สำเนาบทสนทนา 🔗
WEBVTT
00:00:24.719 --> 00:00:27.760
Generative AI, is it cool or what?
00:00:28.879 --> 00:00:31.519
Technologies are cool in my mind.
00:00:31.760 --> 00:00:44.799
The problem with technologies, with new technologies recently is that we just got used to the fact that we have a new AI model, new AI capability almost every week.
00:00:45.039 --> 00:00:47.920
So we sometimes miss the big picture.
00:00:48.159 --> 00:00:57.759
The big picture is that new technologies are not always just faster, better computers.
00:00:57.920 --> 00:01:01.039
They really reshape everything that we're doing.
00:01:01.600 --> 00:01:05.840
So let me take you to the first industrial revolution.
00:01:06.239 --> 00:01:09.760
So we managed to use steam.
00:01:10.239 --> 00:01:10.719
So what?
00:01:10.879 --> 00:01:12.640
It's not just a better machine.
00:01:12.799 --> 00:01:13.040
Okay?
00:01:13.200 --> 00:01:25.280
The steam engine has really made the world smaller with steamboats, steam trains that could transfer people and goods all around the world.
00:01:25.439 --> 00:01:28.959
Now the next generation brought us electricity.
00:01:29.359 --> 00:01:34.560
So now we can also work during night because there is light.
00:01:34.879 --> 00:01:38.959
We can work during winter because there's heating.
00:01:39.120 --> 00:01:45.439
So it's a huge change to our lifestyle, but it even changed our biology.
00:01:45.599 --> 00:01:46.400
Think about that.
00:01:46.560 --> 00:01:48.319
Are you taller than your parents?
00:01:49.120 --> 00:01:51.439
Are they taller than their parents?
00:01:52.079 --> 00:01:54.159
Is the world getting taller?
00:01:54.719 --> 00:01:55.840
What do you think?
00:01:57.040 --> 00:02:08.080
Yes, I mean you don't think about that when you invent electricity, but then you have refrigerators, then you have better food, better health care, and look at that.
00:02:08.319 --> 00:02:10.879
This is the average height of men in Germany.
00:02:11.120 --> 00:02:14.080
They gained like 20 centimeters.
00:02:14.240 --> 00:02:15.280
It's a lot, right?
00:02:15.360 --> 00:02:16.479
It's think about that.
00:02:16.639 --> 00:02:21.199
Too bad they kept me so short, but it's it's pretty amazing.
00:02:21.360 --> 00:02:23.120
And then came computers.
00:02:23.280 --> 00:02:26.319
I don't have to tell you how much this changed everything.
00:02:26.479 --> 00:02:30.560
The way we behave, the way we communicate with each other, everything.
00:02:30.719 --> 00:02:36.479
And the next revolutions are gonna make bigger changes.
00:02:36.560 --> 00:02:39.759
Okay, this is not a linear growth.
00:02:40.080 --> 00:02:45.039
This is um an exponential, exponential revolution.
00:02:45.199 --> 00:02:54.719
Okay, and I think what enabled AI eventually is the fact that we taught the machine to speak our language.
00:02:54.879 --> 00:03:02.159
Okay, you and I have uh language barriers, but computers can understand everything from us, right?
00:03:02.240 --> 00:03:21.919
The minute we were able to represent our language using tokenization, put it in embedding space, attention, and so on, and train large language models, we basically opened the door for computers to all of our knowledge.
00:03:22.080 --> 00:03:34.560
I mean, computers can generate images and videos and audio, it's it's really nice, but text is more important because after all, most of our knowledge is text.
00:03:34.879 --> 00:03:40.639
And then we taught the machines, we taught computers to generate text, right?
00:03:40.800 --> 00:03:46.400
So we can interact with them, we can chat with them, and now we can measure their intelligence.
00:03:46.639 --> 00:03:48.560
Are they smarter than us?
00:03:48.960 --> 00:03:50.159
What do you think?
00:03:50.479 --> 00:03:52.879
Is computer already smarter than us?
00:03:53.199 --> 00:03:58.000
So the average IQ of a person is about 100.
00:03:58.400 --> 00:04:07.919
I want to believe that most of the people in this room are on the right side of this curve, but so as most of the latest models.
