For decades, speedrunners squeezed every last frame from their favorite games. Now AI-powered bots are beating those hard-won human records and, in many cases, discovering optimal ways to blaze through levels.
For decades, speedrunners squeezed every last frame from their favorite games. Now AI-powered bots are beating those hard-won human records and, in many cases, discovering optimal ways to blaze through levels.
The Evolution of Speedrunning
Speedrunning has always been about optimization and discovery. Top runners spend years hunting glitches and refining routes. Many also lean on external tools. For example, tool-assisted speedruns (TAS) use emulators, save states, and frame-by-frame inputs to build a theoretically perfect playthrough.
Because no human presses those buttons live, the finished run is often seen as a research or entertainment tool rather than direct competition with human records.
Today’s AI bots push things further. Instead of a person scripting every move, machine-learning agents train through trial and error, running thousands of attempts without fatigue. They learn strategies their creators never spelled out.
In other words, the process has shifted from people programming perfection to bots discovering it on their own. Below, I dig into some examples where AI agents have already posted times that edge past the best human runs.
1. QWOP
In 2021, researcher Wesley Liao set out to see whether an AI could beat QWOP, a notoriously difficult track-and-field game. At that time, top human players could barely finish the 100-meter dash in just under 49 seconds.
He approached the task by training neural network agents via reinforcement learning, essentially rewarding the AI for making progress and running faster while penalizing it for stumbling.
Through iterating on the training, it gradually learned a more upright, human-like stride. The breakthrough came when Liao modified the reward function to focus purely on speed. After about 40 hours of total training (including some pre-training on human data), the AI started pulling off the advanced technique, a kind of upward leg swing for extra momentum, that top runners use.
As a result, it managed to finish 100m in ~47.34 seconds, edging out the best human time of ~48.34 seconds. In other words, the bot set a new world record in QWOP, a confirmation that a learning algorithm could not only match, but exceed human performance in a speedrun scenario.
2. Super Mario Bros
The fastest SMB players complete the game (World 8-4 warp run) in just under 4 minutes 55 seconds, which is mind-blowing considering the game’s age and the frame-perfect precision required.
In late 2023, a coder set out to train an AI to speedrun SMB at a record pace. The project involved using a reinforcement learning algorithm called Proximal Policy Optimization (PPO) to teach Mario how to zoom through levels efficiently.
Training started in an OpenAI Gym environment that handled the game’s logic. The reward function pushed Mario to move right quickly, stomp enemies, and reach the flag with no wasted frames.
By drilling key stages (1-2 for the warp, then 8-2 and 8-4 for the finish), the bot closed the gap. Thus the final run matched the standing world-record time, which was a surreal moment for a game with decades of speedrunning history.
3. Minecraft
Speedrunning isn’t confined to retro 2D games. Modern open-world games also have speedrunning communities. In Minecraft’s case, the goal is often to defeat the Ender Dragon and finish the game. Humans have gotten pretty quick at this. With glitches and luck, a skilled player can finish in under 20 minutes on a good day, or around an hour in random seed glitchless categories.
However, Minecraft is a complex game for an AI to master. It’s open-ended, 3D, and heavily based on random generation and long-term planning. That didn’t stop intrepid developers from trying to build an AI that could beat Minecraft faster than a human.
In 2023, a YouTuber MCBYT documented an experiment where an AI agent (nicknamed Alto Clef) was pitted head-to-head against a human in a race to complete Minecraft.
Built on the Baritone path-finding framework, Alto Clef layers task logic on top: gather resources, craft gear, build a Nether portal, trade with piglins, reach The End, and bed-bomb the dragon. In a head-to-head race it clocked 2 hours 6 minutes, beating the human’ player’s 2 hours 7 minutes by seconds.
How I See AI Bots Outsmarting Human Speedruns
The jump from TAS scripts to self-learning agents hints at a future where “perfect” stops being a fixed target and starts drifting away from human reach. As such, I see three major shifts that will push the goalposts for good.
Most notably, the next step is in-run adaptation. For instance, a bot that monitors RNG patterns as it plays and rewrites its path on the fly, resetting fewer seeds and sliding straight to the fastest outcome every time.
Soon, evolutionary search will comb input space continuously, flagging physics quirks the moment they emerge after a patch or fan mod. A runner wakes up, and the bot has emailed a new wall-clip that warps straight to the credits. Hence, the discovery loop will collapse from months of testing to a single night of server time.
Finally, the foundation models trained on entire console libraries could learn generalized physics before ever loading a specific game. A brand-new indie release drops, and within hours an AI agent maps momentum quirks, computes the theoretical time, and posts the run.
I love speedrunning for the creativity and grit on display at GDQ streams, yet I can’t shake the sense that AI pushes the hobby toward a fork in the road. AI agents and AI companions are starting to seep into every industry imaginable. They’ve started interacting with every part of our business, and now the more private, leisure part of our life. And I’m not talking about the dirty talk with AI fun type, although that’s also becoming a thing more and more, but this casual, but still competing side of the gaming, where they’ve put themselves between us gamers who are actually enjoying the competitiveness among real fellow gamers. Either we tighten rule sets around what counts as a human run, or we embrace parallel leagues where code races code for pure optimization glory.
Both paths excite me.
Both chip away at the comfort that world records belong to mortal reflexes alone.
Both turn speedrunning into an ongoing dialogue between code and courage, where every new algorithm redraws the map and every daring player tries to prove there is still room for human flair atop machine-found routes.
