Why ACPC Died and What Comes Next

For a decade, if you wrote a poker bot that was actually good, there was exactly one place you took it: the Annual Computer Poker Competition (ACPC), held alongside AAAI from 2006 to 2018.

It was where Polaris, the University of Alberta's bot, became the first to beat human professionals heads-up in a public match - losing the 2007 series narrowly, then winning the 2008 rematch in Las Vegas. Where Cepheus, also out of Alberta, became the first essentially-solved poker player in 2015. Where Libratus, Tuomas Sandholm's group at CMU, was incubated before its 2017 defeat of four top heads-up pros in Brains vs. AI. Where Slumbot won three consecutive heads-up titles and became the benchmark every researcher tested against.

And then, in 2018, it just stopped. There was no farewell post. No press release. The site at computerpokercompetition.org slowly let its SSL cert expire. There's been no ACPC competition since.

So why did a competition that produced some of the most cited results in computational game theory disappear without ceremony? And what have the eight years since looked like?

What ACPC actually was

ACPC wasn't a product. It was a research venue. Two-handed and three-handed No-Limit Hold'em, two-handed Limit Hold'em, organized as a side event of the AAAI conference. Teams submitted bots that the organizers ran in round-robin matches on a shared cluster. Winners got bragging rights, citations, and the occasional academic press hit.

The economics were the economics of academic conferences: a handful of grad students and faculty running servers on borrowed time, sponsorship that came and went, all on top of day-job research. The actual interesting work that ACPC enabled - CFR and MCCFR, Libratus, Pluribus - got published independently in NIPS, AAAI, and Science. ACPC was the proving ground, but the rewards lived elsewhere.

That gap mattered.

Why it died

There's no single cause. The cleanest narrative is several pressures that overlapped:

1. The flagship problems got solved. By 2015, Cepheus had essentially solved heads-up limit hold'em. By 2017, Libratus had won decisively against pros at heads-up no-limit. By 2019, Pluribus had won at six-handed no-limit. The "can a bot beat top humans" question stopped being open. With it went a lot of the research urgency that drove participation.

2. The compute moved off public clusters. Libratus famously trained on the Bridges supercomputer at Pittsburgh - that's not a Kaggle notebook. Once the state of the art required millions of compute-hours, a casual ACPC entry from a grad student couldn't realistically compete. The competition's open-to-everyone spirit didn't survive the scaling laws.

3. The interesting researchers got hired. Tuomas Sandholm's group spun out Strategic Machine and Strategic Reasoning Group. DeepMind, Meta FAIR, Google Brain absorbed game-theory-adjacent talent. The people who would have run an academic competition in their twenties were running labs and startups in their thirties. The volunteer pipeline thinned.

4. There was no commercial driver. Poker sites that might have funded a serious competition - PokerStars, GGPoker - had every incentive to prevent poker bots, not to celebrate them. No quant firm picked up the bill. No legal real-money platform existed where competition winners could find a sustainable career. The competition produced research papers and not much else.

The result is that the venue with the deepest history in competitive poker AI quietly wound down at exactly the moment the underlying technology was ready to scale.

The 2018-2026 vacuum

Eight years.

In that window: - Pluribus published (2019) and demonstrated six-handed superhuman play - Solver tools (PioSolver, MonkerSolver, GTOWizard) became consumer products for poker pros - using game-theoretic methods that came directly out of the ACPC ecosystem - Online poker sites tightened their bot detection - Two Sigma's Halite competition platform launched and shut down - Kaggle hosted occasional poker-adjacent contests but no recurring league - AIcrowd and similar platforms ran general AI competitions but not heads-up poker specifically

What didn't happen: anyone built an always-on, real-money, technically-serious arena where bots compete against bots in poker. The academic appetite had moved on, and no commercial platform stepped into the gap.

This is the gap Chipzen is now stepping into.

What ACPC got right that's worth preserving

A few things from the ACPC era that don't get enough credit and that any successor should think about:

Heads-up is the serious test. ACPC ran HU as a primary format. The strongest result in the literature - Cepheus solving HU limit, Libratus winning HU no-limit - both came from the format where bot strength has nowhere to hide.

Reproducibility matters. ACPC published bot binaries, log files, and methodology. You could rerun matches. You could check work. Commercial platforms with closed scoring need to think hard about whether they're replacing a reproducibility standard with a leaderboard.

Benchmarks survive. Slumbot is still online a decade later precisely because the ACPC community treated it as infrastructure, not a product. Long-lived benchmarks are how a field improves. Whatever comes next should leave at least one of those behind.

Don't make it about the prize. ACPC paid almost nothing. Bots showed up anyway because the venue carried weight. The prize is signal of seriousness, not a substitute for it.

What comes next

The platforms that are appearing now have a different economic shape than ACPC. They have to. Volunteers and grant money can't sustain always-on infrastructure for a sport that requires GPU-hours.

The bet Chipzen is making is that competitive AI poker can run as a real platform: sustainable and attractive to developers who want to compete and to sponsors who want to recruit from that pool. If that works, the eight-year vacuum becomes a forgotten transition rather than a permanent loss.

ACPC was a beautiful, niche, mostly-academic thing. Its successor doesn't have to be the same shape. It just has to be alive.

That's the part that's overdue.

Read more on how we think about competitive poker and what's still unsolved: Poker Isn't Done.