"THIS FLY CAN BEAT YOU AT STAR CRAFT: HAS SCIENCE GONE TOO FAR????"

I haven’t checked for SC:BW but Blizzard has official parsers/replayers for SC2 on GitHub.

Not hard to build. I was shocked at how fast/easy this was to pull together.

Yes, e.g. https://github.com/gulshngill/bwrepanalysis. But I doubt an agent would learn much from them without a lot of additional processing - the replay files are little more than a stream of the orders given during the match (e.g. "at tick 17, player 1 ordered unit 234 to attack-move to 56,78"). They're difficult to make sense of without a lot of additional context, like the map layout, the location and status of other units, what parts of all that are actually visible to each player, etc.

To be fair I’m recalling 2v2 NR20 games from an era where I played talking to my childhood friend on a landline.

So there was enough time. The big risk was that the map would max before you could fully develop the Zerg side.

Defilers

Production speed scaling due to hatches.

If you want 6 carriers at the same time, build 6 starports.

If you want 6 ultralisks or 6 mutas at the same time, build 2 hatches (which you already have)

One consequence of this is that changing army composition as Z is very easy.

Strong and unpredictable early game, solid middle game, strongest late game. It also has some of the best units in the game: zerglings, mutalisks, defilers. Especially defilers, they are arguably the strongest spellcasters in the game

Wow this is super cool. I wonder when LLMs get cheap enough that NPCs will be able to have full conversations with players.

Not sure they’re the best option for raiding, but as a high-level orchestrator for choosing content, that sounds pretty great.

Just curious: do you know the expected XP gain over the first 90 minutes for a human player (either a medium-level one or a speedrunner)?

Try it, you can run your own games on bw.swerdlow.dev

Later we even listened to Eminem and played violent video games. Most of have never even been charged with a crime, much less abused someone.

I agree.

My first time playing StarCraft was at summer camp around a decade after it came out.

All the smartest people played it so I wanted to too. Great decision, I have been continually impressed with the people who StarCraft introduced me to.

So dope

So ... Zerging Man?

That would be cool. It's Captain Pump's Raiders

oh man! missed opportunity.

Likewise really miss those carefree days killing time in servers and making diabolically hard maps to play online with strangers and friends.

I did buy it, all three, loved it. But somehow, broodwar is the one coming back on the scene more than it

Same and Brood War is starting to have a bit of a resurgence. Not anything huge. When Battle.net servers are actually working, tend to play on the ladder a few times a week. Get crushed, but still one of the greatest competitive games ever made.

I mean, if anything, the "great" part is about the specific time and place. Which is incredibly interesting when it comes to this post, because the time and place i am talking about brought me here...

Man I wish I could check in on my old battle.net friends

Battle.net servers was my hangout spot with friends eventually joint clans and playing ranked matches. Making maps to play online was also super fun. Brood War was perfect.

That's Red Alert 2 and Age of Empires 2 for me, and I really long for those times I could gather with friends in an internet cafe and be able to steal their hamburger in the real world when they made me mad in game.

Playing these games autoregressively isn't even the right way to use the LLM for this task (unless it was trained to do so...). It's somewhat like having a creative writing bechmark but requiring that all the input/output be base64 encoded. It can do it-- but no guarantees on the results!

And it's also just bencmaxxing bait: you can get a huge improvement on the task by RLing on it, but make no improvement on anything else. Doing so would just waste model capacity.

If you could tell that every LLM was equally not being exposed to the task then you could justify it as a test of abstract reasoning, but you can't. So it ends up on how much go transcripts ended up in the training, which is ... not a very interesting metric.

This is a great idea. It could learn from its mistakes, repeat things that work, abstract complex situations. I would even want humans monitoring the project.

Don't let yourself get too comfy with 9x9! I stayed on it for too long. No way to know but I feel like it added 2 years to the time it took to get there

Hah, yeah that's definitely interesting, though maybe a general platform for this stuff would be even more interesting.

Although, maybe benchmarking an agent on "how well can you command a swarm to annihilate the Terrans" is how it all starts going downhill...

Yeah. Some of it may just be "thinking" less rather than faster token generation.

Do the humans get to use this auto-stagger too?

The best sc2 bots these days play in the ~50,000 APM range (they could mostly go higher as well, but the game client breaks somewhere around 100k APM). I don't see an LLM-based bot getting up to that sort of speed anytime soon.

On the other hand, I do think LLM-based bots will quickly outperform the decision-making of many of the hand-coded bots, so maybe they won't need so much APM to be competitive.

Should be possible to play this with Jev.

AlphaStar beat one retired professional by cheating.

