Pre-Match Panel Discussion | OpenAI Five Finals (2/6)

OpenAI · Intermediate ·📰 AI News & Updates ·7y ago

Key Takeaways

The video discusses OpenAI Five, a Dota 2 playing AI, and its development, training, and performance, as well as its upcoming match against OG, the world champions.

Full Transcript

yeah thank you so much Greg and indeed we are getting to it because about what's gonna happen today we are also now joined by Brooke and David and David I'm actually I'm gonna start with you because I want to know the history of the project it's a project that's been going on for a while obviously we saw most recently at t i8 but but talk to me about the project in the history of it yeah so we started this project about two years ago in early 2017 we first worked on 1v1 dota as a way to get started just to get off the ground in that ti toys house 2017 we were able to beat the top pros at 1v1 dota then we spent a while getting ready for five you five and we started training open air five in June 2018 by August 2018 we played some tournaments and went to TI ITI pain gaming beat us and was you know open i was not yet ready to play it better than per level it played pretty well we were pretty happy with how it went and now it's been eight months since then here we are opening five has learned as grown has learned some new stuff figured out these strategies and we're excited to see where it is now what are the biggest changes that have happened over the last eight nine months so some of what we worked on was trying to add more heroes to the pool it was it was something that we definitely attempted to address and we did manage to see a lot of improvement there I just kind of didn't get to the level we thought maybe og was going to play back we also worked on a lot of research improvements to just kind of create a more fluid training process in the hopes that we would be able to gain a lot of performance from those metrics now let's say we've seen obviously the events unfold already blitz you have experience firsthand playing against AI both want to be one and also in a team setting what are you what do you expect it today so I was the first person that played it at ti right and I feel like I'm always like the whipping boy like I got brought in I got absolutely destroyed and that was depressing and then they had me played the five on five it was I was casting it at first and then they're like would you like to play against it I feel like I got trapped again like I'll stop it I was pretty confident if I'm honest here III remember I talked to guys write it plays it like max 4k and I just got rekt but I sort of played my teammates and now the fact that you know I was able to beat some teams that it was able to take pain I think it was like a 40 minute game that I casted I think it's really impressive how about you purse with your experience as well it's definitely very interesting to play against it we got to play it match yesterday briefly and it doesn't really feel like the flashiest thing it just kind of feels like it's always reacting perfectly to your movements like at one point I tried to prevent you from dying for the fifth time in a lane fever and every time I would run at the Earthshaker he would run away and then I'd be like okay he's wasting my time when I'm run back and then it would follow me it just kind of did that we just kind of did this for a while and it just feels really annoying it just feels like a step ahead of you it's not necessarily super out playing you but it's always out playing you just because they played so many matches so it's definitely interesting to play against all right David how does AI work cuz you know weird let's say that he was expecting that it's a play at 4k level we have experienced that it's probably a bit higher than that but how how does a I played dota 2 yeah so it's a little bit different from how humans see it and we have a slide to explain how the opening I 5 interprets the game so humans see this image on the left with a nice picture of the characters walking around the AI see the picture on the right which is just a giant pile of numbers and we know that some of these numbers means a hero's health or a some characters position or the items in someone's inventory but all the AI sees is this giant list of numbers it doesn't have any idea about the rules of dota to begin with which what different items are it just sees these different these different numbers and has to figure out what they mean so how did it get to learn nota - is the can't read yeah so that's kind of the question that we get a lot and a lot of people will watch our BOTS play dota and be like well what's the big deal I also learn how to play dota and I think the difference is that essentially that is that the AI takes a completely different path to get there it doesn't get to read the tool tips read about abilities watch purge videos to figure out how to play a new hero it's essentially learning it just by going into the game trying something seeing what happens and even further than that like let's say you have a new ability you're trying to learn and you know that it's done so you know there are other abilities that's done you know what that looks like you know that it's great for you if you're able to stun the enemy hero what the AI does is they go in there they cast the ability and they see that some numbers changed and they don't know if that's good or bad eventually they might learn out they might learn that oh that hero can't move and they see that by it's not moving they have to kind of like recognize that pattern and so that's kind of like how it builds on itself and it has to do this all all the game mechanics so it starting out is pretty inefficient right it takes us a long time to kind of even get it to last-hit which is a pretty basic mechanic that most people are able to pick up quickly the upside of that is that this training is all general so for example if we were to take out the dota lair and swap in a different environment such as the robotic camp which I think we have a slide floor to show and essentially you you take this environment where you've got the fingertip positions you've got camera vision and things like that and you can also translate that into these numbers which is you can then feed into the AI and then they will attempt to solve that problem too so it's something they know how to put they don't know that they're playing dota necessarily they just know that they are solving a task so opening I 5 is focused on dota 2 at the moment but what you're saying now is that the implications for for for use outside of dota 2 that can have some very big impacts on the world yeah and that's definitely one of the reasons why for example we are still working within a limited hero pool is because there are changes like we could learn a hero it's not that we think that it can't be done it's more that we are looking about making the training process general and making changes that we think will will help overall working on AI not just specific to dota how do we think it's gonna impact