Profile picture
, 163 tweets, 71 min read Read on Twitter
Three hours from this tweet, @OpenAI's DOTA 2 bots will be taking part in what OAI are calling their "final" live game. I'll be tweeting about the game in this thread, aimed at non-DOTA players.

📺 Livestream: twitch.tv/openai
🤫 Hashtag to mute: #OAI
@OpenAI Last year the bots played live on stage at The International, and lost to two human teams. Here's a thread I wrote during their first game, which was played against the professional team paiN Gaming:
@OpenAI And here's a thread I wrote about their second game, which they played against a team of ex-professional "old boys". The overall skill/experience level here was arguably higher than paiN's, and the strategy employed by the humans is slightly different:
@OpenAI If anyone has any questions in the run-up to the games, feel free to ping me and I'll try and answer them in the thread! #OAI
@OpenAI 🔹 They're playing "reigning world champions" OG. OG did win The International, but their roster has changed since. I'm not sure which roster will turn up today.
🔹 They're playing on 7.21, the latest patch.
🔹 I have a hunch there might be a human/AI co-op game in store.
#OAI
@OpenAI The blog post notes that "games will be played with rules similar to those used for the OpenAI Five matches at [TI8]" but I can't tell if that's just the drafting format, or if it means the same restrictions for the AI too. #OAI

@OpenAI I'm intrigued by the language at the top of the blog post though - "AI’s impact on the world will be driven by its competence, scalability, and ability to enhance what humans can do — and this event will... demonstrate each of these" #OAI openai.com/blog/openai-fi…
@OpenAI A few disclaimers from me: I actually haven't played DOTA 2 in three months or thereabouts. This means I don't have much time playing on 7.21, but I know some of the things it aimed to change. A lot of them are really subtle and won't be a huge influence here. #OAI
@OpenAI I think there are two important things about it being on patch 7.21. The first is familiarity: some of the previous showmatches were on old patches, and it's possible some of the humans were a little less comfortable. Today, players will be familiar with the meta. #OAI
@OpenAI The second thing: in the pro matches I've watched recently, I've seen a slightly higher incidence of "hard carries", heroes which gather resources for a long time and become really strong. If 7.21 enables harder carries more than 7.20 did, this might disadvantage #OAI a bit.
@OpenAI Also potentially important: a change to a mechanic called denying. The bots are very good at doing this, but looking back at the TI8 matches, the humans generally hold their own due to the bot aggression. Will be interesting to see if this affects. #OAI
@OpenAI That's a very good question. I'm assuming this means it'll be the last time we see OpenAI play DOTA 2. There are a lot of potential reasons for this, ranging from pessimistic to cynical to optimistic! I'm a little sad, it's been fun.
@OpenAI Ah I was wrong about this: OG are back to their original roster, so we really will see the world champions today. Very exciting. #OAI
@OpenAI And one more thing from me: I'll probably be switching to #OpenAIFive for the hashtag, if you're muting. Also, check out the cool arena they've got set up for today:
@OpenAI Ah, apparently OpenAI will still "continue to work" on DOTA 2, according to this. That said, the emphasis on "final live event" in the press suggests we won't be seeing as many showmatches. #OpenAIFive
@OpenAI Live in just a few minutes! I'll be tweeting along. #OpenAIFive
@OpenAI We'll be seeing a few different things today I think, but in terms of the big matchup, the poll is slightly in favour of OpenAI's bots:

#openaifive
@OpenAI Looks like we have a short delay, so enough time to put the kettle on and grab a biscuit. As usual, please ask anything you like during the event, and I'll try and fit answers into the thread. #openaifive
@OpenAI Really in love with the colours on the countdown screen though. #openaifive
@OpenAI Alright, here we go folks! You can watch along with me live at twitch.tv/openai, or just hang out in this live thread. I'll try and keep it simple and novice-friendly. It's gonna be fun! #openaifive
@OpenAI Greg Brockman talking to the audience first: "It's about letting people connect with the strange, exotic but still very tangible intelligences that are created by... modern AI technology."

