Thinking Back to TI 2018, OpenAI Five, and How AI Changed So Fast
Published:
The recent public arguments around OpenAI’s nonprofit roots, its later for-profit structure, and the long-running fallout between Elon Musk and Sam Altman brought back a very specific memory for me: watching the Dota 2 tournament livestream around TI 2018 and thinking, even then, that something unusual was happening. OpenAI started in 2015 as a nonprofit, later created a capped-profit arm in 2019, and the tension between mission, control, money, and scale has stayed attached to the company ever since. 12
What came back to me was not really the corporate side first — it was the feeling of seeing a benchmark suddenly become part of the wider public imagination. At The International 2018, OpenAI Five did not beat the top professional teams, but it did hold its own for long stretches, and OpenAI itself described those matches as close and revealing. A few days earlier, OpenAI Five had also won a benchmark best-of-three against very strong human players in front of a large livestream audience. That whole period felt like one of the first times many people outside research circles could see modern AI as something more than a neat lab demo. 345
Then, less than a year later, OpenAI Five did beat the reigning Dota 2 world champions, OG, in two back-to-back games. OpenAI called it the first time an AI had beaten world champions in an esports game, and also the first time an AI had beaten esports professionals on livestream after earlier high-profile systems had managed private wins but lost public matches. That matters to me not because a game result proves anything magical on its own, but because it showed how quickly progress could compound once a system had the right training setup, enough scale, and a benchmark hard enough to expose real weaknesses. 67
Looking back now, Dota 2 was such an interesting benchmark because it was messy in exactly the way a lot of real problems are messy. It involved partial information, long time horizons, fast decision-making, teamwork, adaptation, and the need to act under uncertainty. OpenAI’s later paper on the project explicitly framed Dota 2 as a task with long horizons, imperfect information, and complex state-action spaces — the kind of environment that makes simple comparisons to chess or Go feel a bit too clean. That is probably why the OpenAI Five story still sticks in my head more than many other AI “milestones” from that era. 76
I think that is also why the memory came back so strongly when the Musk/OpenAI story started surfacing again in the news cycle. In my mind, that whole OpenAI Five period now feels like one of the last times the public story around OpenAI was mainly about bold technical experiments and visible benchmarks. After that, the conversation moved more and more toward structure, governance, funding, productisation, and who exactly gets to steer systems that are becoming economically and politically important. That shift is visible in the company’s own timeline too: nonprofit origins, the 2019 capped-profit structure, then the later restructuring debates once AI moved from impressive research to mass commercial relevance. 21
I do not want to pretend a Dota result directly caused the future, or that one tournament suddenly changed Silicon Valley. That would be too neat, and history is rarely that tidy. But from the outside, it really did feel like that period changed the mood. Suddenly, AI progress looked less hypothetical. It felt faster, more embodied, and harder to wave away. When something can coordinate five agents in a chaotic multiplayer game and then beat world champions on a livestream, people understandably start asking bigger questions about what other domains might move next. 67
What I still find relevant today is not just the nostalgia of that moment, but the benchmark lesson underneath it. Benchmarks matter because they give us a way to notice when a system stops being merely interesting and starts becoming genuinely useful, or at least genuinely hard to dismiss. But benchmarks can also distort how we think if we treat a single win like a full explanation of intelligence. OpenAI itself wrote that the surprising thing in the Dota project was not some elegant new trick so much as the power of scale — training systems at previously unprecedented scale and seeing familiar reinforcement learning methods go much further than expected. That idea has echoed through almost everything that followed in AI. 67
That is probably the real reason this memory keeps coming back for me. TI 2018 and the OpenAI Five run now feel like a marker between two AI eras: one where public milestones still felt experimental and slightly niche, and another where the same organisations became central players in product markets, policy debates, and governance fights. Maybe that is why the recent OpenAI story feels oddly connected to those old tournament streams. Not because esports predicted the future in any direct way, but because it captured a moment when AI capability became visible enough that the questions around ownership, direction, and purpose were always going to get louder. 362
I still remember watching those matches and feeling that mix of curiosity and unease: curiosity because the technical achievement was clearly real, and unease because once a benchmark starts falling, the next question is never just “what else can this system do?” It is also “who decides where it goes next?” That question feels even more relevant now than it did back then. And if there is one reason I still think about OpenAI Five, it is probably that: not just because it was impressive, but because it made the future feel a little less abstract.
