It started, as these things do, with a genuinely interesting fact. On September 3, 2026, science-and-tech account Massimo (@Rainmaker1973) posted a simple breakdown of chess's combinatorial explosion: 20 opening moves, roughly 400 positions after one full move, about 8,900 after three, nearly 200,000 by move four, and by move 40 the count of possible games reaches something like 10⁴⁰ — comparable to the number of atoms in the observable universe.
Elon Musk replied: "The actual number of moves that are not utterly stupid in chess is tiny and chess will be fully solved one day." Chess.com's account pushed back with its own number — the full game tree at roughly 10¹²⁰ possible games, "about 40 orders of magnitude" more than atoms in the observable universe — and added, dryly, "To solve chess, you'd have to map them all. Good luck." What followed was several rounds of increasingly technical sparring, Musk invoking a different number (legal positions, not games), Chess.com clapping back with one-liners ("I'm a full-time employee", "even my friends don't text me back that fast"), and Musk eventually turning to Grok to run the actual math on a joke he made about storage limits.
It's a fun thread. It's also a real, checkable computer science question underneath the jokes — and it connects directly to why modern game-playing AI works the way it does. This isn't independently verified beyond the public tweets themselves, so treat the exchange as reported from X, but the underlying combinatorics are well-established computer science that we can check independently.
TL;DR
| Claim | Number cited | What it actually measures | Correct? |
|---|---|---|---|
| Chess.com: "more possible games than atoms in the universe" | ~10¹²⁰ games vs. ~10⁸⁰ atoms | Full game-tree size (the Shannon number) | Yes |
| Musk: "all legal chess positions are ~40 oom less than atoms" | ~10⁴⁴ positions | Distinct legal board positions, not games | Yes — but a different metric than Chess.com's claim |
| Musk: "chess will be fully solved like checkers" | — | Whether a compact proof of the game's outcome exists | Unproven either way |
| Checkers precedent | ~5×10²⁰ positions, ~10¹⁴ examined | Weakly solved in 2007 by Chinook | Solved; chess is not analogous in scale |
| "Storing a value for every legal chess position" | ~10⁴⁴ bits | Physically possible in principle, engineering-impossible today | Not ruled out by physics |
| "Storing the full game tree" | ~10¹²⁰ leaves vs. ~10¹²² bit holographic bound | At or beyond physical storage limits of the observable universe | Ruled out |
The two numbers that aren't actually in conflict
The entire disagreement hinges on a distinction chess players know well and casual observers often miss: a game is a full sequence of moves from start to finish; a position is a single board state, reachable by any number of different move sequences. These are wildly different quantities.
- Games (Chess.com's number): the Shannon number, commonly cited as roughly 10¹²⁰ possible games, based on an average branching factor of about 35 legal moves per turn across a typical ~40-move game. This is the number Massimo's original post and Chess.com's reply were both describing.
- Positions (Musk's number): the count of distinct, legal board arrangements — commonly estimated around 4.8 × 10⁴⁴. This is a completely different, much smaller number, because vast numbers of different game sequences funnel through the same intermediate positions.
Several replies in the thread — including from working chess players like Golf Takes (@theunplayable) and Daniel Kennedy (@danieldklogics) — flagged exactly this: Musk's rebuttal cited positions while responding to a claim about games, which is comparing two different measurements as if they contradict each other when they don't. Chess.com's original 10¹²⁰ figure and Musk's ~10⁴⁴ figure are both defensible numbers; they just aren't answering the same question.
What "solved like checkers" actually required
Musk's underlying claim — that chess will eventually be "solved" the way checkers was — is worth taking seriously on its own merits, separate from the games-versus-positions mix-up. Checkers was weakly solved in 2007 by the Chinook project, led by Jonathan Schaeffer at the University of Alberta, proving that checkers is a draw with perfect play from both sides.
The number that matters here isn't checkers' full position count (~5×10²⁰) — it's how few of those positions the proof actually had to examine: roughly 10¹⁴. Chinook didn't build an exhaustive lookup table. It used alpha-beta search, endgame tablebases, and aggressive pruning to construct a much smaller proof tree that establishes the outcome without visiting anywhere near the full state space. That gap — 5×10²⁰ total positions but only 10¹⁴ examined — is the entire reason "solved" doesn't mean what a lot of people assume it means, and it's the crux of whether Musk's prediction has any teeth.
What Grok's follow-up math actually settled
When Musk joked about a different storage question — riffing that "there are also more ways to piss in a pot than atoms in the universe" — he asked Grok to check it, and the answer is a useful, precise breakdown of a question people conflate constantly: "can something be stored" is not one question, it's at least three.
- Storing every full game explicitly. At ~10¹²⁰ leaves against a holographic bound of roughly 10¹²² bits for the observable universe, this is at or beyond what physics allows — genuinely, not just practically, impossible.