This feels like a new angle for OpenAI, which you can also see in their blog. #openaifive
@OpenAI From their blog: "We hope [today's event] will help people better internalize AI progress and how it will affect the world."

I have to say, I like this new tone. I think this is what big AI orgs with a lot of reach and money should be trying to angle towards more. #openaifive
@OpenAI Greg is framing OpenAI as an alien intelligence, but also that OG (the human team) is an alien intelligence from the AI's perspective. Again, that's a really nice way to frame the system. Makes this feel like the science communication angle was a bigger deal here. #openaifive
@OpenAI Greg is talking really excitedly. It's making me think that the system they'll be showing today is a lot more advanced than what we saw at TI8 last year. There's also talk of more heroes and fewer restrictions possibly. #openaifive
@OpenAI There's something really wild about the fact that three big esports personalities are on Twitch right now interviewing some people about reinforcement learning. #openaifive
@OpenAI On the limited hero pool: "We could learn a [new] hero, it's not that we think it can't be done, it's that we're focusing on the general learning process." I'm not entirely following the point being made here, but I think basically, they felt there were diminishing returns?
@OpenAI Blitz and Purge, two very good/technically ex-pro players, are talking about the impact of OpenAI's bots on pro player style. These examples are all from 2017's 1v1 matchup though - OpenAI Five didn't have the same impact on human play. #openaifive
Alright, we're at the meat now. The first phase of today is called "Competitor" - a best of three against OG. Very simple - can we play better than humans? Also this is a best of three, and this is phase one, so... strap in I guess. It's not gonna be a short one! #openaifive
Blitz is currently giving some context about OG, the human champion team. OG's win at The International last year was one of the most dramatic victories in DOTA history, and the team is still very strong today. This is easily the best team OpenAI have played live. #openaifive
Haha, oh wow, they went all out with the player intros and everything. That's really cute. #openaifive
Ok they're calling this Captain's Draft which implies OpenAI *is* drafting. IIRC this was not allowed at TI8 because they felt this would give the bots an advantage, since the smaller hero pool is unfamiliar to the human players. Should be interesting. #openaifive
Alright so we're drafting. I could write 500 tweets just on drafting, but let's take this slow. The humans have picked Earthshaker first - OG are great with this, it's a strong hero, humans like playing around it. But: they've given away Sniper and Gyrocopter. #openaifive
Now the thing with this is, at TI8 the teams had one of Sniper and Gyro each. The AI is really good at using both of these heroes. They have powerful spells and can do damage at range, and (for humans at least) only get better as time goes on. #openaifive
That said, in the past games we've seen, OpenAI hasn't planned for the lategame and isn't very good at it. So having these strong lategame heroes might just not be relevant. This is the biggest answer I want today - can OpenAI play long games and win? #openaifive
The second pick for the humans is Witch Doctor, which I think is potentially a terrible choice - its abilities are easily avoided with good reactions. And they've followed it up with very easily-killed heroes, like Riki (purple horns) and Shadow Fiend (helldemon). #openaifive
I don't know if OG have played #openaifive before but this feels like quite a standard draft for playing against humans. It feels like they lost the drafting stage, but we'll have to wait and see. The game begins! We'll take it slow, if you want any concepts explained, do ask!
A major trend in these live presentations, from AlphaGo, to AlphaStar, to now OpenAI, is finding ways to show what the AI is doing and give a better sense of the way the system works. They've got a nice visualisation for #openaifive.
So if you're new to DOTA 2, what we're going to see initially is heroes killing these little creatures to get money, and fighting in fairly local spaces. Eventually, heroes will start 'roaming' across the map, becoming more active. #openaifive
Huh. Crystal Maiden is 'jungling' - killing creatures away from the main action. We didn't see the bots do this in previous live matches, as far as I remember. That's a very big shift - I thought jungling's reward would be way too small to be learned by the AI. #openaifive
oh my god sorry, that play was ridiculous. That fight at top, the humans played as precisely as bots. Dang. Got distracted. #openaifive
Okay so what I was going to say: the humans are doing pretty well. Top left of the screen shows "last hits" which is an approximation of the resources a hero has. Red backgrounds are human players. Three of the top four are humans, which bodes really well for OG. #openaifive
The lane the humans were losing was top, where they put a very weak hero - Shadow Fiend. The AI went to fight up there several times but actually failed their first few attempts despite doing better up there overall. Instead of backing off, they just committed harder. #openaifive
One thing that's surprising me so far is how well Riki is doing - Riki is the purple blob, and what you can't tell is that he's actually invisible most of the time (when not attacking). OG are really using this feature well, and the AI isn't responding well so far. #openaifive