dota 2 I mean there was definitely things that we learned by watching the 1d ones for example there were things like oh is it worth it to use fairy fires just for region or there's other smaller get the the mango like bring mangoes eat them instantly your gold is lost but you have mana back which ultimately helps you and now when we see one we once that's very consistent it even happens in regular games right now so there are little advantages that you do get by watching something very different play it from outside the box basically it always helps same thing happens in pro players like if you see a pro player do something new they've never thought it before you're willing to trust it because it's different it's something you respect doing it so it's there's always going to be similar benefits to is it is it gonna be coexisting or are the AI taking over I don't know about that but there's a lot that you can learn like when when you play a high level game with dota 2 and you see somebody do something cool it should intuitively make sense and the first time I played the 1v1 bought like Kevin said I was wondering I was still like on a meta that was way too far behind like I'd buy tangos and I'd buy clarity's and I'd realized that every single time I did that I was playing slower and slower and slower and so the only way to match it was to copy it and so he started developing this meta where it was pretty much just salves and mangoes you didn't even buy any other items because it was less efficient to buy other rate bands or other boots in the shadowfiend versus shadowfiend matchup you would just keep varying region until somebody want or lost and that's something that's replicated nowadays not maybe not to like that same extreme because you know the game progresses past that 1v1 phase but you're still going for that quick regen playing as fast as humanly possible a quick question so based on the current patch in some of the previous tournaments we were playing on an older version as the game but right now we're playing on the current version of the game how difficult it is when a new patch comes out for the for the bots to it adjusted that you guys start off from scratch or you take like the previous knowledge and then let them read them yeah so we take the previous spot and just throw it in the new environment we don't tell it or anything that changed but it eventually learns that the numbers don't change and quite the way I thought they would change and it slowly adapts it takes a week or a few days after we've done the work to write the update the code and make the code good so how long have it they've been training on the current patch it's is like a new patch comes out you guys instantly put them to work to play or is it like maybe at the late thing typically it's not instant just because it's work on our part we really don't know what another patch is gonna come out after that we don't want to bring them up to speed I think that they've been training on this patch for at least a few weeks now and there there has been some interesting feedback from our test teams who when we switched to a new patch they for example when we switched in 721 the AI started really liking viper and so we asked like you know what were the changes with Viper that they really liked and they were like oh it was no big deal I was like minor AoE improvements and then a couple weeks after that they're like oh you can fight for jungle jungle now it's really really strong and so it was kind of like they they were talking to us and saying like they'd catch on pretty quickly there's definitely some things that they have more trouble like switching over to but for the most part yeah one of the things they have trouble switching to like other more complicated heroes or something yeah it's typically when abilities change they have a little bit of trouble with that that's kind of like while it didn't make it past our patches just because there were there were significant changes there so when they have to relearn their playstyle that becomes a lot harder for them now today's event is gonna be a three part event we're gonna start with with with part one and David Carr part one is cool competitor so what did open a I do to prepare for today so when you have a very high skill dota 2 bot you want if you guys how good it is so the natural thing to do is play against some of the very best teams so we've invited og the world champions from ti last year to push our bots to limit and see just how good it is so we'll play a best-of-three against them and see how it goes and Brooke normally when a professional dota 2 team plays a tournament you know there's a little bit of screaming before had a little bit of practice and opening I did that too yeah so coming out of Ti we played three other semi pro and pro teams to kind of gauge where we're at because for the most part that is training just with itself so it's a little bit hard to see once you put a human team in there if it's going to play the same if it's going to you know have a lot of issues with it so throughout the event wall we'll talk about the other teams that we played and also go over how those games went yeah that will be very exciting to see and also exciting to see how og is gonna do today because we will have ot playing on stage against AI what uh what can we say about og og obviously they want ei8 they are back with Ana again after a long hiatus they have shown that with Anna back they are also back qualifying for the major blitz looking like like a very strong team again yeah felt like they were I mean all respect to iti ltw who they were using as a stand-in but it felt like they were just waiting for and it come back at some point because I mean this guy it just completes their lineup you can see instantly that everyone's a lot more energized these guys want TI off of maybe the most magical run that I've ever seen before like that was the underdog story of underdog stories for them to go through what they had to go through they force the game five against LGD I don't know how strong they are right now I mean they were able to go through which i think is the hardest qualifier make it to the major I feel like they're coming into their own maybe it's a bad time for hoop in the i-5 to play against them because I feel like they're peaking at the right moment I mean you saw right there in terms of overall winnings no tails got I mean no-tell completes most of that but 17 million dollars for dota I think that makes them like the second or third highest team in terms of prize pool so they've definitely been some of the most storied teams they had crazy runs during the major system when it first came out they won the the vast majority of those and winning TI this last year was really the the key stone that they needed to be like were probably the most winning team of all time so it's very exciting to have them here it is very exciting to have them here and I believe also we have them are ready to take the stage
Watch on YouTube ↗ (saves to browser)
Sign in to unlock AI tutor explanation · ⚡30

