1 00:00:00,000 --> 00:00:02,760 All right, so today we're going to be looking at this paper about 2 00:00:02,760 --> 00:00:07,240 logicity. It's all about this virtual city and how they're using it to push 3 00:00:07,240 --> 00:00:08,400 the limits of AI. 4 00:00:08,480 --> 00:00:13,600 Yeah, it's pretty wild stuff. Like imagine a city where AI is not just dodging 5 00:00:13,600 --> 00:00:17,380 things, but actually like getting the whole concept of traffic laws and 6 00:00:17,380 --> 00:00:20,620 interacting safely and smoothly with everything around it. 7 00:00:20,720 --> 00:00:24,820 Yeah, it's not just like self-driving cars, right? It's about AI. They can 8 00:00:24,820 --> 00:00:29,280 figure out any complex environment where there are rules and it needs to work 9 00:00:29,280 --> 00:00:32,120 with other things in that environment. Exactly. So what's this 10 00:00:32,120 --> 00:00:33,880 Neurosynvolic AI all about? 11 00:00:34,040 --> 00:00:38,640 So it's often called NECI AI. Okay. It's trying to combine the strengths of 12 00:00:38,640 --> 00:00:42,600 deep learning, which is really good at recognizing patterns with symbolic 13 00:00:42,600 --> 00:00:47,480 reasoning, which is great at logic and rules. It's like, uh, it's kind of like 14 00:00:47,480 --> 00:00:51,080 how humans use both intuition and experience and logical thinking when 15 00:00:51,080 --> 00:00:52,040 we make decisions. 16 00:00:52,240 --> 00:00:55,240 That makes sense. The paper mentions first order logic. 17 00:00:55,240 --> 00:00:56,240 First order logic, yeah. 18 00:00:56,240 --> 00:00:58,840 Is that how they're building this virtual city and its rules? 19 00:00:58,840 --> 00:01:04,840 Yeah. So first order logic or FOL is like a set of building blocks for 20 00:01:04,840 --> 00:01:08,800 creating rules. So it lets the researchers define the rules of the 21 00:01:08,800 --> 00:01:12,920 city in a way that's flexible and powerful. And the best part is it uses 22 00:01:12,920 --> 00:01:16,560 concepts that apply no matter what the city looks like or what kinds of things 23 00:01:16,560 --> 00:01:16,960 are in it. 24 00:01:16,960 --> 00:01:20,880 Okay. So instead of programming a ton of if then statements, they're 25 00:01:20,880 --> 00:01:24,480 creating a system that can understand and use general rules. 26 00:01:24,480 --> 00:01:28,400 Yeah. And it doesn't matter what situation it's in. Like teaching an AI, 27 00:01:28,400 --> 00:01:32,600 the idea of yield rather than saying, if you see a red octagon and stop, 28 00:01:33,080 --> 00:01:36,640 so then the AI can use that yield rule, even if it runs into some weird new 29 00:01:36,640 --> 00:01:37,800 sign it's never seen. 30 00:01:37,800 --> 00:01:40,800 So it's all about teaching it the logic behind the rules, not just memorizing 31 00:01:40,800 --> 00:01:45,440 commands. I see why this is so cool. How do they even test if the AI is 32 00:01:45,440 --> 00:01:46,480 learning these rules? 33 00:01:46,600 --> 00:01:49,600 So they've got two main testing grounds inside logicity. 34 00:01:49,640 --> 00:01:50,000 Okay. 35 00:01:50,000 --> 00:01:54,480 Safe path following or SPF and visual action prediction or VAP. 36 00:01:54,480 --> 00:01:58,640 Okay. With SPF, the AI has to navigate a long path through the city trying to 37 00:01:58,640 --> 00:02:01,800 keep costs down and most importantly not break any rules. 38 00:02:01,800 --> 00:02:02,080 Oh, wow. 39 00:02:02,080 --> 00:02:05,000 Yeah. It's like a driver's test for AI. 40 00:02:05,000 --> 00:02:05,360 Okay. 41 00:02:05,360 --> 00:02:07,560 But in an environment that's always changing. 42 00:02:07,560 --> 00:02:09,360 That's a good analogy. What about VAP? 43 00:02:09,360 --> 00:02:11,600 VAP. That's all about prediction. 