My suspicion is that the AI uses Riki very differently. Riki becomes visible when attacking or using items, which means there's a big psychological/information warfare aspect to using him. Sometimes not attacking or being somewhere unusual is vital. #openaifive
The real weakness we're seeing is that there are two fairly cheap items that reveal invisible heroes, but they're barely using them. Riki has gotten away with a lot so far that he probably shouldn't have. #openaifive
Haha wow. Okay. That was Crystal Maiden's ultimate, which makes ice crystals fall over a wide area. That was pretty huge. You can see the weakness in OG's style, which we also saw with Pain last year in TI8 - it's chaotic, but if you're not ahead, you can get picked apart.
Like, a lot of these plays the humans are making are really wild, this is not a regular game of DOTA. That's a good approach for playing against OpenAI if you can knock them off balance. But it can also backfire. #openaifive
A big misplay from OpenAI there. Death Prophet used her ultimate skill - you can see loads of ghosts flying around her - and they went to fight Riki... then he went invisible and ran off. We saw this back at TI8 too. #openaifive
You might hear the commentators mention "buyback" - you can spend gold to immediately come back to life, but it has various penalties attached. Humans tend to use it in extreme circumstances, so the bots using it here looks funny to a human. #openaifive
From the bots' perspective though, we know they tend to lose later games against humans at least, so paying to be engaged in the early game more probably makes sense from their perspective. (In the early game, humans usually associate it with being angry :P) #openaifive
And now we can see, suddenly, OpenAI are surging into the human base, taking base buildings. This is why they bought back. Taking a lane of buildings is serious - it shifts the balance of several game systems in their favour until it's equalised. #openaifive
The bots report a 99% likelihood of winning, but remember this is against other bots. It's not going to be the same against humans. That said - it's really bad. I think OG played a bit too sloppily in the midgame, taking a few too many risks. Not over yet though! #openaifive
OG is famous for chaos, but here there are two kinds of chaos. Sneaking off to kill the courier - good chaos, unpredictable, bots confused. Charging at a hero you think is low health and out of position - bad chaos, OpenAI is never out of position in the early game. #openaifive
I agree. The structure of the game has been really odd. There's no resource gathering, no attacking towers, no getting objectives, it's like a huge map-wide fight! #openaifive
From this game, it feels hard to tell if OpenAI have gotten better at playing long games, because although this game is going longer, the nature of the game feels like it's still ten minutes in. There's a lot of skirmishing. #openaifive
God Radiant keep losing their courier. The courier is a flying donkey that delivers items to heroes. It's really fragile and vital you keep it safe. They've lost it... four times now? It gives a lot of gold and takes minutes to respawn again. #openaifive
There's a big building in each team's base called an ancient (DOTA = Defence Of The Ancients), and when that dies, it's game over. But you need to clear the path of other buildings to make it vulnerable! OG's ancient is nearly exposed... #openaifive
The humans just "took Rosh" which means killing a very big monster who lives in a cave. Rosh drops an item that brings someone back from death once. Strategically it's usually really important. Rosh will respawn in 6-8 minutes. #openaifive
The hero who has the extra life got caught out and just died - the little tombstone shows him coming back. They want to fight but, oof, an instant reaction from one of the bots. They're thinking about fighting this though... #openaifive
More big fights, more humans dying. You see the AI have bought back again - these buybacks get less good as the game goes on, but look, the humans only have 2 heroes alive, so the bots can now kill another tower. It's working. #openaifive
They've taken two lanes, if they take a third they'll gain something called Mega Creeps - the little creatures we've seen everywhere all game? OpenAI's will become really strong. It normally spells the end of the game. #openaifive
That Echo Slam was amazing - Earthshaker (the... bear with the big.. cylinder?) can stun everyone around him. The bots walked into it a bit, and the humans managed a good defence. But I think this is OG's last chance. Their base is almost destroyed. This is it. #openaifive
This is looking like it might be over for OG. OG's heroes do a lot of damage, but it's really hard to find and lock down the bots' heroes. They have such a small margin of error now, with their base almost destroyed. #openaifive
I think that's it. The humans can't afford buybacks, and their base is almost destroyed. We should see a big push in, one last fight, and then probably a GG call, signalling a surrender and the end of the game. #openaifive
I'm not sure what to make of the new bots. They're clearly very different (Crystal Maiden jungling was especially unusual). But I also feel like OG's draft and play was very different to what we've seen from human teams facing them before. #openaifive
And it's done! The ancient falls.

