Re: unboxed structure fields (or class slots)

===================================================
From: Marco Antoniotti <[email protected]>
Date: Tue, 13 Jan 2026 07:09:00 +0100 
Plus, awful Common Lisp code….

Marco Antoniotti
https://dcb.disco.unimib.it
===========================

1) Yes, that was surprising, where did such code come from? (cf. “swords made of damp cardboard”).

The point is that LLMs do not appear to learn from their own prompted inventions at all.
They take the final IN grid, and extract the answer from their primary training data.


2) The 'arc-transformer' function is intended as an example of 'inductive logic programming' (ILP).
Perhaps Prof. Marco Antoniotti’s lab would be interested in following/collaborating on this project.

https://x.com/fchollet/status/2009657443396227307
https://jobs.helsinki.fi/job/Helsinki-Postdoctoral-Researcher-in-Logical-Reasoning-and-Machine-Learning-%28ARC-Challenge%29/1349801957/


3) Finally:
===================================================
From: David McClain <[email protected]>
To: Tim Bradshaw <[email protected]>

But I honestly don’t see a path for corrections to their database (i.e., Perceptron weight matrices) based on daily
interactions. Perhaps a local database is held whereby future interactions in the same login account will receive some
corrections along the way?
==========================


A possible partial answer may be found here: https://arxiv.org/abs/2601.06851  ;; (01.11.2026).

'While our results provide compelling evidence for a synergistic core in LLMs, several limitations
of our current approach warrant discussion. First, our analysis focused primarily on attention
heads and Mixture-of-Experts (MoE) modules as the fundamental units of information processing.

However, Multilayer Perceptrons (MLPs) constitute a significant portion of transformer
parameters and are hypothesized to store factual knowledge. Future work should apply
information decomposition techniques to MLP layers – potentially extending recent
methodologies which may reveal additional or complementary synergistic structures.'



p.s.
Sorry for merging threads, just trying to make it brief.



Kind Regards,
Adrian W. Pasieka