44 00:02:11,600 --> 00:02:11,880 Yeah. 45 00:02:11,880 --> 00:02:17,280 So the AI gets a single snapshot of the city and it has to predict what each 46 00:02:17,280 --> 00:02:21,320 thing in that scene will do next based on where they are now, the rules of the 47 00:02:21,320 --> 00:02:23,280 road and any other important info. 48 00:02:23,280 --> 00:02:27,760 Oh, so it's like if you pause a busy intersection and try to guess what will 49 00:02:27,760 --> 00:02:32,320 happen next, does the AI get a clear view of everything? 50 00:02:32,320 --> 00:02:32,760 Nope. 51 00:02:32,760 --> 00:02:33,160 Really? 52 00:02:33,160 --> 00:02:35,080 And that's what makes VAP so interesting. 53 00:02:35,080 --> 00:02:39,400 They actually added realistic noise to what the AI sees, like what we would see 54 00:02:39,400 --> 00:02:43,600 in the real world. Stuff like, you know, a person mostly hidden behind a tree or a 55 00:02:43,600 --> 00:02:47,840 car in the shadows. It has to deal with not having all the information, just like 56 00:02:47,840 --> 00:02:48,560 we do sometimes. 57 00:02:48,560 --> 00:02:50,600 Oh, well, that's an extra layer of difficulty. 58 00:02:50,600 --> 00:02:50,960 Right. 59 00:02:50,960 --> 00:02:56,120 So it's like the AI has to make smart guesses using what it can see and what it 60 00:02:56,120 --> 00:02:57,920 knows about the city's rules. 61 00:02:57,920 --> 00:02:58,320 Yeah. 62 00:02:58,320 --> 00:03:01,920 So we've got our AI learning in this city taking on these challenges. 63 00:03:01,920 --> 00:03:05,520 What were the results? Did the neuro symbolic AI actually do well? 64 00:03:05,520 --> 00:03:07,040 Well, that's what we'll get into next. 65 00:03:07,040 --> 00:03:11,360 The findings are pretty interesting, especially when you compare Nessie AI to 66 00:03:11,360 --> 00:03:12,320 other approaches. 67 00:03:12,320 --> 00:03:14,320 Ooh, exciting. Let's get into that. 68 00:03:14,320 --> 00:03:20,280 So the results showed that in both SBF and VAP, these neuro symbolic AI 69 00:03:20,280 --> 00:03:23,880 approaches, they did better than just deep learning, especially when things are 70 00:03:23,880 --> 00:03:24,920 more complex. 71 00:03:24,920 --> 00:03:29,560 So like mixing logic and learning really gives AI an advantage, huh? 72 00:03:29,560 --> 00:03:32,200 But I bet there's still some challenges, right? I mean, these are pretty complicated 73 00:03:32,200 --> 00:03:32,520 tasks. 74 00:03:32,520 --> 00:03:36,600 Oh, for sure. One of the biggest hurdles was something called compositional 75 00:03:36,600 --> 00:03:41,480 generalization. It's basically that the AI struggled when it had to adapt to 76 00:03:41,480 --> 00:03:46,200 totally new combinations of agents, even if it already knew the basic rules. 77 00:03:46,200 --> 00:03:50,120 So like it could understand right of way, but got confused if like a new kind of 78 00:03:50,120 --> 00:03:51,080 vehicle showed up. 79 00:03:51,080 --> 00:03:55,560 Exactly. It shows how hard it is for AI to take what it knows and use it in 80 00:03:55,560 --> 00:03:56,920 totally new situations. 81 00:03:57,560 --> 00:04:00,840 Kind of like if you learned chess, but then had to play with pieces that moved 82 00:04:00,840 --> 00:04:01,480 differently. 83 00:04:01,480 --> 00:04:02,040 Oh, yeah. 84 00:04:02,040 --> 00:04:05,960 The basics are the same, but using them in a new context that takes a whole 85 00:04:05,960 --> 00:04:07,800 other level of understanding. 86 00:04:07,800 --> 00:04:12,280 That's a great way to put it. It makes me wonder, how did the AI do with the visual 87 00:04:12,280 --> 00:04:14,840 noise in that VAP task? 88 00:04:14,840 --> 00:04:19,320 I feel like that real world messiness would really mess things up, even if it 89 00:04:19,320 --> 00:04:20,360 gets the rules. 90 00:04:20,360 --> 00:04:23,160 Yeah, you're right. That was another big challenge. Telling the difference between 91 00:04:23,160 --> 00:04:28,360 similar looking things, like a police car versus a regular car that was tough for 92 00:04:28,360 --> 00:04:33,320 the AI when the images were blurry or hidden. It really struggled to use what 93 00:04:33,320 --> 00:04:35,720 it knew when the visual data was messy. 