AlphaZero and Leela Chess Zero Explained

For decades, chess engines improved primarily through incremental refinement of hand-coded evaluation functions — human programmers explicitly telling the engine how much a knight outpost or a doubled pawn was worth. AlphaZero and its open-source successor Leela Chess Zero represented a genuinely different approach: engines that learn to play chess almost entirely through self-play, discovering their own understanding of the game rather than following rules written by human programmers.

How Self-Play Learning Actually Works

Rather than being given explicit evaluation rules, these engines start with only the basic rules of chess itself and improve by playing enormous numbers of games against themselves, using the outcomes to gradually refine a neural network that predicts good moves and evaluates positions. Over millions of self-played games, this process produces an engine with genuinely strong chess understanding, arrived at without ever being told which specific features of a position matter or how to weigh them — the engine essentially rediscovers chess strategy from first principles through pure trial and adjustment.

AlphaZero's Landmark Demonstration

Developed by DeepMind and publicly demonstrated in 2017, AlphaZero made headlines throughout the chess world by convincingly defeating Stockfish, at the time the strongest traditional engine, after training for only a matter of hours purely through self-play, with no access to human game data or hand-coded chess knowledge at all. This result was widely discussed not just for the outcome itself, but for the distinctly different, more intuitive-looking style AlphaZero displayed — frequently favoring long-term piece activity and attacking chances over the more materially cautious calculations traditional engines were known for.

Leela Chess Zero: An Open-Source Successor

Because AlphaZero itself was never released publicly, the volunteer-driven Leela Chess Zero project set out to replicate its self-play training approach in a fully open-source engine anyone could run, train, and study. Leela has since become a genuinely elite engine in its own right, competing at the very top of computer chess rating lists and giving the broader chess community direct, hands-on access to a neural-network-based engine with a similar training philosophy to AlphaZero, rather than relying solely on DeepMind's original, unreleased demonstration.

A Genuinely Different Playing Style

Neural network engines like AlphaZero and Leela are frequently described as playing with a noticeably more "human" style than traditional engines — favoring long-term piece activity, king safety, and space over the more materially precise, sometimes visually awkward-looking play that classical evaluation-function engines were known for. This isn't a coincidence: because these engines learned through self-play rather than explicit programmed rules, their resulting style emerged organically from what actually worked across millions of games, producing an intuitive feel that many strong human players found genuinely inspiring to study.

How This Changed Elite Opening Preparation

Following AlphaZero's public demonstration, several of its games were closely studied by elite players and their preparation teams for genuinely novel strategic ideas, particularly involving long-term piece sacrifices and space advantages that traditional engine analysis had previously undervalued. Some ideas first popularized through AlphaZero's games found their way into serious top-level opening preparation, representing a rare instance of an engine's own discovered style directly influencing human opening theory rather than simply verifying existing human ideas.

Modern Engines Now Blend Both Approaches

Following the success of neural-network-based engines, even Stockfish, long the standard-bearer for traditional evaluation functions, incorporated neural network evaluation into its own search process, blending the deep tactical calculation traditional engines were already known for with the more intuitive positional judgment neural networks bring. This convergence means the sharp distinction between "traditional" and "neural network" engines has become considerably less meaningful in practice today, since the strongest modern engines now draw on both approaches simultaneously rather than representing two genuinely separate, competing schools.

Why This Matters for Everyday Players

Even for players who will never train their own neural network engine, the broader shift these projects represent is worth understanding: modern engine analysis increasingly reflects genuinely learned, intuitive-feeling chess understanding rather than purely mechanical, rule-based evaluation, which is part of why studying strong engine games today often feels more directly instructive for human strategic understanding than it did with earlier generations of purely calculation-driven engines.

Running Leela Chess Zero Yourself

Unlike AlphaZero, which remains a closed research demonstration never released for public use, Leela Chess Zero can genuinely be downloaded and run locally, similarly to Stockfish as covered elsewhere in this pipeline, though it typically benefits more from graphics-card acceleration than traditional engines do, given its neural network architecture. Players curious to experience this style of engine directly, rather than only reading about its games secondhand, can pair it with the same kind of UCI-compatible graphical interface used for any other local engine setup.

The Broader Lesson for Human Players

Beyond the specific technical details, the AlphaZero and Leela story offers a genuinely useful broader lesson for human players at any level: strong chess understanding doesn't have to come from memorizing explicit rules handed down by someone else — it can also emerge from sustained, deliberate practice and honest evaluation of what actually works, a process not entirely unlike how a human player's own intuition develops over years of serious study and play, just compressed into a vastly larger number of repetitions than any human could complete in a lifetime.

Where to Find Well-Documented Games From These Engines

Several of AlphaZero's original demonstration games against Stockfish were published in full by DeepMind and remain widely available for study, alongside a large and continuously growing library of Leela Chess Zero's own games from ongoing testing and competitive engine tournaments. For a player wanting to see this distinctive playing style firsthand, working through a handful of these well-documented, publicly available games is far more instructive than any secondhand description of their strategic tendencies alone.

A Note of Caution About Blind Imitation

It's worth remembering that these engines' apparent style emerged from optimizing purely for winning probability across millions of games, not from any conscious aesthetic or pedagogical intent, and some of their specific decisions reflect a depth of calculation no human can genuinely replicate over the board. Studying their games for broad strategic inspiration is genuinely valuable; attempting to imitate specific sacrifices or piece placements without understanding the calculated justification behind them risks copying the surface style without the substance that actually makes it sound.

The Ongoing Rate of Engine Progress

Since AlphaZero's original demonstration, neural-network-based engines have continued to develop rapidly, with Leela Chess Zero and successive versions of Stockfish's own neural network integration each pushing playing strength further, in a pace of improvement that shows little sign of meaningfully slowing down. Following this ongoing progress, even casually, gives a genuinely useful sense of how quickly computer chess continues to advance well beyond the specific historical moment AlphaZero's initial announcement represented, and what that pace suggests about where the technology heads next. Players who found AlphaZero's original demonstration compelling have plenty of more recent material to explore, since the underlying research direction it opened has continued producing genuinely new results year after year, rather than settling quietly into a fixed, unchanging state.

Looking Back at a Genuine Turning Point

Years on, AlphaZero's 2017 demonstration still holds up as a genuine turning point in how the chess world thought about engine strength and engine style alike — not simply a stronger opponent for Stockfish to beat, but a meaningfully different way of arriving at chess understanding in the first place. That distinction between raw strength and genuine novelty of approach is worth keeping in mind whenever a new engine or training tool claims to represent the next major leap forward.