Playlist

Uploads from OpenAI · OpenAI · 26 of 60

1 Robots that Learn
Robots that Learn
OpenAI
2 Emergence of Grounded Compositional Language in Multi-Agent Populations
Emergence of Grounded Compositional Language in Multi-Agent Populations
OpenAI
3 OpenAI + Dota 2
OpenAI + Dota 2
OpenAI
4 Dendi vs. OpenAI at The International 2017
Dendi vs. OpenAI at The International 2017
OpenAI
5 Competitive Self-Play
Competitive Self-Play
OpenAI
6 Learning a Hierarchy
Learning a Hierarchy
OpenAI
7 Physical Spam Detection
Physical Spam Detection
OpenAI
8 Ingredients for Robotics Research
Ingredients for Robotics Research
OpenAI
9 OpenAI Five
OpenAI Five
OpenAI
10 OpenAI Five: Dota Gameplay
OpenAI Five: Dota Gameplay
OpenAI
11 Learning Dexterity
Learning Dexterity
OpenAI
12 Learning Dexterity: Uncut
Learning Dexterity: Uncut
OpenAI
13 OpenAI Five Benchmark: Post-Game Analysis
OpenAI Five Benchmark: Post-Game Analysis
OpenAI
14 Investigating Model Based RL for Continuous Control | Alex Botev | 2018 Summer Intern Open House
Investigating Model Based RL for Continuous Control | Alex Botev | 2018 Summer Intern Open House
OpenAI
15 Generative Modelling | Sadhika Malladi | 2018 Summer Intern Open House
Generative Modelling | Sadhika Malladi | 2018 Summer Intern Open House
OpenAI
16 A pathway to more efficient generative models | Will Grathwohl | 2018 Summer Intern Open House
A pathway to more efficient generative models | Will Grathwohl | 2018 Summer Intern Open House
OpenAI
17 Learning Dexterity | Alex Ray | 2018 Summer Intern Open House
Learning Dexterity | Alex Ray | 2018 Summer Intern Open House
OpenAI
18 Robust Vision-Based State Estimation | Hsiao-Yu 'Fish' Tung | 2018 Summer Intern Open House
Robust Vision-Based State Estimation | Hsiao-Yu 'Fish' Tung | 2018 Summer Intern Open House
OpenAI
19 Using Semantic Trees In Place of Sentences | Munashe Shumba | OpenAI Scholars Demo Day 2018
Using Semantic Trees In Place of Sentences | Munashe Shumba | OpenAI Scholars Demo Day 2018
OpenAI
20 Reinforcement Learning with Prediction-Based Rewards
Reinforcement Learning with Prediction-Based Rewards
OpenAI
21 OpenAI Spinning Up in Deep RL Workshop
OpenAI Spinning Up in Deep RL Workshop
OpenAI
22 Arena Announcement and Closing | OpenAI Five Finals (6/6)
Arena Announcement and Closing | OpenAI Five Finals (6/6)
OpenAI
23 Co-Op Match | OpenAI Five Finals (5/6)
Co-Op Match | OpenAI Five Finals (5/6)
OpenAI
24 OpenAI Five vs. OG, Game 2 | OpenAI Five Finals (4/6)
OpenAI Five vs. OG, Game 2 | OpenAI Five Finals (4/6)
OpenAI
25 OpenAI Five vs. OG, Game 1 | OpenAI Five Finals (3/6)
OpenAI Five vs. OG, Game 1 | OpenAI Five Finals (3/6)
OpenAI
Pre-Match Panel Discussion | OpenAI Five Finals (2/6)
Pre-Match Panel Discussion | OpenAI Five Finals (2/6)
OpenAI
27 Opening Keynote | OpenAI Five Finals (1/6)
Opening Keynote | OpenAI Five Finals (1/6)
OpenAI
28 OpenAI Robotics Symposium 2019
OpenAI Robotics Symposium 2019
OpenAI
29 OpenAI Scholars Demo Day 2019
OpenAI Scholars Demo Day 2019
OpenAI
30 Multi-Agent Hide and Seek
Multi-Agent Hide and Seek
OpenAI
31 Solving Rubik’s Cube with a Robot Hand: Uncut
Solving Rubik’s Cube with a Robot Hand: Uncut
OpenAI
32 Solving Rubik’s Cube with a Robot Hand: Perturbations
Solving Rubik’s Cube with a Robot Hand: Perturbations
OpenAI
33 Solving Rubik’s Cube with a Robot Hand
Solving Rubik’s Cube with a Robot Hand
OpenAI
34 Music Generation | Christine Payne | OpenAI Scholars Demo Day 2018
Music Generation | Christine Payne | OpenAI Scholars Demo Day 2018
OpenAI
35 Deephypebot | Nadja Rhodes | OpenAI Scholars Demo Day 2018
Deephypebot | Nadja Rhodes | OpenAI Scholars Demo Day 2018
OpenAI
36 Physics Net | Ifu Aniemeka | OpenAI Scholars Demo Day 2018
Physics Net | Ifu Aniemeka | OpenAI Scholars Demo Day 2018
OpenAI
37 Art Composition Attributes + CycleGAN | Holly Grimm | OpenAI Scholars Demo Day 2018
Art Composition Attributes + CycleGAN | Holly Grimm | OpenAI Scholars Demo Day 2018
OpenAI
