Okay, this is the first time energy-based models actually clicked for me. I wrote a while back about moving from deterministic to probabilistic problem-solving, and honestly Bodnia's "can you define your rules?" says it better than I did.
The new approach is interesting but seems more complex than a standard BFS or DFS approach which may only cause 1/100 of the new approaches. Certainly, love the concept but the example is controversial. We may need a better and real examples to better illustrate the new approach.
Prior to LLM, there are tons of research like the new ideas in the article - most of these researches were paired with realistic problems in a more challenging spaces.
A model that returns no sentence is what I run every day, so the Sudoku result reads like home. Mine sorts public posts into rooms and returns a label with a probability, and the day I checked it against a language model on the same hundred posts, the funniest miss was a post saying "Thank you for giving me a good laugh" filed under AI. No language, so no explanation to argue with, only a score to check. That turned out to be the feature: I audit counts, not prose. Does Logical Intelligence expose a confidence per cell, or only the solved grid?
A plain checker catches most of this today. Whatever model produces the schedule or ledger entry, run a program that tests the invariants, like money in equals money out, and reject any output that fails before it reaches a downstream system.
The left-to-right + discrete-token point is the sharpest part of this. Once you commit to digit 4 in row 2, discovering the column collision in row 7 means regenerating the whole prefix — there's no continuous downhill step toward a valid grid. That's why an energy landscape over continuous cell scores fits CSPs better than next-token prediction, even when the LLM is 1000× larger. Curious how Kona handles soft constraints (cost, latency) vs hard ones (Sudoku rules) — does the energy function stay separable, or do soft terms flatten the valleys?
Language is not the same thing as intelligence. Children play Sudoku where they use spatial reasoning, logic, and flexible planning to solve a puzzle. An LLM only manipulates recorded text.
For a constraint-heavy game like Sudoku, you need a system that navigates logic, which is why humans (and smaller, logic-based models like Kona) easily beat the biggest text-based AI models ever built.
Okay, this is the first time energy-based models actually clicked for me. I wrote a while back about moving from deterministic to probabilistic problem-solving, and honestly Bodnia's "can you define your rules?" says it better than I did.
The new approach is interesting but seems more complex than a standard BFS or DFS approach which may only cause 1/100 of the new approaches. Certainly, love the concept but the example is controversial. We may need a better and real examples to better illustrate the new approach.
Prior to LLM, there are tons of research like the new ideas in the article - most of these researches were paired with realistic problems in a more challenging spaces.
Amzaing
A model that returns no sentence is what I run every day, so the Sudoku result reads like home. Mine sorts public posts into rooms and returns a label with a probability, and the day I checked it against a language model on the same hundred posts, the funniest miss was a post saying "Thank you for giving me a good laugh" filed under AI. No language, so no explanation to argue with, only a score to check. That turned out to be the feature: I audit counts, not prose. Does Logical Intelligence expose a confidence per cell, or only the solved grid?
A plain checker catches most of this today. Whatever model produces the schedule or ledger entry, run a program that tests the invariants, like money in equals money out, and reject any output that fails before it reaches a downstream system.
Businesses don’t lose money because they lack AI. They lose it because the wrong constraints go unsolved.
A bottleneck in a workflow becomes lost capacity, slower delivery, and limited growth.
The real intelligence is knowing which constraint to remove—and how much business value removing it creates.
The left-to-right + discrete-token point is the sharpest part of this. Once you commit to digit 4 in row 2, discovering the column collision in row 7 means regenerating the whole prefix — there's no continuous downhill step toward a valid grid. That's why an energy landscape over continuous cell scores fits CSPs better than next-token prediction, even when the LLM is 1000× larger. Curious how Kona handles soft constraints (cost, latency) vs hard ones (Sudoku rules) — does the energy function stay separable, or do soft terms flatten the valleys?
If an LLM can see the problem and write a script to solve it (well within the abilities of frontier models now) then what’s the difference here?
Why LLMs are poor in game like Sudoku?
Language is not the same thing as intelligence. Children play Sudoku where they use spatial reasoning, logic, and flexible planning to solve a puzzle. An LLM only manipulates recorded text.
For a constraint-heavy game like Sudoku, you need a system that navigates logic, which is why humans (and smaller, logic-based models like Kona) easily beat the biggest text-based AI models ever built.
In Eve Bodnia words, “you’re navigating the understanding, rather than the data.” https://blog.bytebytego.com/p/how-to-build-models-without-language