00:04:08.080 --> 00:04:12.159
Okay, so yes, AI is already smarter than us.
00:04:12.240 --> 00:04:20.079
We can argue about the details, how do you measure IQ and stuff like that, but in any possible test, they are better.
00:04:20.240 --> 00:04:24.240
In IQ, in EQ, they're even more funny than us.
00:04:24.399 --> 00:04:24.720
Okay?
00:04:24.959 --> 00:04:26.319
That's really insulting.
00:04:26.560 --> 00:04:31.600
But for me, the shocking fact is how fast this has all happened, right?
00:04:31.759 --> 00:04:42.000
It took AI one year, one year, to move from an average person with an IQ of 100 to a genius with an IQ of 140.
00:04:42.079 --> 00:04:43.759
Okay, that's pretty fast.
00:04:43.920 --> 00:04:45.439
Now, what is next?
00:04:45.600 --> 00:04:54.000
I mean, it when we're moving so fast, it's very difficult to understand where technology is heading, what comes next.
00:04:54.319 --> 00:05:01.920
So if we zoom out a little bit, I think we can see the pattern, okay?
00:05:02.000 --> 00:05:09.600
And there is a pattern because technology evolved in, in my mind, in predictable directions.
00:05:09.680 --> 00:05:15.040
Okay, so let's take this device, sickle, right?
00:05:15.199 --> 00:05:24.720
One blade, and I believe that when people stopped uh picking up crop with their their hands and start using blades, it was a huge technology advancement, right?
00:05:24.959 --> 00:05:30.240
But then somebody said, okay, using the same stroke, I can use two blades, right?
00:05:30.319 --> 00:05:34.879
So multiple blades will be better, will be more efficient, so why not?
00:05:35.040 --> 00:05:40.240
Now, if this machine can be mobile, I can cover more ground.
00:05:40.399 --> 00:05:45.279
If this machine is autonomous, then I'm free up to do other stuff.
00:05:45.519 --> 00:05:53.680
And if I have one machine like that that is autonomous, why not having many of them working like a swarm, right?
00:05:53.759 --> 00:05:58.560
And if one
Load full transcript
of them is getting you know broken, they can cover for me for each other.
00:05:58.879 --> 00:05:59.759
This is nice.
00:05:59.920 --> 00:06:06.399
So this is technology trends, and you can see them in many domains, not only in agriculture.
00:06:06.560 --> 00:06:10.160
You know, here you can see fishing, you can see it in many places.
00:06:10.399 --> 00:06:16.160
Recently, unfortunately, we also see that in military technologies, right?
00:06:16.240 --> 00:06:29.839
You have one rocket, multiple missiles, or one missile with multiple heads, and then mobile launchers, and then guided missiles, you know, guided by GPS, by vision, and then a swarm.
00:06:30.079 --> 00:06:42.639
In all of these technologies, in all of these examples, parallelism means efficiency, mobility means opportunities to cover more.
00:06:43.279 --> 00:06:46.560
If it's autonomous, I'm free up to do other stuff.
00:06:46.720 --> 00:06:58.319
And if it's distributed like a swarm, it opens up new opportunities and new advantages like redundancy, scalability, and more.
00:06:58.720 --> 00:07:00.800
And AI is not different.
00:07:01.360 --> 00:07:08.000
Last year's we've tried to improve AI by adding more and more layers to the models.
00:07:08.160 --> 00:07:14.480
Okay, so more parallelism, more parallelism to our GPUs, and this is how we made it.
00:07:14.639 --> 00:07:17.360
Then we tried to make it mobile, right?
00:07:17.439 --> 00:07:22.000
We want to take these models and put them in more devices.
00:07:22.079 --> 00:07:23.920
Okay, so how do you do that?
00:07:24.240 --> 00:07:32.800
You invent all kinds of protocols like MCP, so it would be easier to put them in many places and communicate with them.
00:07:32.959 --> 00:07:47.120
You try to come up with smaller language models, you're coming up with better, more efficient accelerators like NPUs, the neural processing unit, or you come up with better algorithms like speculative decoding.
