Deep Learning Advances: DeepMind's Robot AI Revolutionizes Learning

0:00What can 160 years of experience teach you about the future? When it comes to protecting what matters, Pacific Life provides life insurance, retirement income, and employee benefits for people and businesses building a more confident tomorrow. Strategies rooted in strength and backed by experience. Ask a financial professional how Pacific Life can help you today. Pacific Life Insurance Company, Omaha, Nebraska, and in New York. Pacific Life and Annuity, Phoenix, Arizona. Today on the podcast, I want to cover three really interesting stories that are happening in AI. One is with Cisco, one is with DeepMind, and one is with a slightly smaller AI company.

0:39So without further ado, let's dive into it. So the news out of Cisco is the fact that on Tuesday, they launched a networking chip for AI supercomputers that they say is going to compete with a bunch of offerings from Broadcom and Marvel Technologies. So why is this important in the industry? Well, first off, essentially with the rise of the popularity in AI right now, you know, chat GPT and everything else, it's made the need for GPUs or graphics processing units is what they're actually called. It's made these really important because these GPUs are what is used to train AI models. Now, one GPU is, you know, has its limitations.

1:17NVIDIA is making a lot of money building these GPUs. But how you actually train these AI models is by chaining multiple of these GPUs together to get a lot of processing power to train these really powerful models. How many GPUs, you might ask? Well, there are thousands and thousands. So Cisco is coming onto the scene right now by introducing what they are calling the G200. And essentially, they say it is the most powerful networking chip in the market, which is going to fuel AI and machine learning workloads to enable the most power efficient networks. So what this is actually doing is this new networking chip from Cisco is actually going allow you to link together 32 ,000 GPUs.

1:59This is twice as many as previous networking chips were able to do. And essentially what this is doing is it's decreasing the latency and it's decreasing a lot of the time that it's taking for these GPUs to communicate with each other. This is actually going to really allow these AI model workloads to be super power efficient and to be super fast. So Cisco said that these chips are actually going to help in carrying out AI and machine learning tasks with about 40 % fewer switches and a lesser leg, and that is going to make it a lot more efficient. So I think this is interesting because they've also noted that chips from their Silicon One series, and that is what this is under, are being tested by five of the six major cloud providers.

2:45So they didn't actually name names, but obviously five of the six, we know that these are going to be someone from Amazon Web Services, Microsoft Azure, Google Cloud, and there's a handful of other ones. But these three ones inevitably are going to be integrated in there. And this makes perfect sense, right? These guys that are running these massive data centers, these massive cloud providers are going to want the fastest networking switches. and this seems to be a really big play out of Cisco, which is going to be enabling faster AI model training and making this more energy efficient, which I think is really great overall.

3:19The second story I wanted to cover on today's podcast is the fact that DeepMind has developed a new AI for helping robotic arms train. And this is actually super interesting how they've done this and what I think the implications are. But essentially, I guess the first thing to say is obviously AI that trains these robot arms is nothing new, right? This is what a lot of factories have been working on for quite a while. Like when people say, oh, AI is not new. It's been around for, you know, 10 years actively integrating into things. This is one of those areas that has been pretty popular. But what they've actually done that is unique is that they've said, we demonstrate that a single large model can solve a diverse set of tasks on multiple real robotic embodiments and can quickly adapt to new tasks and embodiments.

4:08So essentially what they're saying is they had an AI model watch someone with like a controller, right? You can imagine someone with like a joystick controlling a robotic arm to go pick up an object from a bucket, essentially. so they had the the human do this a thousand times they went and picked the object out of the bucket a thousand times and they let the ai train off of that and then practice and they were able to refine that training telling it yes you picked the object up correctly or no you didn't eventually the robot could you know copy what they were doing and could pick the objects up now the impressive part of this is that they were then able to use the same trained model and have the AI work with new claws or new robotic arms.

4:56So, you know, originally they had one, or actually I believe they had four different robotic arm models that it was trained on, and they had two-pronged arms. They were then able to switch and have the model adapt to a more complex arm with three fingers and also had twice as many controllable inputs. So essentially, the AI was able to learn how to manipulate any robotic arm to go pick up the object based off of what it learned from the first which is really impressive um and i you can i'm like essentially going to extrapolate and assume that the ai is learning how to learn it's learning how to use different um arms and if you wanted to think of like a really dystopian like terrifying aspect you could say like okay well what if you release an ai like this that understands how to like use different machinery like what if it moves beyond robotic arms and all of a sudden it moves into like i don't know autonomous driving or all sorts of other tasks where essentially it's trained to do one thing and because of that it can kind of figure out how to learn how to do another thing similar to how a human mind would work so imagine right this is the um this is like the terminator example is all of a sudden the robot it got trained in a factory to pick up something then it gets mad and it learns how to take over all the machinery and all the equipment and operate all of it to do whatever its nefarious task of ending the world is, right?

6:15Okay, dystopian. But you can see where this goes. This is a really powerful technology where essentially it's being trained to do one thing and then it's being able to adapt to a whole bunch of different tooling, a bunch of different, you know, a factory might have one arm and it can learn how to do the same task with a bunch of different types of arms. So really powerful technology and I see a lot of implications. The third story I wanted to talk about on the podcast today is actually just a feature that I think is absolutely amazing out of by Otter AI. So Otter AI, it's essentially just kind of like a, it's like a meeting oriented AI, where essentially, like, if you have a meeting, it's going to record the whole thing.

6:54And then it has it knows like who's talking, and it pretty much generates a transcript of your entire meeting. So the idea being, you can go back and look at what was said. And the AI part is that it's listening to your voice, and it's transcribing it and then it's also knowing who's talking it saves all the time stamps it's really useful for meetings essentially it just records your meeting and transcribes it and you can go get data but what they've just introduced that is super cool in my opinion is an AI powered chatbot that integrates with those meetings so imagine you're 10 minutes late to a meeting instead of having to you know do the question of shame and say hey sorry guys I just got here what did I miss you just go to the chat bot on the side of this AI meeting transcriber and say, you know, I just got here right now.

7:40What did I miss from the meeting? Can you give me like a breakdown? It gives you a breakdown. Or imagine if the meeting's over, you didn't go to that meeting because you had a conflicting, you know, conflicting meeting, you go back to the link to the meeting that happened. And you can ask any questions. Was there any, you know, deliverable deliverables requested from myself? Was I ever mentioned in the meeting? Is there anything that mentions my department or what the projects I'm working on. You can ask it all sorts of questions and it can give you information about that meeting without you actually having to go and talk to another person or bother another person to kind of get a recap on what's going on with the meeting.

8:15I think this could save people a lot of time. I'm a really big fan of a lot of these tools that essentially can spare you a meeting if it's not necessary. Of course, meetings aren't bad, but I really like tools like Loom where you may not actually need a full meeting. Maybe you just need to record yourself explaining how to do a process for like three minutes and send it to a co-worker so that they can do that. You know, if it was an actual meeting, you might have sat down for like 30 minutes and talked about it and talked about a bunch of different stuff. But instead, you just recorded a quick like three to five minute screen recording, sent it over, and it saves both of you a lot of time.

8:50So I think that a lot of these technologies doing that are going to be powerful. And I think this is a really powerful use case of chat bots and AI in general within the corporate and kind of enterprise space. So this is an area I'm excited to see more features and tools like this that are going to make our lives a lot simpler. I'm going to cut down on, you know, things that we had to do in the past that now we no longer have to do and hopefully make us all more productive and make our companies more profitable and successful.