- Storing one value per legal position. At ~4.8×10⁴⁴ positions, one bit each is "only" ~10⁴⁴ bits — far beyond any storage humans can build, but nowhere near the ~10⁸⁰-atom physical ceiling. Not ruled out by physics; ruled out by engineering, for now.
- Storing a compact algorithm that computes the answer — the Chinook approach. This is the one that's actually open. Nobody has shown chess lacks a Chinook-style proof tree; generalized n×n chess being EXPTIME-complete is a hint that no small closed-form solution exists for the general case, but that's a statement about the hardest instances as the board scales up, not a proof about the fixed, finite 8×8 game most people mean by "chess."
Grok's own summary line is worth quoting directly: "A short program, circuit, or decision diagram that outputs the right move (or [win/draw/loss]) from any position would 'store' the solution in Kolmogorov-complexity form. We do not know that complexity." That's the honest state of the science — not solved, not proven unsolvable, genuinely open.
Why this matters if you build game-playing AI
This debate isn't just internet trivia — it's the exact problem that shaped how self-play reinforcement learning works today. Nobody built AlphaZero-style systems, or Stockfish's search-plus-evaluation engine, by trying to enumerate the game tree. They exist because brute-force traversal at chess's scale was never viable, going back decades before this thread.
The actual engineering insight is the one buried in Chinook's 10¹⁴-out-of-5×10²⁰ gap: a system doesn't need to visit every state to reach a strong or even perfect policy over the whole space, it needs a way to generalize from a tractable subset of experience to states it's never directly seen. That's precisely the bet behind self-play and function approximation in modern reinforcement learning — Richard Sutton's long-standing argument for general-purpose learning algorithms over hand-built search is built on the same core intuition. A model doesn't need a lookup table the size of 10⁴⁴ to play chess well; it needs an evaluation function and a search procedure good enough that the astronomical remainder of the state space never has to be visited to make a strong move.
Musk's ASI aside — that "advanced AI could compress [the solution] beyond what we could possibly comprehend" — is speculative but not incoherent: it's a claim about discovering a much smaller Kolmogorov-complexity representation than anyone has found today, which is exactly category 3 above, and exactly the kind of open question debates over whether current systems have reached superintelligence tend to gesture at without resolving.
What people are asking
So is chess "basically already solved" the way people casually mean it? No. Modern engines like Stockfish play at superhuman strength, but "very strong" and "solved" (a mathematically proven outcome with perfect play) are different claims entirely. Checkers is solved; chess, as of this writing, is not, and there's no known timeline for it to become so.
Does a bigger game tree mean a game is definitionally harder for AI to master? Not directly — Go has a vastly larger state space than chess (roughly 10¹⁷⁰ legal positions) and was still mastered to superhuman level by AlphaGo and AlphaZero years before this exchange, because search-plus-learned-evaluation scales very differently than exhaustive enumeration. Raw state-space size correlates with difficulty but isn't the whole story; structure and prunability matter enormously, which is exactly what makes the verifiable-reward self-play approach tractable at all.
Why did this become a huge thread instead of a five-tweet exchange? Partly the personalities involved (Musk's account has 200M+ followers; Chess.com's account has built a following specifically for fast, funny replies to public figures), and partly because the underlying question — games versus positions versus proof trees — actually does confuse people, including some very online, technically literate ones, which kept the replies going.
Did Grok get anything wrong in its answers? Grok explicitly corrected itself mid-thread — "No. I overstated that" — when Musk pushed back on an initial claim, walking through the three-category breakdown above instead of restating the original overstatement. That kind of visible self-correction, happening live in a viral public thread, is itself a small data point about how people are increasingly treating AI chat assistants as real-time fact-checkers in public arguments, not just private tools.
Related reading on explainx.ai
- RLSVR / SPyRL: Self-Play with Verifiable Rewards — how self-play RL sidesteps exhaustive game-tree search
- Richard Sutton's Oak Lab and the Case for General AI Algorithms — the reinforcement-learning argument this whole debate echoes
- Has AI Reached Superintelligence? The Astra Debate — more on the ASI claims Musk gestures at in this thread
- Paul Graham: LLMs, Math vs. Writing, and Verifiable Answers — on why some domains have checkable ground truth and others don't
- AI Benchmarks: The Complete Guide — background on how "solved" and "state of the art" get conflated across AI domains
- Grok 4.5 Private Beta: SpaceX, Tesla, Cursor v9 — background on the Grok model that fact-checked this thread live
Sources: Public X posts from @Rainmaker1973, @elonmusk, and @chesscom, September 3–4, 2026. Checkers-solving reference: Schaeffer et al., "Checkers Is Solved," Science, 2007.
This post reflects public X posts as of September 6, 2026. The underlying combinatorics (Shannon number, legal position counts, the 2007 Chinook proof) are established, citable computer science; the exchange itself may continue to evolve on X.