00:07:47.279 --> 00:07:51.120
Feel free to take a look at my video explaining exactly that.
00:07:51.759 --> 00:07:53.279
What is the next step?
00:07:53.519 --> 00:07:54.720
Autonomous.
00:07:55.279 --> 00:07:55.759
Right?
00:07:55.920 --> 00:07:58.000
We all heard about AI agent.
00:07:58.160 --> 00:07:59.519
What is an AI agent?
00:07:59.600 --> 00:08:04.879
It's basically an AI-powered system or application, right?
00:08:05.040 --> 00:08:06.079
That is autonomous.
00:08:06.240 --> 00:08:13.439
It can consume data and make decisions for us without the constant need for input, right?
00:08:13.680 --> 00:08:14.959
So it's autonomous.
00:08:15.120 --> 00:08:17.360
And then the question is what comes next?
00:08:17.680 --> 00:08:18.800
What comes next?
00:08:19.439 --> 00:08:20.399
The edge.
00:08:20.800 --> 00:08:26.240
Distributed, autonomous, AI systems everywhere.
00:08:26.560 --> 00:08:33.279
This is just logical, you know, trend that will bring us exactly there.
00:08:33.440 --> 00:08:45.759
We want to take the intelligence, we want
to take the wisdom of AI and distribute it and put it closer to human, closer to the edge, closer to where the data is.
00:08:46.000 --> 00:08:48.240
This is exactly what we're trying to do.
00:08:48.559 --> 00:08:52.000
And you're gonna see that anywhere.
00:08:52.159 --> 00:09:00.159
So an LLM, a large language model running in the cloud is great, but I want it to run on my laptop.
00:09:00.320 --> 00:09:00.559
Okay?
00:09:00.960 --> 00:09:05.440
I don't need generic answers and advices.
00:09:05.600 --> 00:09:29.120
I want something, I want advices and answers customized to my style, to my you know intelligence, to my needs, and only a system that is familiar with with what I'm doing, okay, that have access to all my communication, documentation, mail, etc., can do that.
00:09:30.000 --> 00:09:34.240
We want to bring AI closer to the data.
00:09:34.799 --> 00:09:36.879
The foundation of AI is data.
00:09:37.039 --> 00:09:38.960
Where is that data being generated?
00:09:39.120 --> 00:09:40.559
Not in the cloud, right?
00:09:40.639 --> 00:09:43.600
The data is being generated in a factory.
00:09:43.840 --> 00:09:52.320
So a robot in a factory can detect right away that this part has a defect and it has to be discarded.
00:09:52.480 --> 00:09:55.840
It doesn't need to go to the cloud for that, right?
00:09:56.000 --> 00:10:00.879
You need to bring the wisdom to the edge, to where things are happening.
00:10:01.120 --> 00:10:08.960
And if you have a smart city camera, it doesn't need to stream all the video to a command center.
00:10:09.120 --> 00:10:14.559
It can detect by itself that has accidents that just occurred, right?
00:10:14.960 --> 00:10:26.480
It can alert the police, the hospital, it can, you know, collaborate with other intersections to create a green wave for for the ambulance, and so on.
00:10:26.639 --> 00:10:30.480
And I haven't even started to talk about humanoids, right?
00:10:30.639 --> 00:10:39.840
They're gonna be distributed, they're gonna be mobile, they're gonna be autonomous.
00:10:40.080 --> 00:10:41.279
Exactly this.
00:10:42.320 --> 00:10:45.440
The question is, why is this not happening?
00:10:45.519 --> 00:10:55.919
Okay, we have millions or billions of devices waiting to be to get smarter, and we got lots of AI models want to be deployed.
00:10:56.080 --> 00:10:59.039
Okay, so what is preventing us from getting there?
00:10:59.200 --> 00:11:03.039
There are many problems, but just to mention a few.
00:11:03.200 --> 00:11:05.200
The first problem is hardware.
00:11:05.279 --> 00:11:10.159
Okay, edge devices have unique hardware requirements.
00:11:10.240 --> 00:11:17.840
Okay, when you're working in the cloud, you theoretically can have endless compute power, endless memory space, right?
00:11:17.919 --> 00:11:20.879
And this is not the case in the real world.
