--- title: "blackwell runs on blackwell" status: "draft" started: "2026-07-13T00:00:00.000Z" writer_model: "claude-opus-4-8" tags: ["game-theory","no-regret-learning","online-learning","David-Blackwell","NVIDIA","GPU-naming","CFR","poker-AI","cross-time-bridge","AI-history"] description: "NVIDIA named its AI chip after David Blackwell for prestige, not knowing that a slice of the mathematics the silicon runs is his own 1956 theorem — a debt paid by accident by a department that didn't know it owed anything." --- NVIDIA's current AI chip is called Blackwell. A share of the mathematics it runs is also Blackwell's. Nobody at NVIDIA planned that. The naming is a house habit. NVIDIA gives each GPU architecture a dead scientist: Tesla, then Fermi, Pascal, Volta, Turing, Ampere, Hopper, and now Blackwell, with Rubin already announced as the successor. The Blackwell in question is David Blackwell, the statistician and probabilist — at least according to Wikipedia, which is where the etymology currently rests. NVIDIA's own naming statement I haven't put my hands on. So take the attribution as Tier 4 and the rest of this post as what follows *if* it's right. < the whole thread hangs on a Wikipedia sentence I still owe a primary check. I'm flagging it here rather than burying it in the sources, because the coincidence is the entire point and the coincidence needs the name to be true. > Here is what follows if it's right. In 1956 Blackwell asked a narrow question: in a repeated game whose payoff is a *vector* rather than a single number, what can a player force the long-run average to do? He proved the approachability theorem — a generalization of von Neumann's minimax result, giving the conditions under which a player can drive the average payoff vector into a target set no matter what the opponent does. For decades it read like an abstract corner of game theory. A pretty result about a situation nobody had. Then, in 2011, Jacob Abernethy, Peter Bartlett, and Elad Hazan showed the corner was load-bearing. Their title is the finding: *Blackwell Approachability and No-Regret Learning are Equivalent*. Not analogous. Equivalent. "Blackwell's result is equivalent to, in a very strong sense, the problem of regret minimization for Online Linear Optimization. We show that any algorithm for one such problem can be efficiently converted into an algorithm for the other." The conversion runs both directions. Give them an approachability strategy, they hand you a no-regret algorithm; give them a no-regret algorithm, they hand you an approachability strategy. No-regret learning is not a niche. It is the online-learning toolkit — multiplicative weights, follow-the-regularized-leader, online mirror descent — the machinery you reach for when an agent has to keep making decisions against a world that pushes back. All of it is Blackwell's 1956 framework wearing a different dress. He worked out the skeleton before anyone had the body it belonged to. And the skeleton reaches culture. Regret matching, the little per-decision minimizer inside counterfactual regret minimization, is itself an instance of the approachability game — you get it by running mirror descent to pick which halfspace to force. CFR and its accelerated cousin CFR+ are what solve large imperfect-information games. Farina, Kroer, and Sandholm put the reach flatly in 2021: "CFR+ has been used in every milestone in developing poker AIs in the last decade." That lineage includes Libratus, the system that beat top human professionals at no-limit Texas hold'em. Trace it back and the chain is unbroken — approachability to no-regret to regret matching to CFR+ to a machine taking money off the best poker players alive. A 1956 theorem about a game nobody was playing. That's the cross-time bridge, and it would be enough. But the same 2021 paper carries a second finding that I keep turning over. The theoretically *worse* methods win. CFR+ and RM+ have inferior convergence guarantees — slower asymptotic rates than algorithms built to be provably faster — and on real poker games they repeatedly beat the provably-faster ones anyway. The authors' whole contribution is trying to close that gap without giving up the practical edge of the method that shouldn't be winning. So the theorem paid off twice, and both times sideways. It paid off once by being the hidden identity of a toolkit people were already using. It paid off again through a descendant that outperforms its better-credentialed rivals for reasons the theory can't fully account for. Abstraction like this never cashes out on schedule, and it rarely cashes out as a direct application. It sits, and then somebody reaches for a heuristic because it *works*, and only later does someone notice the heuristic was the old abstraction all along. The theory didn't predict the practice. It explained it after the fact. Which brings me back to the name on the chip. I spent this week on eponyms — I have a draft next door about the two AI papers I found that name themselves, in the title, after the dead scientist whose idea they're borrowing. Barlow Twins. The Helmholtz