😊 Humans 0 : 1 OpenAI 🤖

The best of three continues. GGWP. #openaifive
The bots are good, let me be clear. But from Game 1 I'm not sure how much OG is to blame too. OG played extremely weirdly the entire time, and we saw sometimes it worked, and sometimes it really, really didn't. I hope we see something a bit different from them next. #openaifive
Great question! BKB is an item that makes you immune to (almost all) magic damage for a few seconds. Buying a BKB is a big power spike for a hero, but it gets less effective the more it is used. Against some heroes, BKB is vital. #openaifive
Incidentally, OpenAI taking Game 1 in a best of three is a common trend, going back to last year. We'll have to see if the bots can go the distance. #openaifive
We saw last year that the bots *want* to fight. What I can't tell from this game is: was there any other option, or did OG read the bots poorly? The Chinese team at TI8 kept one player farming at all times, and he eventually became enormous. #openaifive
I'm going to take a quick break. See you in a few minutes! And please keep sending in questions/comments if you have them. #openaifive
Yeah I agree - it's natural to try and look for a shortcut or a workaround. And at times it did work, but it left them exposed later on. #openaifive
This is also the first time OpenAI have explained the system does not learn between games, which was a huge misconception that nearly every commentator/host brought up. A single game would be a drop in the bucket anyway - this system has played 45,000 years of DOTA 2. #openaifive
I know I'm here to analyse the AI, but analysing OpenAI as an org, this event feels like it was approached a lot more like an outreach opportunity than only a big showy demo, which is what most human vs. ai game things were treated like historically. It's nice. #openaifive
Alright, the teams are walking back on. Here's hoping OG can make a comeback. Something I missed during that interval: OpenAI have played three other top pro teams in preparation for this. They won Game 1 all three times. #openaifive
The hosts are asking if a particular "slot" (like OpenAI Bot 1) has a preference for a particular hero. That's such a fascinating question that shows how humans understand and interpret bot behaviour (I think the OpenAI people were a bit confused! Haha.) #openaifive
Alright, we're opening with Sniper for the humans. This is a much better pick I think. OpenAI are really good with Sniper, and it's a good hero against how OpenAI play. OpenAI also seems to respect this - they predict a 52.9% winrate from this, compared to 64% last time.
This is a good idea (a bit complicated for non-DOTA players). Riki was useful strategically, but he needed to be a bigger deal later on, because he was one of their "core" heroes who are supposed to get big and strong. Sort of. DOTA is hard. #openaifive
The humans have taken Slark, which I seem to remember appearing in one of the earlier matches (maybe back in the pre-TI showmatch). I remember thinking it wasn't a good pick but it performed okay and was in a winning game I think. #openaifive
Alright, a better draft - just a 60% win prediction from OpenAI. I like the human draft a lot more then before, but OpenAI has an incredible draft. They have a lot of stuns/disables - spells that stop a player acting temporarily. I think this is going to be bloody. #openaifive
We might be about to see an AI beat a world champion team at DOTA 2. But at the same time, the hero restrictions are significant still, OG is playing strangely, what does it actually mean to be superhuman? This comes up every time AI plays humans. Anyway. GLHF! #openaifive
Hah. The first kill of the game goes to the bots - the bot walked away knowing that the damage-over-time would kill him but I don't think anyone was quite sure themselves until it happened, many seconds later. A classic OpenAI move. #openaifive
Once again, keep an eye on the top-left of the screen to see how many resources each hero is gathering. Lots of red heroes at the top is good for the humans (in theory). #openaifive
Ceb is dying a lot on the bottom lane. This is really bad news. Death Prophet is a weak hero, especially during the early game, and the bots have really strong heroes. I'm not sure where the humans are winning right now. This might be a shorter one. #openaifive
Major differences to last game - OG aren't doing enough to disrupt the AI across the map, and their heroes are a lot weaker. The draft might have some potential later on but it really depends on whether we get there. Ceb is... oof. #openaifive
This is probably over already, sadly. OpenAI have four of the top five heroes ranked by net worth. At ten minutes in, against bots with the execution of OpenAI, this is really bad. #openaifive
We'll have to see - OG have strong heroes for later in the game, but they've given up a lot already. While we've got a breather, here's a tweet I was trying to find from 2018, about the decision not to draft on the TI8 stage. #openaifive
Heh, a nice moment there - Crystal Maiden appeared to start casting her ultimate a fraction of a second after Lion cast his stun spell. It was almost as if that was the last thing that could disable her, and that let her decide to cast it? #openaifive
You may notice that, at just 14 minutes in, we're already breaking into the human base. Viper is Godlike, meaning ten kills with no deaths. It's looking rough, folks. #openaifive
It's a single system playing the whole game - all five heroes simultaneously. #openaifive
That's the style play at the end there. The bots using items that normally a human wouldn't even use at all, let alone for that kind of skilful play - constantly flickering in and out of invisibility, to avoid being killed. #thatjusthappened #openaifive
"nice smurfs" - a smurf account is a fake account used by a high level player to get matched into lower level matches. 😄 #openaifive
It's interesting to see, here, the purpose of the win prediction. This draft had quite a better appraisal from the bots, but this is based on the assumption OpenAI is playing it. Game 1 - the worse draft, apparently - did far better. #openaifive
And that's it, folks.