94 00:04:35,720 --> 00:04:40,200 Makes sense. So it's like it aced the written test, but the practical exam was 95 00:04:40,200 --> 00:04:42,840 hard when things got visually complicated. 96 00:04:42,840 --> 00:04:45,960 What about that comparison to large language models in humans? 97 00:04:46,680 --> 00:04:50,280 The paper talked about testing them on a simpler version of the VAP task. 98 00:04:50,280 --> 00:04:54,360 Right. They wanted to see how different AI approaches could do logical reasoning 99 00:04:54,360 --> 00:04:57,080 without the extra difficulty of visual input. 100 00:04:57,080 --> 00:05:01,240 So they took away the visual noise and just gave the AI and the humans text 101 00:05:01,240 --> 00:05:04,920 descriptions of what the city looked like, and then asked them to predict what a 102 00:05:04,920 --> 00:05:06,520 specific thing would do next. 103 00:05:06,520 --> 00:05:09,400 Ah, so it's like giving them a written driving test to see if they get the 104 00:05:09,400 --> 00:05:12,760 theory before putting them behind the wheel what they find. 105 00:05:12,760 --> 00:05:17,800 So the large language models, even the advanced ones like GPT-4, they showed some promise, 106 00:05:18,520 --> 00:05:23,640 but they still had a harder time with the tougher logical puzzles compared to people. 107 00:05:23,640 --> 00:05:27,560 They often relied on common sense or were super cautious in their predictions. 108 00:05:27,560 --> 00:05:30,840 They didn't quite get all the little details of Ligocity's rules. 109 00:05:30,840 --> 00:05:36,200 It's interesting that even AI that's so good at language and information can trip up 110 00:05:36,200 --> 00:05:41,000 when faced with complex reasoning. What about the humans? How they do on this logic test? 111 00:05:41,000 --> 00:05:44,200 Humans did a lot better overall. Not surprising, right? 112 00:05:44,200 --> 00:05:47,960 But even they made mistakes, especially when the rules got super complicated 113 00:05:47,960 --> 00:05:50,760 and you needed a bunch of steps to figure things out. 114 00:05:50,760 --> 00:05:55,000 It just shows that logical induction, even for us, can be pretty tough. 115 00:05:55,000 --> 00:05:58,760 Yeah. It makes you realize how much we do without even thinking about it when we're 116 00:05:58,760 --> 00:06:02,840 navigating everyday stuff. Something like understand the rules of the road is 117 00:06:02,840 --> 00:06:04,360 actually a pretty amazing mental thing. 118 00:06:04,360 --> 00:06:07,400 It is. And this opens up some cool research avenues. 119 00:06:07,400 --> 00:06:12,040 Imagine combining those large language models, which are awesome at processing info and coming 120 00:06:12,040 --> 00:06:18,120 up with ideas with a formal logic engine that can double check and refine those ideas. 121 00:06:18,120 --> 00:06:23,320 It'd be like having a super smart brainstorming partner who also happens to be a logic genius. 122 00:06:23,320 --> 00:06:27,240 Now that's a powerful combo. It's like those AI systems that are chess champs. 123 00:06:27,240 --> 00:06:30,440 But instead of a game, they're mastering real-world logic. 124 00:06:30,440 --> 00:06:35,720 But before we go too far down that road, I'm curious, what are your big takeaways from this 125 00:06:35,720 --> 00:06:39,800 whole deep dive into logicity? What are the key points we should remember? 126 00:06:39,800 --> 00:06:45,160 So what does it all mean? Where does logicity fit into the bigger picture of AI research? 127 00:06:45,160 --> 00:06:50,120 Well, I think logicity is a really useful tool for researchers who are studying this 128 00:06:50,120 --> 00:06:55,240 neurosymbolic AI. It gives them this rich environment that they can change up and use 129 00:06:55,240 --> 00:06:59,080 to test how well AI can learn and use complex rules. 