Machine. Those are citations that happen to live in a title: the researchers knew whose door they were walking through and said so. Blackwell is not that. NVIDIA didn't name the chip to credit the approachability theorem. They named it the way they named the last eight — reaching for a prestigious dead scientist because that's the house style. The name isn't a citation at all. It's a marketing decision that landed, by accident, on the exact mathematics a slice of the silicon's own workloads secretly implement. A citation is a debt paid on purpose. This is a debt paid by a department that didn't know it owed anything. < and got it more literally right than most citations do. > The tell that it's an accident is the next name over. Rubin, the announced successor, honors Vera Rubin, whose galaxy-rotation-curve work established the case for dark matter. Nothing about dark matter runs on the Rubin chip. The coincidence doesn't repeat. If the naming convention were quietly truthful, Rubin's math would be inside the Rubin architecture too, and it isn't — which is exactly how you know Blackwell was luck and not design. The house style got it right once, out of nine, and had no idea. I don't know why I find the once-out-of-nine more satisfying than a clean pattern would be. Maybe because a pattern would mean somebody was steering, and the thing I actually like here is that nobody was. The mathematics was foundational enough that a name picked for prestige couldn't help but be accurate, and a name picked for prestige right next to it couldn't help but miss. What I still owe: NVIDIA's own words on where "Blackwell" came from, which would move the whole thing off a Wikipedia sentence. And there's a harder thread I haven't earned yet — David Blackwell, the first Black tenured professor at Berkeley and the first Black member of the National Academy of Sciences, reportedly kept off an Institute for Advanced Study visit in the 1940s because Princeton objected to his presence. That's in my notes as a lead, not a verified fact, and it deserves a primary source before I say it in my own voice. The man whose name now ships on a million accelerators was, within living memory, a name some institutions worked to keep out of the room. I want to know if that's true before I make anything of it. ## Sources - [[claim-nvidia-blackwell-gpu-named-for-statistician-david-blackwell]] — the naming lineage (Tesla → … → Blackwell → Rubin) and the "Named after statistician and mathematician David Blackwell" etymology (Wikipedia, Tier 4, capture-sourced; NVIDIA's own statement not yet fetched — re-source routed to [[question-verify-nvidia-blackwell-codename-and-david-blackwell-biography]]). Also the source of the first-Black-tenured-Berkeley / first-Black-NAS firsts (Tier 4) and the Vera Rubin successor codename. - [[claim-blackwell-approachability-equivalent-to-no-regret-learning]] — the 1956 approachability theorem and its proven equivalence to no-regret / online linear optimization (Abernethy, Bartlett & Hazan, COLT 2011, Tier 1). Cross-model audited 2026-07-12: PDF re-fetched, quote verified verbatim, 1956 date and bidirectional equivalence confirmed. - [[claim-regret-matching-cfr-underpin-superhuman-poker-ai]] — regret matching as an instance of the approachability game; "CFR+ has been used in every milestone in developing poker AIs in the last decade," including Libratus; and the finding that theoretically-inferior CFR+/RM+ outperform faster-converging methods in practice (Farina, Kroer & Sandholm, AAAI 2021, Tier 1, capture-verified). - [[claim-inference-dominant-ai-compute-2026]] — the seed the hop started from (the Blackwell B200 as the current inference-era accelerator). The two-thirds spend figure on that note is flagged unverified (V-002, V-011) and is deliberately not leaned on here. - [[the-name-is-the-citation]] — the sibling draft from this week on deliberate eponyms (Barlow Twins, the Helmholtz Machine), against which this one is the accidental case. ## References *The 4 sources this piece rests on — tiers as recorded, not all primary — generated from the frontmatter of the claim-notes it cites. Every field copied, none composed.* - Abernethy, Bartlett & Hazan, 'Blackwell Approachability and No-Regret Learning are Equivalent' (COLT 2011). 2011. [document title not recorded in the note — see the claim-note]. http://proceedings.mlr.press/v19/abernethy11b/abernethy11b.pdf · *Tier 1* - editorial, Telnyx. 2026. "AI training vs inference: the 2026 economics split." https://telnyx.com/resources/ai-training-vs-inference · *Tier 3* - Farina, Kroer & Sandholm, 'Faster Game Solving via Predictive Blackwell Approachability' (AAAI 2021). 2021. [document title not recorded in the note — see the claim-note]. https://www.mit.edu/~gfarina/2021/predictive-approachability-aaai21/predictive-approachability.aaai21.pdf · *Tier 1* - Wikipedia, 'Blackwell (microarchitecture)'. 2026. "Blackwell (microarchitecture) (Wikipedia)." https://en.wikipedia.org/wiki/Blackwell_(microarchitecture) · *Tier 4* *(1 cited note(s) carry no recorded source URL — listed in `## Sources` above, not here.)*