Re: unboxed structure fields (or class slots)

Plus, awful Common Lisp code….


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


On Mon, 12 Jan 2026 at 22:29, wojciech.pasieka (as wojciech dot pasieka at
ai dot pressiton dot com) <[email protected]> wrote:

>
> The following exercise may help Martin decide if and how to implement LLMs.
>
> There is an ARC-AGI challenge, and the simplest task seems to be b7999b51.
> https://anokas.substack.com/p/o3-and-arc-agi-the-unsolved-tasks
>
>
> This is my attempt to evaluate LLMs' 'reasoning'.
>
>
> 1) LLM-s can't synthesize shapes from raw JSON data, so they are given a
> lists of shapes.
>
> The first IN-OUT.
>  (DEFPARAMETER input-1-facts '(((2 2 C) (2 3 C) (2 4 C) (3 2 C) (3 4 C) (4
> 2 C) (4 4 C) (5 2 C) (5 4 C) (6 2 C) (6 4 C) )
>                                ((3 3 O) (4 3 O) (5 3 O) (6 3 O) (7 3 O) (8
> 3 O) (9 3 O) (10 3 O) (11 3 O) (12 3 O) (13 3 O) (14 3 O))
>
>                                ((4 6 F) (4 7 F) (4 8 F) (4 9 F) (4 10 F)
> (5 6 F) (5 7 F) (5 8 F) (5 9 F) (5 10 F) (6 6 F) (6 7 F) (7 6 F) (7 7 F)
>                                (8 6 F) (8 7 F) (9 6 F) (9 7 F) (10 6 F)
> (10 7 F))
>                                ((6 8 N) (6 9 N) (6 10 N) (6 11 N) (7 8 N)
> (7 9 N) (7 10 N) (7 11 N) (8 8 N) (8 9 N) (8 10 N) (8 11 N) (9 8 N) (9 9 N)
> (9 10 N)
>                                (9 11 N) (10 8 N) (10 9 N) (10 10 N) (10 11
> N) (11 8 N) (11 9 N) (11
>                                10 N) (11 11 N) (12 8 N) (12 9 N) (12 10 N)
> (12 11 N) (13 8 N)
>                                (13 9 N) (13 10 N) (13 11 N))))
>
>  (DEFPARAMETER output-1-facts '(((0 0 F) (0 1 F) (0 2 F) (0 3 F) (0 4 F))
>                                 ((1 0 N) (1 1 N) (1 2 N) (1 3 N))
>                                 ((2 0 C) (2 1 C) (2 2 C))
>                                  ((3 0 O))))
>
>
> 2) Gemini generates the function which maps input-1-facts into
> input-2-facts correctly (with human help).
>
>  (DEFUN arc-transformer-1 (input-facts)
>   (LET* ((colors '(C O F N))
>          (objects (MAPCAR (LAMBDA (col) (REMOVE-IF-NOT (LAMBDA (p) (EQ
> (THIRD p) col)) input-facts))
>                                      colors))
>
>          (object-data (MAPCAR (LAMBDA (obj) (LET ((cols (MAPCAR #'SECOND
> obj)))
>                                               (LIST :color (THIRD (CAR
> obj))
>                                                     :width (1+ (- (REDUCE
> #'MAX cols) (REDUCE #'MIN cols))))))
>                                objects))
>
>          (sorted-objects (STABLE-SORT (COPY-LIST object-data)  #'> :KEY
> (LAMBDA (x) (GETF x :width))))
>
>          (final-output (LOOP FOR obj IN sorted-objects FOR row-idx FROM 0
>                           COLLECT (LOOP FOR col-idx FROM 0 TO (1- (GETF
> obj :width))
>                           COLLECT (LIST row-idx col-idx (GETF obj
> :color)))) ) )
>
>              final-output))
>
>  (DEFUN flatten-level (level list-1)
>   "Flatten possibly nested LIST a given number of LEVELs (or to the end)."
>  (LABELS ((rec (x  acc depth)
>    (COND ((NULL x) acc)
>          ((ATOM x) (CONS x acc))
>          ((AND depth (ZEROP depth)) (APPEND x acc))
>          ((rec (CAR x) (rec (CDR x) acc depth) (WHEN depth (1- depth))))
> )  ))
>           (rec list-1 NIL level) ))
>
>
> RESULT:  (arc-transformer-1 (flatten-level 1 input-1-facts)) ;; correct
>
>          (((0 0 F) (0 1 F) (0 2 F) (0 3 F) (0 4 F))