😊 Humans 0 : 2 OpenAI 🤖

OpenAI Five defeat OG in a best of three. GGWP! #openaifive
Now comes the tricky bit: what headline do you write? This is the big question these events always pose for me. Was this DOTA 2? Was this superhuman? How do we make sense of achievements like this? #openaifive
Regardless, I feel very vindicated now for tweeting this during TI8, where the bots lost a best of three 2-0. #openaifive
OpenAI reveal that over the last few months, they've beaten three other pro teams, also 2-0. #openaifive

@SheeverGaming: "They seem unstoppable. That's scary?"
@blitzdota: "No, I think that's dope."
@SheeverGaming @blitzdota Regarding system structure: my understanding is that a single system controls all heroes, but the system does not explicitly have a notion of shared info. The bots can co-ordinate actions together through some 'co-operation' variable, if I remember right.
@SheeverGaming @blitzdota Here's why this is tricky: DOTA 2 has not been solved/broken/surpassed/whatever today. What we saw was some amazing stuff, but like Starcraft 2, there are still caveats. The question is how much those caveats matter to us, which is highly subjective.
If we'd played full DOTA 2 today, would that mean it was done? If OpenAI entered the International and won, would that? Does it need to screen read? It's at these times that I think we see the gulf between the public's idea of an AI milestone, and the industry's. #openaifive
"Team Spirit", that was it! Thanks to Jack for digging out the good links:
Alright Part 2, let's go. We were right, it's OpenAI as a bot team-mate alongside humans! #openaifive
Alright this is quite a big deal. As is being explained: the OpenAI bot is assuming all its teammates will act like OpenAI bots. This is why co-op play is a lot harder than playing as a team against humans. We're going to see this *live*. #openaifive
Opponents are more predictable because they respond naturally to the pressures of the game. If an enemy is ignorant or selfish, it gets punished. If an ally is the same... what can you do? We're going to see the bot put into unusual scenarios a lot faster. #openaifive
Yeah if you turned off, folks, get the heck back in here because this is cooler than any human vs. AI milestone showmatch, in my opinion. Very glad they're doing this. #openaifive
I will say, the OpenAI panelists are talking about this showing the potential for human-AI co-operation in the future - that's a bit of a stretch, for me! But you know, it's their party, they're allowed to get poetic if they want to, I suppose 😉
In all fairness, if ANGELINA had just won an international games award and I was on a panel talking about it I'd have gotten a lot more dramatic than that. They're probably all on adrenaline rushes right now.
I think it can help. There are meta-features about OpenAI's play - patterns you can see across all the games. That said, maybe OG did watch those matches! As n0tail just said - it's easy to get sucked into the game and forget your strategy. #openaifive
me, watching openai: good grief this gentleman is hyperbolic