38 Generating Emotional Landscapes | Hannah Davis | OpenAI Scholars Demo Day 2018
Generating Emotional Landscapes | Hannah Davis | OpenAI Scholars Demo Day 2018
OpenAI
39 Looking For Grammar In All The Right Places | Alethea Power | OpenAI Scholars Demo Day 2020
Looking For Grammar In All The Right Places | Alethea Power | OpenAI Scholars Demo Day 2020
OpenAI
40 Semantic Parsing English to GraphQL | Andre Carerra | OpenAI Scholars Demo Day 2020
Semantic Parsing English to GraphQL | Andre Carerra | OpenAI Scholars Demo Day 2020
OpenAI
41 Long term credit assignment with temporal reward transp… | Cathy Yeh | OpenAI Scholars Demo Day 2020
Long term credit assignment with temporal reward transp… | Cathy Yeh | OpenAI Scholars Demo Day 2020
OpenAI
42 Social learning in independent multi-agent reinfor… | Kamal N’dousse | OpenAI Scholars Demo Day 2020
Social learning in independent multi-agent reinfor… | Kamal N’dousse | OpenAI Scholars Demo Day 2020
OpenAI
43 Quantifying Interpretability of Models Trained on Coi… | Jorge Orbay | OpenAI Scholars Demo Day 2020
Quantifying Interpretability of Models Trained on Coi… | Jorge Orbay | OpenAI Scholars Demo Day 2020
OpenAI
44 Towards Epileptic Seizure Prediction with Deep Network | Kata Slama | OpenAI Scholars Demo Day 2020
Towards Epileptic Seizure Prediction with Deep Network | Kata Slama | OpenAI Scholars Demo Day 2020
OpenAI
45 Universal Adversarial Perturbations and Language M… | Pamela Mishkin | OpenAI Scholars Demo Day 2020
Universal Adversarial Perturbations and Language M… | Pamela Mishkin | OpenAI Scholars Demo Day 2020
OpenAI
46 Introductions by Sam Altman & Greg Brockman | OpenAI Scholars Demo Day 2020
Introductions by Sam Altman & Greg Brockman | OpenAI Scholars Demo Day 2020
OpenAI
47 Introduction by Sam Altman | OpenAI Scholars Demo Day 2021
Introduction by Sam Altman | OpenAI Scholars Demo Day 2021
OpenAI
48 Breaking Contrastive Models with the SET Card Game | Legg Yeung | OpenAI Scholars Demo Day 2021
Breaking Contrastive Models with the SET Card Game | Legg Yeung | OpenAI Scholars Demo Day 2021
OpenAI
49 Large Scale Reward Modeling | Jonathan Ward | OpenAI Scholars Demo Day 2021
Large Scale Reward Modeling | Jonathan Ward | OpenAI Scholars Demo Day 2021
OpenAI
50 Words to Bytes: Exploring Language Tokenizations | Sam Gbafa | OpenAI Scholars Demo Day 2021
Words to Bytes: Exploring Language Tokenizations | Sam Gbafa | OpenAI Scholars Demo Day 2021
OpenAI
51 Learning Multiple Modes of Behavior in a Continuous… | Tyna Eloundou | OpenAI Scholars Demo Day 2021
Learning Multiple Modes of Behavior in a Continuous… | Tyna Eloundou | OpenAI Scholars Demo Day 2021
OpenAI
52 Scaling Laws for Language Transfer Learning | Christina Kim | OpenAI Scholars Demo Day 2021
Scaling Laws for Language Transfer Learning | Christina Kim | OpenAI Scholars Demo Day 2021
OpenAI
53 Contrastive Language Encoding | Ellie Kitanidis | OpenAI Scholars Demo Day 2021
Contrastive Language Encoding | Ellie Kitanidis | OpenAI Scholars Demo Day 2021
OpenAI
54 Characterizing Test Time Compute on Graph Structur… | Kudzo Ahegbebu | OpenAI Scholars Demo Day 2021
Characterizing Test Time Compute on Graph Structur… | Kudzo Ahegbebu | OpenAI Scholars Demo Day 2021
OpenAI
55 Studying Scaling Laws for Transformer Architecture … | Shola Oyedele | OpenAI Scholars Demo Day 2021
Studying Scaling Laws for Transformer Architecture … | Shola Oyedele | OpenAI Scholars Demo Day 2021
OpenAI
56 Feedback Loops in Opinion Modeling | Danielle Ensign | OpenAI Scholars Demo Day 2021
Feedback Loops in Opinion Modeling | Danielle Ensign | OpenAI Scholars Demo Day 2021
OpenAI
57 Creating a Space Game with OpenAI Codex
Creating a Space Game with OpenAI Codex
OpenAI
58 “Hello World” with OpenAI Codex
“Hello World” with OpenAI Codex
OpenAI
59 Talking to Your Computer with OpenAI Codex
Talking to Your Computer with OpenAI Codex
OpenAI
60 Data Science with OpenAI Codex
Data Science with OpenAI Codex
OpenAI

