puzzles or meta-games. But I don't know quite what I'm trying to say. Although I loathe the term, Systemic Games <https://the-artifice.com/systemic-games-philosophy/> comes to mind. Reasoning engines have (at least) 2 modes, maybe akin to Kahneman's systems 1&2, where some input simply clicks or doesn't and is tossed away, but other inputs *modify* the lattice ... change the game. I say "at least" because there's a distinction between something like self-modifying code - where an execution can modify, add, or delete axioms or even the language - and "emergent play" where nothing fundamental changes, but one plays games atop or within the base game. So I guess there are at least 3 modes. All 3 are appropriately called "reasoning". But along with the gist of Hullman's post, failing to distinguish them is lazy. But we need generalized, non-jargonal nicknames for them, otherwise every mention requires a detailed glossary ... or perhaps an entire, pickled runtimeworldclosure, attached to every message passed. On 5/14/26 8:54 PM, Roger Critchlow wrote: > https://statmodeling.stat.columbia.edu/2026/05/14/as/ > <https://statmodeling.stat.columbia.edu/2026/05/14/as/> > > "Previously it didn’t feel like such a crime to talk about intelligence or > learning in machines because nothing really worked that well, so the labels > were clearly aspirational. But now it’s much easier to believe the simulacra. > And so it becomes harder to tell when we are using human-oriented terms as a > predictive convenience versus a scientific claim versus a marketing device." > > > "Too much casualness with words is unscientific. There was no good reason in > the first place to call the token sequences a model produces when we ask it > to “explain its reasoning” reasoning, other than that’s what we wish we could > see." -- 8647 ⊥ ɐןןǝdoɹ ǝ uǝןƃ ὅτε oi μὲν ἄλλοι κύνες τοὺς ἐχϑροὺς δάκνουσιν, ἐγὰ δὲ τοὺς φίλους, ἵνα σώσω. .- .-.. .-.. / ..-. --- --- - . .-. ... / .- .-. . / .-- .-. --- -. --. / ... --- -- . / .- .-. . / ..- ... . ..-. ..- .-.. FRIAM Applied Complexity Group listserv Fridays 9a-12p Friday St. Johns Cafe / Thursdays 9a-12p Zoom https://bit.ly/virtualfriam to (un)subscribe http://redfish.com/mailman/listinfo/friam_redfish.com FRIAM-COMIC http://friam-comic.blogspot.com/ archives: 5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/ 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/ .- .-.. .-.. / ..-. --- --- - . .-. ... / .- .-. . / .-- .-. --- -. --. / ... --- -- . / .- .-. . / ..- ... . ..-. ..- .-.. FRIAM Applied Complexity Group listserv Fridays 9a-12p Friday St. Johns Cafe / Thursdays 9a-12p Zoom https://bit.ly/virtualfriam to (un)subscribe http://redfish.com/mailman/listinfo/friam_redfish.com FRIAM-COMIC http://friam-comic.blogspot.com/ archives: 5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/ 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