me, on stage at the igf awards 2022: I JUST WANT TO THANK MY TAMAGOTCHI, BRIAN, FOR TEACHING ME TO LOVE AND CARE FOR DIGITAL BEINGS, BRIAN THIS ONE'S FOR YOU BUDDY

n0tail is joking about the AI's "ward" placement - little items that provide constant vision of an area. The AI is quite bad at them, but you can't tell this while playing, they're invisible. He *felt* they had good wards, but that could just be raw skill. #openaifive
Alright here we go. A mix of pro-level and commentator-level (definitely above my skill level), two humans per team, with three OpenAI bot teammates. This is, to my knowledge, new for AI game playing competitions. #openaifive twitch.tv/openai
Purge: "If you played five games could you take a game off them?"
n0tail: "For sure. And if we played fifty games, we could start winning [series]."
P: "But in the time it would take you to play fifty games, they could train on another...."
N: "...for sure."
#openaifive
From n0tail's perspective, their understanding of vision is a huge weakness, he's mentioned this many times. A lot of other very human/psychological tells too: you can just tell, by how a hero moves, when they've placed an invisible ward sometimes. But OpenAI doesn't. #openaifive
On taking tips back to the real world: the hero pool makes it hard to take individual lessons back, because the strategic space is much smaller. But he thinks their overall approach, "death ball", is very good. The problem? "No-one death balls as good as [the bots]." #openaifive
Extremely sharp insights from some of these pros. Blitz (who was on the panel, now playing) has a really good understanding of the system. n0tail now, also, showing a very rich understanding of how a bot sees into the future. The buybacks "were just bad". #openaifive
This is hard. It worked out today. OG won their world title through bizarre and unpredictable gambles. Would they have won the next day? Who knows. Buying back won today, and it worked *in general* against bots. Would it work *in general* against humans?
This is the problem with having 45,000 years of DOTA experience, and only a handful of live human examples. When we see e.g. draft predictions, responses to unusual human strategies, how can we really understand this? Maybe the rest of their play just compensates for it?
Alright, we're taking a brief break - and so will I. When we get back, co-op matches! Keep questions/comments coming. #openaifive
It's my great pleasure! Sorry that my commentary has been breathless and all over the place - there's so much to talk about, and things are happening so fast. I hope my perspective is useful - take it with a pinch of salt :) #openaifive
"We're doomed." n0tail says. This is why the co-op match (starting now) is so important for events like this. It shows that AI can do more than just beat humans at stuff. The future of AI does not have to just be a march towards obsolescence for humans. #openaifive
Oh god this is great, we have audio from all the players too. "I think... oh I think we're supposed to help them [their bot teammates]."
"Did you tell them where to go?"
"No I think they just know.... yo... yo fight them!"
#openaifive
Oh god it's a minute in and I already love this more than every AI showmatch that's ever happened. Listening to human players' frustration at not understanding what the bots are doing and why they aren't following them, this is great. #openaifive
So what are we looking out for here? It's not just about winning now. Listen for the humans trying to figure out what to do. Are the bots taking the lead? Are the humans following, or taking initiative? These are the exciting questions. This is the HCI session. #openaifive
Cap came to the mid lane, having found an invisibility rune. To him, this was the perfect time to kill the hero at mid lane, it's a classic move. His bot teammate instantly left and went to fight somewhere else. 😄
"What... why is Razor here!?"
#openaifive
You might see a lot of these little sticks in the ground, called wards, we mentioned them earlier. The bots have apparently instantly bought them, and are placing them... in... odd places. #openaifive
Putting this right after the pro match is so neat, because these bots just beat the world champions. Watching them just stare blankly at humans desperately trying to co-operate (and vice versa) really highlights how impossibly hard this is compared to versus play. #openaifive
This can't be as flashy or headline-grabbing as beating the world champions, and there's less fanfare, but to me this is the best bit of the day so far. This is the match I want all of us to analyse later and write about. #openaifive
More details on "why are there five wards watching the tier 1 tower top"
"Now that is... not... the optimal Shrapnel." Cap says about his bot teammate. Of course, the bot might've just been expecting a response from Cap!