The video discusses OpenAI Five, a Dota 2 playing AI, and its development, training, and performance, as well as its upcoming match against OG, the world champions. The AI has learned and grown, figuring out new strategies and improving its performance. The video provides insights into AI development, training, and performance, and highlights the potential of AI in gaming and other applications.

Key Takeaways
  1. Start with a basic AI model
  2. Train the AI on a large dataset
  3. Fine-tune the AI for specific tasks
  4. Test the AI against human opponents
  5. Analyze the AI's performance and adjust its training
💡 The AI's ability to learn and adapt to new situations and environments is a key factor in its success, and this ability can be applied to other areas beyond gaming.

Related Reads

📰
Learn AI Training from Industry Experts at Visualpath
Learn AI training from industry experts at Visualpath to future-proof your career
Dev.to · kalyan visualpath
📰
AI Theatre: The Gap Between Talking About AI and Actually Using It
Learn to identify the gap between discussing AI and actual implementation, and why it matters for professionals
Medium · Cybersecurity
📰
How to Structure Content for AI-First Indexing: 7 Rules That Get You Cited
Learn to structure content for AI-first indexing to increase citation chances, as Google AI Overviews now appear in 47% of US search results
Medium · AI
📰
The Last Premium: What Stays Expensive When Thinking Gets Free
Discover what stays expensive when AI commoditizes human thinking and how it impacts various industries
Medium · Data Science
Up next
Matt and Ryan Have a Chat: Microsoft's Big Week, UGREEN NAS & Rivian R1S
Matt Williams
Watch →