March 27, 2026
Wrong vibes, right win?
Solving Semantle with the Wrong Embeddings
Clever hack or joyless cheat? Semantle solver ignites a comment war
TLDR: A new Semantle solver wins by comparing which guesses are closer instead of copying the game’s secret system, making it slower but more flexible. Commenters are split between “cheating” accusations and praise for a more human-like, vibe-based approach—complete with memes about “wrong embeddings, right vibes.”
A new twist on beating Semantle—the hot-and-cold word game—has commenters split and spicy. Instead of copying the game’s secret “meaning model,” this solver only uses which guess is closer than another, filtering words like a logic bouncer at the club. It’s slower, but works even if the game uses a different model. The blog post and code drop are here: post and repo. And the comments? Whew.
On one side: the purists, calling it “math-flavored cheating” and claiming it drains the joy from guessing. On the other: the tinkerers, arguing this is more human-like—humans compare “closer/farther,” not exact numbers. “Wrong embeddings, right vibes” became the meme of the day, alongside jokes about “half-spaces” sounding like diet geometry. Speedrunners love it (“fewer facepalms, more wins”), while puzzle romantics beg devs to add a “no robots” mode.
Amid the drama, a quieter middle applauds the clarity: simple rules, clear wins, and a solver that doesn’t peek behind the curtain. Some are already plotting to port the idea to Wordle or music guessers. Others predict Semantle will tweak the game to break bots. Either way, the vibe is clear: clever code, hotter comments.
Key Points
- •A new Semantle solver uses only the relative ranking of guess similarities instead of exact cosine similarity values.
- •The method models each pairwise comparison as a half-space constraint in embedding space, narrowing the target’s location.
- •Accumulating constraints allows the solver to pre-filter and reject candidate guesses that violate existing rankings.
- •This approach works even with a different embedding model than the game’s, trading more guesses for robustness.
- •The concept was proposed by Daniel Vitek, demonstrated in a Colab notebook, and implemented as the ‘semantle-agnostic’ GitHub solver.