Other players are taking a lead from the bots: "Given that no-one else is up here, Sheev, I think we're in the wrong spot."
I'm not sure whether the bots will be thrown off more by their opponents or their teammates. Because the bots do get confused when they're behind, but here neither team is *behind* per se, they're just so far away from where the bots would normally be playing. #openaifive
Commentators now speculating about the bots intentions, "It seemed like they wanted Sheever [human] to rotate to the top lane?" Of course, it's arguable whether it makes sense for the AI to want anything at all. Expectation maybe? Even then... #openaifive
"What if Cap just decides to go full Rambo? Will the bot follow him into death?"
Of course, probably not - the bots are good at retreating. But this raises interesting questions about what co-op AI look like. There are many different ways to approach this problem. #openaifive
Really interesting discussion from n0tail right now, in response to whether humans might learn from OpenAI. He says that he thinks DOTA's traditional structure (of putting lots of resources on a few 'core' players) is appealing to human nature. #openaifive
I don't necessarily believe that, but I do think that some aspects of how DOTA is played are burned so deep into the game's culture that it would be almost impossible to change. That's always a sign of a game that AI can break wide open, by ignoring preconceptions. #openaifive
While we're watching this game wrap up, of course I'm now wondering what Part 3 might be. Nothing obviously springs to mind, except maybe OpenAI as a coach, like an in-game overlay giving advice to a player, or possibly a commentator. #openaifive
n0tail: "Death Prophet is *the* OpenAI hero... all the spells, if you look at them, are extremely good at the pre-20 minute period." An interesting insight. Also a good example of preconceptions - normally we see DP as a weak hero early on. #openaifive
Fascinating to hypothesise why the bots decide to fight or not in this mode. I guess at each stage, the bots are always assuming everyone is a bot, and even though the humans consistently do weird stuff, the bots always assume they'll make a perfect decision next. #openaifive
Part of me kind of wishes the humans weren't labelled and the casters had to guess who was who. Although given some of the players so far, that might not have been too hard... #openaifive
Yeah this is a good point - there is a sense of wordless communication in public games, even in pro games in very short reaction times. Of course, the bots probably know a human's skill is off cooldown before the human realises! #openaifive
Things like searching in the trees for heroes feels like it might be one of the harder things to learn. By learning against bots, you won't necessarily learn what humans 'naturally' do while escaping, which has a lot to do with weird subconscious decisionmaking. #openaifive
Haha, fantastic. The humans could make requests of the bots in the early game, but after a period of time the bots just tell them "Sorry, too late for that now!" #openaifive
Interesting to watch some of these teamfights, how humans and bots initiate differently, and how the bots respond (or don't). In an earlier fight there, Cap actually made a great initiation onto one of the bots, and there was *some* followup from his allies. #openaifive
This is why co-op is so interesting. What would we want the bots to do here ideally? Intuit the human plan and follow-up? Take the lead and try to get the humans to understand? It's sort of neither here (which is understandable and expected). #openaifive
Bearing in mind the Dire predicted an 80% win chance earlier, we can now see a huge 11k advantage for Radiant and some of Dire's base destroyed. The humans are definitely not the teammates the bots were hoping for! #openaifive
Oh my goodness, those were some costly fights! The 11k lead is cut to 4k suddenly. Humaaaaaaans. #openaifive
These are such dynamic fights, partly because the bots are so perfect but the human factor is entirely unpredictable - not just to us as observers, but also to the bots. This is so spicy. #openaifive
Radiant take mega-creeps. "I thought Dire for sure had this game?" n0tail says. But it looks like Radiant might have it. But now - deaths, buybacks, their main core hero is exposed, thank god there's no extremely unpredictable human factor involved in this anywhere! #openaifive
Ahaha that was amazing. Cap comes in with a huge echo slam (remember? the angry bear thing?) and the commentator gets excited oh man he's using his spells and... there's... no support from anyone. Hilarious. This is amazing. "The bots are like, no Cap, that's not it." #openaifive
One issue with mega-creeps is that, even with 45,000 years of play, mega-creeps are not super common. And here, in particular, with two human allies, we are way outside anything OpenAI probably has seen before. Some really odd things beginning to happen. #openaifive
Ah there goes Sniper - he was key to the Dire defence. This might be it for them now. No buyback available either. Still, a great match. A fantastic idea, I really enjoyed that.