AlphaStar won a showmatch against TLO. TLO was never one of the strongest players in the world. He had been retired for over three years by the time of the match. Google set the rule that their system would have human-like mechanics, but it played several times faster than any human, never issued a wasted action, had an inhumanly fast reaction time, issued commands with perfect accuracy using an API, and could see the entire map at once.

It was later released to the open ladder with more human-level mechanics. Even strong amateurs regularly beat it. I have beaten it myself. It was strong, but not even close to the level of the strongest human players. It had obvious and easily-exploitable deficiencies in strategy and building placement.

I think even the cheater version would have lost handily to Serral or any of the strongest players.

(It apparently beat MaNa as well as TLO, but those matches were never released to my knowledge. I see no reason to assume Google cheated less flagrantly in private than they did in public.)

There's a bot data stream already; it's probably hooked up to that rather than screencap. Yes, I believe these were playing in real time.

It plays from a special API. Some things that are impossible with normal UI are possible with API, like selecting a unit under other units. Invisible units are also reported via the API.

dota wouldn't exist without Aeon of Strife (the UMS map from Starcraft that was later reworked into WC3 dota)

It’s not “endless complaining” when the bot is functionally cheating. Yes if you let a computer use special mechanisms it will operate faster than a person who is using a mouse and a keyboard. A computer computers better than a person, what a revelation.

I don’t applaud people using aim assist in CS for the same reason. We literally don’t have access to the same tools. That’s why chess has been such a great target for years. One’s ability to manipulate the pieces has no bearing.

I’m a little bummed it did have map hack on. Would love to see this play “legit”. It really reminds me of my pastime, chess. When you play bots, if you don’t mess up early, it can often feel like you’re winning in the early game even though you’re just hanging yourself slowly with the rope it’s giving you.

bot is at the top of the kr ladder right now (excluding drop hackers) https://www.youtube.com/watch?v=1vsTqNwHquE

If you want to see human written bots in action or compete in the bot ladder yourself, try https://aiarena.net/

I've been running these with friends recently, it's very fun. I ran irl bot tournaments for board games a couple times but the agent era opens up a huge amount of possibilities.

Must recently I built out a MMORPG puzzle box thing, I wrote a general game architecture doc but left the specific puzzle design up to Fable. Nobody is actively playing rn but I left it up at bot.willmorrison.net.

In the early days of SC2, I remember people using genetric programming to optimize build orders. I remember a slightly unorthodox Zerg Roach Rush which was _really_ fast.

All the hobbyists were using proxybot because it allowed you to use more fun languages than C/C++ but proxybot lacked the features to effectively play Zerg. I really wanted to play Zerg so I built a really sweet API+DSL in Ruby around proxybot and then used that to give the other newbies a hard time with a zergling rush.

Unfortunately the proxybot limitations precluded me from expanding its capabilities so I tried to switch to having ruby embedded in C++ and I basically got mired there and was eventually distracted by real world concerns like actually finishing my degree.

I think I played against Krasi0's bot a couple times in the early days. Hopefully they stuck with AI and have suddenly become crazy rich after 2016. It certainly wasn't a given that AI was going to lead to a fruitful career back then, let alone to unimaginable riches.

I also entered the competition as an undergrad—I contacted Ben Weber 10 years later (2020) and interviewed him on my podcast:

https://theaccidentalengineer.com/adversarial-machine-learni...

At last something that feels properly orthogonal to pelicans on bicycles.

Gemini wasn’t included, but I’m guessing its performance would have been similar to Grok’s performance despite having roots in DeepMind.

+1 because StarCraft

Wondering what it would look like if you allowed them to write scripts. You could throttle the number of clicks to make it interesting.

It can play doom in realtime so maybe it can play this.

Committed to the bit; context aliased “change plans” to “pylons” and I have a command to /construct them

AlphaStar was absolutely dominant. I don’t know if it played exhibition matches after it got its view restricted, but surely it still performed at a very high level. The APM restriction they put on is actually a handicap in favor of the human since humans are allowed to have 2k APM, just incapable of doing so.

If you just mean general LLMs can’t beat humans yet that’s one thing, but it’s not the case that no system can do so.

https://www.youtube.com/@CholeraSC/videos

I would love all the NukeTheStars matches in 4k, his commentary was fantastic

Maybe one could try to reconstruct very closely matched replay files?

Kespa or Gomtv remastered/upscaled their old televised games about a year ago. Just using the upscalers that were the hype at the time.

working on a similar project for street fighter currently https://www.youtube.com/watch?v=dJYTV3ZiT6o

I would love to create one of these benchmarks for age of empires 2, but I have no idea how to make the AIs play it. Maybe I can get Claude to do it anyway.