00:11:21.200 --> 00:11:26.159
Edge devices has limited amount of power.
00:11:26.320 --> 00:11:31.360
They can dissipate limited amount of compute power, limited memory, and so on.
00:11:31.519 --> 00:11:39.600
If you want to bring AI capabilities to the edge, you want to have specialized machines, specialized hardware for that.
00:11:39.840 --> 00:11:42.559
So, just like our brain, think about our brain.
00:11:42.720 --> 00:11:47.919
Our brain is an excellent edge device, uh, example for an edge device, right?
00:11:48.000 --> 00:11:51.919
It can dissipate around 20 watts, so low power.
00:11:52.159 --> 00:12:09.200
It has unique regions to decipher, to process vision, unique special regions to process audio, special specialized regions to take decisions, and so on.
00:12:09.360 --> 00:12:09.679
Okay?
00:12:09.919 --> 00:12:15.120
Now look at the latest um microprocessor from Intel.
00:12:15.440 --> 00:12:20.000
This is codenamed the Ultra Core III, codenamed Panther Lake.
00:12:20.159 --> 00:12:25.120
It has a CPU, it has a GPU, it has an NPU, lots of acceleration.
00:12:25.200 --> 00:12:31.519
It's a highly heterogeneous system, and we need all of that for the future edge devices, right?
00:12:31.759 --> 00:12:44.320
Think about robots that need to process multiple cameras to construct a 3D view of the world around him and navigate through it just like we humans do.
00:12:44.559 --> 00:12:51.120
So you need specialized hardware, and you need here, you got here uh CPUs with multiple cores, right?
00:12:51.279 --> 00:12:53.039
Efficient core, performance core.
00:12:53.120 --> 00:12:57.840
You got here image processing and media processing cap acceleration inside.
00:12:58.000 --> 00:13:02.159
You got here uh very large GPU and an NPU.
00:13:02.320 --> 00:13:10.799
The problem is, and this brings me to the problem number two, that it introduces software problems, okay?
00:13:10.960 --> 00:13:16.559
I mean, we all know that programming a CPU, a GPU, or an NPU is totally different.
00:13:16.799 --> 00:13:17.840
Which makes sense, right?
00:13:17.919 --> 00:13:24.240
It's a different architecture, different instruction set, different memory mapping, and so on.
00:13:24.399 --> 00:13:27.360
Okay, so this is problem number two.
00:13:27.519 --> 00:13:32.240
And for this, we have introduced what we call open-edge platform.
00:13:32.320 --> 00:13:43.759
It's an open source, free-to-use bundle of software that basically have everything you need in order to get AI into edge devices.
00:13:43.919 --> 00:14:04.320
So it starts with the lower level manageability and operating system and everything, but it goes to what we call edge AI libraries, like in with tools like OpenVino, if you want to use um highly efficient uh inference on CPU, GPU, NPU.
00:14:04.480 --> 00:14:09.759
By the way, it's open source, it runs also on ARM, on Raspberry Pi, whatever you want.
00:14:09.919 --> 00:14:14.399
If you want to process full video pipelines, we have the deep learning streamer.
00:14:14.559 --> 00:14:25.279
If you want to take multiple cameras into constructing your 3D world, okay, and understand how it operates,
use Intel Scenescape.
00:14:25.440 --> 00:14:26.960
We have a nice demo over there.
00:14:27.120 --> 00:14:30.000
Just go to the Dell booth and take a look.
00:14:30.240 --> 00:14:35.919
How we take multiple cameras and just track real objects in the world and Getty.
00:14:36.480 --> 00:14:48.080
Getty is for in my mind is a great tool that can help people train models, custom models, without the need to be an expert, right?
00:14:48.159 --> 00:14:49.039
It's all GUI.
00:14:49.600 --> 00:14:59.600
You get in your data, annotate it really quickly, and you get a model which you can use in the cloud or on any edge device.
00:14:59.679 --> 00:15:00.480
So try it out.
00:15:00.559 --> 00:15:02.480
There's many other tools, of course.
00:15:02.720 --> 00:15:15.679
And there is an AJI suite on top with ready-to-use solutions for multiple use cases for retail manufacturing, for smart city, for robotics.