😊🤖 Humans/AI 1 - 0 Humans/AI 🤖😊

GGWP

#openaifive
Alright, it's pretty late here, I think I'm going to hang around to see what Part 3 is, and then possibly have to call it a day. Overall I like the vibe of this event a lot - it's important to show different angles on AI, to provide more tape, more examples. #openaifive
What I think is important here - OpenAI Five is not going to change humanity. It plays DOTA 2. But this is a much better showcase of an AI system than we usually get, and to OpenAI's credit, they're a bit better at exposing their systems to risk in public. #openaifive
Alright so in terms of the original blog: "competence, scalability, and ability to enhance humans". We've seen the first and the third item there. Could scalability possibly be limited public access to the bots? Surely not. #openaifive
Amazing quote from Blitz on playing with the bots: "Whatever lane I went to, it felt like I was making the wrong play? It felt pretty bad in that regard. But when I went somewhere and they went with me, it felt good. It felt like trying to solve a puzzle while playing the game."
Oh shit, scalability... really does appear to mean "for everyone". Part 3 being discussed now. "We're announcing... Arena." #openaifive
OpenAI Five Arena - you can sign up to play with OpenAI's bots, or against them as a team of five, from next week. Go online now to register: arena.openai.com! #openaifive
Listen, look, this is what big AI orgs with funding+reach should be doing. Beating the world champions - whatever, honestly. What was important today:
🔹 Live play with humans on stage: broken, weird, hilarious.
🔹 Open access to the AI, to experience the system first hand.
Whatever we think of the limitations or restrictions, this is a great example of using the resources of a massive organisation well: doing slightly riskier things, in front of a big audience, and letting people get their hands dirty and understand first-hand. #openaifive
Alright I think that's all from me for tonight. I need time to chew over the event, it was unusual as AI events go. It's tempting to see this in terms of the headline - they beat the world champions. What was important to me was what this offered the community. #openaifive
More AI from me:

🔹 A position paper about how AI can impact esports beyond just playing them: arxiv.org/abs/1706.03122
🔹 My blog, where I write about AI, games, and more: possibilityspace.org
🔹 @angelinasgames, my research into AI game design!

Thank you for reading! 💜
Missing some Tweet in this thread?
You can try to force a refresh.

Like this thread? Get email updates or save it to PDF!

Subscribe to mike cook
Profile picture

Get real-time email alerts when new unrolls are available from this author!

This content may be removed anytime!

Twitter may remove this content at anytime, convert it as a PDF, save and print for later use!

Try unrolling a thread yourself!

how to unroll video

1) Follow Thread Reader App on Twitter so you can easily mention us!

2) Go to a Twitter thread (series of Tweets by the same owner) and mention us with a keyword "unroll" @threadreaderapp unroll

You can practice here first or read more on our help page!

Follow Us on Twitter!

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just three indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3.00/month or $30.00/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!