From: janburse@fastmail.fm   
      
   Hi,   
      
   No joke, I have a new CLP heuristic coded,   
   and made a blind test. Explaind to the AI   
   what I coded, but not how coded it,   
      
   omitted the crucial scoring function, and   
   the AI was exactly proposing this scoring   
   funcion out of the blue.   
      
   I was quite bamboozled. The AI could not tell   
   me how it recalled or halucinated the formula,   
   I spent like 1-2 hours trying to find a   
      
   precedent paper. Will blog about it later.   
   So whats going on . Why can AI suddently   
   propose Heuristics. Because they are now   
      
   differently trained:   
      
   Phase 1: Early LLMs (~2018–2020)   
   Phase 2: LLMs with Prompted Reasoning (~2020–2022)   
   Phase 3: Scaling & Reasoning Engineering (~2023–2025)   
      
   Training an AI to look for heuristics,   
   can help in self programming. The generative   
   part of the AI would generate some code,   
      
   which then the AI would use for itself. See also:   
      
   How it started:   
      
   AI for Low-Code for AI   
   Nikitha Rao, Jason Tsay, Kiran Kate,   
   Vincent J. Hellendoorn, Martin Hirzel - 2023   
   https://arxiv.org/abs/2305.20015   
      
   How its going:   
      
   Successor-Generator Planning with LLM-generated Heuristics   
   Alexander Tuisov1, Yonatan Vernik2,   
   Alexander Shleyfman - 2024   
   https://arxiv.org/abs/2501.18784v4   
      
   Bye   
      
   Mild Shock schrieb:   
   > Hi,   
   >   
   > Good Morning Vietnam, the HPC-AI Convergence   
   > doesn't sleep. Here a friendly reminder of   
   > the Sudoku leader board (Topn87 Challenge):   
   >   
   > #1: jczsolve / Rust WASM   
   >      Solving 87 Sudokus in 0.006 seconds (0 sec / Sudoku)   
   >      https://emerentius.github.io/sudoku_web/   
   >   
   > #2: Kudoku / JavaScript   
   >      Solving 87 Sudokus in 0.043 seconds (0.0004 sec / Sudoku)   
   >      https://attractivechaos.github.io/plb/kudoku.html   
   >   
   > #3: Picat / import cp. solve([ff],L)   
   >      CPU time 0.175 seconds   
   >      https://picat-lang.org/   
   >   
   > #4: Picat / import sat. solve(L)   
   >      CPU time 0.373 seconds   
   >      https://fmv.jku.at/kissat/   
   >   
   > Tested on Windows 11, with a AMD Ryzen AI 350   
   >   
   > Didn't test yet GNU Prolog, ECLiPSe Prolog or   
   > Ciao Prolog. So whats next? Well beat #1 by   
   > tapping into an NPU of Copilot+ PC.   
   >   
   > Have Fun!   
   >   
   > Its Winner Winner Chicken Dinner time again...   
   >   
   > Bye   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)