00:15:15.840 --> 00:15:27.360
It's just if you look there, I'm pretty sure you're gonna find uh a reference design, an example, a sample code you can use right away, copy, learn how to use it.
00:15:27.679 --> 00:15:30.559
So this is the open edge platform.
00:15:30.799 --> 00:15:54.480
I think the hardware solutions with the right software solutions will bring us to the right place we're looking for, but AI still has some mileage to go, and I just want to talk in two, three more minutes about how do I see the next trends in AI, what's going next.
00:15:54.639 --> 00:16:07.200
So in the near future, we're gonna see models getting more nimble, lighter, getting more cognitive, more autonomous, and more personal.
00:16:07.360 --> 00:16:09.440
And I'll try to elaborate on that.
00:16:09.600 --> 00:16:14.559
So right now we see how models are progressing.
00:16:14.720 --> 00:16:18.879
They are getting smarter and smarter, but they're also getting bigger and bigger.
00:16:18.960 --> 00:16:20.559
They're getting getting too big.
00:16:20.720 --> 00:16:25.519
Okay, and if you look inside them, right, some of them are really sparse.
00:16:25.679 --> 00:16:30.639
I don't want to say they're empty, but they're really not very efficient.
00:16:30.799 --> 00:16:39.759
Okay, and you know, we can apply all kinds of um algorithms on large models, quantization.
00:16:39.919 --> 00:16:45.679
We're quantizing now to 4-bit, and we hardly see any accuracy degradation.
00:16:45.919 --> 00:16:50.159
We're running pruning, distillation, and so on.
00:16:50.480 --> 00:16:56.720
So the potential is there, and we're gonna see more and more smaller models.
00:16:56.879 --> 00:17:00.320
We're gonna see much more targeted models, okay?
00:17:00.559 --> 00:17:06.880
Right now, what you're doing is you're getting a model that knows everything about everything, right?
00:17:07.119 --> 00:17:10.799
But I'm only wanna ask about a medical question, right?
00:17:11.119 --> 00:17:18.160
So when you want to read about, I don't know, dinosaur or something, you don't read the whole encyclopedia,
right?
00:17:18.319 --> 00:17:28.319
So we're gonna see more and more targeted, focused models for doctors, for lawyers, I don't know, for engineers.
00:17:28.480 --> 00:17:35.519
They're gonna be smaller, they're gonna be smarter, okay, and they're gonna be focused to exactly what I need.
00:17:35.680 --> 00:17:38.400
We're gonna see more retriever-based, okay?
00:17:38.559 --> 00:17:53.599
Why hold all the information inside the model when I can just go and look it up using you know techniques like RAG in the most updated real-time resources, right?
00:17:53.759 --> 00:18:00.400
You don't have to hold everything in in your in your belly, in your mind, in your brain, in the model.
00:18:00.559 --> 00:18:07.920
So we're gonna see much many more retrieval-based, and we're gonna see much more on fine-tuning.
00:18:08.079 --> 00:18:11.599
Today, when you get a model, this is a model, okay?
00:18:11.920 --> 00:18:18.400
But I'm pretty sure in the near future we're gonna see lots of models getting improved over time.
00:18:18.559 --> 00:18:21.680
It doesn't have to be fast, you don't need a big machine for that.
00:18:21.839 --> 00:18:23.279
It can take 10 hours.
00:18:23.440 --> 00:18:24.160
I don't care.
00:18:24.400 --> 00:18:26.880
Overnight, just fine-tune the model.
00:18:27.039 --> 00:18:39.440
And then my smart surveillance camera knows that this strange person who stopped, you know, who walked by every every day at 10 o'clock is the mailman, right?
00:18:39.519 --> 00:18:43.519
It's not somebody that I should suspect, right?
00:18:43.759 --> 00:18:50.960
Or think about two cars, two Tesla cars that just left the factory in Texas.
00:18:51.119 --> 00:18:55.920
Okay, one of them was shipped to India and the other was shipped to Australia.
00:18:56.079 --> 00:19:02.400
Okay, and these two cars are driving every day, taking millions of photos, right?