Good example of how these things sometimes spend way too much time thinking to be useful

I don't normally anthropomorphize the AIs but this is super cute, lol.

Yeah, what's the deal with this error message?

AlphaGo, Watson, Stockfish, Eliza. I can name so many of the interesting stuff and each new LLM just comes and goes.

No model will ever figure out how to make a ling tight wall.

This thing kept me sane through university. I had a shitty computer which could barely run this and no Internet so I just played against the computer, which was both frustrating and educational.

I will always love this and now I'm going to play it again. Remastered and on a fancy modern machine.

*wonders if running startcraft in wasmjs uses less CPU than anubis

Eh… I'm fairly sure that most of the people here are old enough to have played Star Craft and Brood War when it first came out, and are fully aware of e-sports becoming huge in South Korea. And I have a hard time imagining younger generations that grew up with e-sports being an established thing instead of a novelty to not be aware of how South Korea looks at Brood War differently.

Interesting that it benches down to Haiku but doesn't bench any Chinese models (which are at least between Sonnet and Opus, when they aren't beyond Opus).

Maybe just speed? TFA said that many agents failed because they were thinking too much and doing too little.

You must construct additional pylons

Your APM is too low. If you click a bunch sometimes the pause button works.

Game developers cheat like there's no tomorrow, though. In the videogame 3D graphics space, the old mantra was, "if it looks right, it's right".

That barrel you shoot, is really half a barrel when you're up close, a flat rectangle when you're far, a point-with-mass + a vector for purposes of physics, and not even there for purposes of AI because pathfinding uses a precomputed graph of nodes that's carefully aligned with the map so you don't notice the enemies can noclip through everything other than floors and walls. Etc.

And yes, many games would have scripted enemies or other events come out at you so you don't linger in particular areas too long, lest you spot some of the shortcuts they made.

I grew up wanting to make games, spent my teenage years in hobbyist gamedev communities, and to date, this remains to me the most enjoyable and pure form of exercising software development skills.

It's very different to have a generalist AI be able to win rather one trained for starcraft

This is a very interesting benchmark, and I think it has a lot of potential to make the speed of a model quantifiable.

I'm often asking myself is it better to use higher or lower effort levels, or to maybe drop down to a "dumber" but faster model. And so using a real-time based competition as a benchmark could shed some light on this, I think.

In this vein, here are what I would love to see added in this benchmark:

- Include Google's Gemini models. I keep hearing Gemini being praised for its speed, and I would like to see whether that gives it a big enough edge over the bigger but slower models. - How does a Cerebras-accelerated open source model fare against a much larger but much slower frontier model?

I also feel like in general there is a lot of very low-hanging fruit to start benchmarking models across the spectrum of real-time vs batch-style workloads. Perhaps Brood War sits somewhere quite near the "real-time" end of the spectrum, but what about something like a game of speed chess, or a turn-based game with time limits?

I think what I would like to see the most is for someone to come up with a benchmark that supports tuning the "real-timeliness" of the benchmark, and then running a sweep of a model across the whole spectrum. That could get result in real nice graphs with multiple models on the pareto-frontier, varying based on the hosting provider and the model dimensions.

“Older models tended to play the RTS as a turn-based game, leading them to get destroyed while they were thinking. ”

That’s me. I’ve always struggled with real-time games because I need to pause and think. While I excel at chess and board games, I’m just no good at real-time ones. At last I can only manage by sticking to a fixed set of tactics for a game, which minimizes the need for on-the-fly thinking. Seeing current models face the same difficulty leaves me with mixed feelings.

Check out pluto, RL trained sc bw bot. Somehow it is now going rogue on the Korean ladder and flattening pros. There are some yt games by it, search for ^333^

It was the late 90s, my very first day at a new school, I was asked to introduce myself at the front of the class, I mentioned I like computers, one guy at the back of the class blurts out “En Taro Adun” and without skipping a beat I replied “J’tokoh zohl”, and we instantly became best friends.

So they aren’t intelligent.

Like if a model can’t handle a task it hasn’t been trained on extensively, that’s not intelligence it’s memorization.

Fable's effort level is a significant omission, given it was ranked 3rd behind Astra xhigh and Astra medium.

Hmmmmm, almost like these models aren't generally intelligent

"another match is already active"

Really looking forward to playing- are you all limiting boxes?

Astra had an explicit medium/xhigh levels, Fable - just Fable. What reasoning level was used? Why not multiple were tested?

I’ve scrolled the article, but haven’t noticed any remarks about Fable’s levels.