A look inside the AI review: heatmap, key squares and attention links

The AI tab on the analysis page (/analysis) now has three new visual layers. Each one draws directly on the board what Kingside's neural network "sees", and they work independently — each has its own toggle, so turn on one or all at once.

The network runs locally, in your browser: nothing is sent to a server, and the review stays free and ad-free. It's a tool for exploring on your own — it shows where the network looks, it doesn't interpret the position for you.

1. Heatmap — where the network's representation is loaded

The heatmap colors the background of the squares by how "loaded" the network's internal representation is on each one. Brighter means more work is happening there for this position.

Two controls add depth:

• layer (1…10) — the network processes a position in ten passes; the slider picks which layer to view (the last, tenth, by default);
• feature (0…256) — "0" shows the overall load per square (overview mode, a dark-to-light scale); numbers 1…256 show individual internal features (an explorer's mode: the features are unnamed, formed during training and can be signed, so the scale is two-sided — color encodes the sign).

A separate switch picks the side — white, black or both — so the picture stays in one frame of reference as you step through moves.

2. Key squares of a mistake — where your move and the best one diverge

When a move is marked a mistake (?) or a blunder (??) and the "key squares" toggle is on, the board outlines in green up to six squares and numbers them by significance (1 is the most significant).

This isn't a rating of a single position, it's a comparison of two: the squares where the network's understanding diverges most between what you played and what it recommended. The outline shows where to look, not a verdict on why the move is bad. On accurate moves, or with the toggle off, there is none.

3. Attention links — which squares the network ties together

Lines between squares show the network's attention: which squares it considers together in this position (computed on the position before the move, with the game's move history taken into account). Line thickness is the strength of the link. Each key square gets its strongest link across the whole board drawn, and the board shows at most six lines so it stays readable. The channel has its own toggle.

This is a structural link — which squares "work together" — not the cause of the move. It reads best alongside the key squares: you see not just where the divergence is, but what the network relates those squares to.

Why it matters for a player

Together the three channels turn a review from a "bad move" verdict into a clear map: where the network is focused, which squares your move and the best one diverge on, and how those squares connect. We don't interpret the position for you — we hand you a tool to explore it yourself.

Open any of your games in analysis, go to the AI tab and switch the layers on.