00:19:02.480 --> 00:19:08.720
And at night they plug it in, upload everything, retrain, get a new model.
00:19:08.880 --> 00:19:19.839
So after a month, the car in India will know how to avoid cows on the road, and the car in Australia will be careful of kangaroos, for example.
00:19:20.000 --> 00:19:30.400
Okay, but the models will be local, will be you know, ongoing, um improving themselves as time goes by.
00:19:30.559 --> 00:19:33.279
Okay, models are gonna be more cognitive.
00:19:33.359 --> 00:19:48.720
We see this already happened with models combining text and vision and others, they will be more adaptive, more aware to the context and to the situation, and we already see models that can explain themselves.
00:19:49.519 --> 00:19:52.240
And models are gonna be much more personal, okay?
00:19:52.400 --> 00:19:59.119
Think about there are already um services like that, paid services like that, but think about that.
00:19:59.200 --> 00:20:03.759
You have a student who wants to solve an equation.
00:20:03.920 --> 00:20:08.640
I don't need the my model to just tell him what the answer is, right?
00:20:08.720 --> 00:20:11.440
I need the model to take him hand by hand.
00:20:11.519 --> 00:20:15.519
So I want the model to ask me, do you know what is the equation?
00:20:15.839 --> 00:20:16.799
No, let me tell you.
00:20:16.880 --> 00:20:18.559
Do you know what is a differential equation?
00:20:18.720 --> 00:20:19.359
Let me show you.
00:20:19.440 --> 00:20:25.279
So just take me the path to understanding so I can really solve it the next time, right?
00:20:25.359 --> 00:20:26.640
I don't need the answer.
00:20:26.799 --> 00:20:31.920
And for this, you gotta have a system that actually knows me, right?
00:20:32.079 --> 00:20:35.359
Not giving me some generic answer.
00:20:35.519 --> 00:20:41.359
Here is a great example of personal use by Microsoft CEO, right?
00:20:41.519 --> 00:20:51.680
And you can see how he's asking AI to prepare him to the next meeting based on all his interaction with this person.
00:20:52.000 --> 00:20:56.799
In order to do that, the system really has to know him, right?
00:20:56.960 --> 00:21:04.079
It needs to have access to all of his emails, to all of his communication with this person, right?
00:21:04.559 --> 00:21:11.039
And read maybe all his presentation, all his um whatever he wrote, right?
00:21:11.359 --> 00:21:20.480
This way, you know, the system can really give you something very targeted, very stitched to your need, okay?
00:21:20.720 --> 00:21:26.720
But again, it has to be on the edge, it has to know what are your needs, what is your style, and so on.
00:21:26.880 --> 00:21:30.400
And it can customize to my specific special needs.
00:21:30.559 --> 00:21:34.319
For example, if I have a dyslexia, just just the thought, right?
00:21:34.480 --> 00:21:38.720
It can present me the text in the way that it's easier for me to read.
00:21:39.200 --> 00:21:47.359
Just one example, and you can think about so many of these such customization, and it's gonna be autonomous.
00:21:47.519 --> 00:22:02.400
In the near future, AI models or agents will work for us, will find for us the next vacation, will read, you know, will schedule meetings for us, etc.
00:22:02.799 --> 00:22:08.079
Until then, we need to be, you know, before it happens, we need to trust them.
00:22:08.240 --> 00:22:19.920
And right now, I don't feel so comfortable giving an AI agent the keys to my bank account or to my stock account to do whatever he wants.
00:22:20.079 --> 00:22:23.519
Okay, it will take us some time to really trust them.
00:22:23.759 --> 00:22:25.599
So, but it will happen.
00:22:25.759 --> 00:22:31.279
And until then, until then, feel free to find me on LinkedIn.
00:22:32.640 --> 00:22:34.640
Feel free to find me on YouTube.
00:22:34.960 --> 00:22:45.920
I have a YouTube channel called AI with Guy, um, where I post like very small short videos about you know AI and and stuff.
00:22:46.079 --> 00:22:48.559
And feel free to use the Open Edge platform.
00:22:48.640 --> 00:22:50.079
We built it for you.
00:22:50.319 --> 00:22:52.480
So enjoy it and thank you.