130 00:06:59,080 --> 00:07:02,360 And it's way closer to the real world than a lot of other testing methods. 131 00:07:02,360 --> 00:07:05,320 So it sounds like this could be useful for way more than just self-driving cars. 132 00:07:05,320 --> 00:07:11,560 Exactly. The potential uses are huge. Like, think about using this type of AI to make traffic flow 133 00:07:11,560 --> 00:07:17,880 better in cities. Or design smarter transportation systems. Or even build robots that are safer 134 00:07:17,880 --> 00:07:20,040 and more adaptable so they can work better with us. 135 00:07:20,920 --> 00:07:26,200 What about AI that can understand those super complicated legal contract? 136 00:07:26,200 --> 00:07:26,760 Oh, yeah. 137 00:07:26,760 --> 00:07:31,480 Or medical guidelines. There are so many areas where understanding and applying rules is super 138 00:07:31,480 --> 00:07:36,840 important. Absolutely. If we can teach AI to understand something as crazy complicated as a city, 139 00:07:37,400 --> 00:07:43,560 then using that same tech for other complex systems is like, it has enormous potential. 140 00:07:43,560 --> 00:07:48,280 It's like we're giving AI a crash course in how human society works. That could totally 141 00:07:48,280 --> 00:07:50,920 change how we design and interact with AI in so many ways. 142 00:07:50,920 --> 00:07:55,800 Yeah, for sure. And that brings us to a really important point. This research isn't just about 143 00:07:55,800 --> 00:08:00,600 building smarter AI. It's about building AI that gets the systems we've already created and can 144 00:08:00,600 --> 00:08:07,400 work within them. From traffic laws to social norms to legal stuff to economic models, 145 00:08:07,400 --> 00:08:09,720 these systems control so much of our lives. 146 00:08:09,720 --> 00:08:14,440 So it's not just about intelligence, but about AI that can actually function in our world, 147 00:08:14,440 --> 00:08:15,480 working with people. 148 00:08:15,480 --> 00:08:21,880 Exactly. And that leads us to a question for you to think about. If we can teach AI the logic of 149 00:08:21,880 --> 00:08:27,800 a whole city, what other complex systems could we use this for? What are the good things and the 150 00:08:27,800 --> 00:08:29,240 bad things that could happen? 151 00:08:29,240 --> 00:08:33,800 Those are some really interesting questions to think about. I'm already imagining how this AI 152 00:08:33,800 --> 00:08:38,920 could change healthcare, education, and even how we manage the environment. There's so many 153 00:08:38,920 --> 00:08:39,880 possibilities. 154 00:08:39,880 --> 00:08:45,720 It is really exciting. But we do need to remember that logicity is still like a simplified version 155 00:08:45,720 --> 00:08:51,640 of the real world. It's a big step though. It gives us a powerful tool to explore what AI 156 00:08:51,640 --> 00:08:57,480 can do and to figure out what roadblocks we need to clear to build AI systems that are 157 00:08:57,480 --> 00:08:59,160 really reliable and adaptable. 158 00:08:59,160 --> 00:09:02,760 It's pretty mind blowing to think about all the ways this research could change things. But for 159 00:09:02,760 --> 00:09:08,200 now, we want to hear from you. What areas do you think this type of AI could have the biggest impact 160 00:09:08,200 --> 00:09:08,440 on? 161 00:09:08,440 --> 00:09:10,840 Yeah. Let us know down in the comments. 162 00:09:10,840 --> 00:09:14,760 We're really interested to hear your thoughts on this groundbreaking research. After all, 163 00:09:14,760 --> 00:09:18,600 figuring out the future of AI is a conversation we all need to be part of. 164 00:09:18,600 --> 00:09:23,240 Well, that wraps up our deep dive into logicity. Thanks for joining us on this exploration of 165 00:09:23,240 --> 00:09:49,240 Neurosymbolic AI. Until next time, keep those AI brains buzzing.