>           ((1 0 N) (1 1 N) (1 2 N) (1 3 N))
>           ((2 0 C) (2 1 C) (2 2 C))
>           ((3 0 O)))
>
>
> 3) Logically, this successful template should be reused for the next steps.
>    But Gemini ignores it and generates a brand-new function for the second
> IN-OUT task.
>
> (DEFUN arc-transformer-2 (input-facts)
>   (LET* ((flat-facts (REDUCE #`APPEND (REDUCE #`APPEND input-facts)))
>          (all-colors (REMOVE-DUPLICATES (MAPCAR #`THIRD flat-facts)))
>          (color-counts (MAPCAR (LAMBDA (col)  (LIST :color col  :count
> (COUNT col flat-facts :KEY #`THIRD)))  all-colors))
>          (sorted-colors (SORT (COPY-LIST color-counts) #`> :KEY (LAMBDA
> (x) (GETF x :count))))
>          (output-map `((H . 4) (F . 3) (S . 1)))
>          (final-cells `()))
>     (LET ((current-row 0)
>           (current-col 0)
>           (final-output (LOOP FOR color-entry IN sorted-colors
>             FOR color = (GETF color-entry :color)
>             FOR target-count = (CDR (ASSOC color output-map))
>             DO (LOOP REPEAT target-count
>                      DO (PUSH (LIST current-row current-col color)
> final-cells)
>                         (INCF current-col)
>                         (WHEN (>= current-col 4)
>                           (SETF current-col 0)
>                           (INCF current-row)))))
>
>       (LIST (LIST (REVERSE final-cells)))))
>
> This arc-transformer-2 function has a different, incorrect structure and
> fails entirely.
>
>
> SUMMARY:
> 1) Illusion of Continuity: Prompts create a fake sense of "ongoing
> thought."
> 2) No Reflective Thinking: LLMs do not review or validate their previous
> logic.
> 3) Data Over Logic: LLMs restart every task using training data, not
> established facts.
>
> There are attempts to bridge this gap via Recursive Reasoning Models and
> other concepts.
>
>
> I’m calling 2026 'the year of the CLI and REPL', defined by the history
> and beauty of Lisp.
>
> ___________________________________________________________________________________________
>
>
>   https://code.claude.com/docs/en/cli-reference
>   Command             Description
>   claude              Start interactive REPL
>   claude "query"      Start REPL with initial prompt.
>
>
>
>
> Kind Regards,
> Adrian W. Pasieka
>
>
>
>
>
> ---------- Forwarded message ----------
> From: "David McClain (as dbm at refined-audiometrics dot com)" <
> [email protected]>
> To: Tim Bradshaw <[email protected]>
> Cc: "Adrian W. Pasieka" <[email protected]>,
> [email protected]
> Bcc:
> Date: Mon, 12 Jan 2026 07:42:09 -0700
> Subject: Re: unboxed structure fields (or class slots)
> My experiences so far with LLMs is that they are essentially deaf to
> constructive criticism - by that I mean, they profusely apologize and often
> reframe the discussion usefully, but that the interaction of correction
> does not become assimilated into their long-term character.
>
> I suppose I could waste an hour to verify this by logging into another
> account and asking the original question again…
>
> 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?
>
>
>
> On Jan 12, 2026, at 06:30, Tim Bradshaw (as tfb at tfeb dot org) <
> [email protected]> wrote:
>
> On 8 Jan 2026, at 11:02, wojciech.pasieka (as wojciech dot pasieka at ai
> dot pressiton dot com) (Adrian W. Pasieka) <[email protected]> wrote:
>
>
> Gemini's says both: 'YES' and 'NO'.
>
>
> LLM gives useless answer: news at 11.
>
>
>
>
>