The AI News You Might Have Missed This Week
00:00:00.000 | The goal of this video is simply to show you 7 AI advances that you might have missed this week.
00:00:06.440 | Sam Altman recently said that in a world of AGI, everything happens much faster.
00:00:11.440 | But as far as I can see, AI developments are already almost impossible for a human to keep up with.
00:00:17.520 | So, in no particular order, let's get started.
00:00:20.740 | First, video calls look like they're about to get 3D.
00:00:24.220 | Let's take a look at how NVIDIA Aerial and NVIDIA Maxine 3D, running on the NVIDIA Grace Hopper Super Chip,
00:00:31.180 | can enable 3D video conferencing on any device without specialized software or hardware.
00:00:36.880 | This brings a new dimension to video conferencing with Maxine 3D visualization.
00:00:41.820 | Engage with others more directly with enhanced eye contact.
00:00:45.700 | And personalize your experience with animated avatars, stylizing them with simple text prompts.
00:00:52.200 | And it isn't just NVIDIA.
00:00:53.600 | Here's Google.
00:00:54.200 | Google's new Project Starline prototype.
00:00:56.740 | You know, you were so used to seeing a two-dimensional little, you know, box.
00:01:01.200 | And then we're connecting like this.
00:01:03.020 | And that feeling of being in front of a person is now replicated in Starline.
00:01:08.680 | Speaking of connecting the world, here is GPT-4 doing geography in a paper you might have missed from this week.
00:01:15.180 | The paper proves that even without access to the internet,
00:01:18.260 | GPT-4 knows a lot more granular detail about the world than you might first imagine.
00:01:23.460 | I'm not saying that it's a bad thing.
00:01:23.780 | I'm not saying it knows where you live, but it's not too far off.
00:01:26.880 | Take this example.
00:01:27.920 | It could recreate the Hong Kong Mass Transit Railway from memorization.
00:01:33.120 | This wasn't through using web browsing.
00:01:35.080 | It could recreate this diagram, giving the latitude and longitude coordinates of each of the stations in this transit line.
00:01:42.740 | Obviously, it's not perfect, but it's pretty incredible that it's got this mental map of the world.
00:01:47.840 | GPT-4 can do elevations as well.
00:01:50.060 | And here is it trying to recreate the topography of the atmosphere.
00:01:53.080 | And here's the map.
00:01:53.360 | It gets pretty close.
00:01:55.300 | One of the ways they tested GPT-4 was to ask it something like this.
00:01:59.160 | Please provide the latitude-longitude coordinates for the outline of X,
00:02:04.020 | where X was a continent or a river or a country,
00:02:06.900 | as a Python list of tuples consisting of approximately 50 points arranged clockwise.
00:02:12.040 | And they describe how it did really well for quite a few countries and rivers,
00:02:16.760 | but kind of flopped on Africa.
00:02:18.880 | But honestly, when I read this paper, I was skeptical that GPT-4 knew that list,
00:02:23.340 | because it was a little about Africa.
00:02:24.300 | So I gave this exact question to GPT-4 with Code Interpreter.
00:02:28.160 | Now, interestingly, it would sometimes deny that it had the ability to do this,
00:02:31.880 | but with enough encouragement, it outputted these coordinates.
00:02:35.260 | And here is the end result in Google Earth.
00:02:38.180 | I think that's a pretty impressive outline.
00:02:40.400 | Obviously, a few points are a bit off, like this point here isn't really on the coast, nor is this point.
00:02:46.640 | But it really knows the outlines of countries, continents, rivers.
00:02:50.660 | So I'm not sure if Code Interpreter had an impact there,
00:02:53.320 | or a model update, but the researchers kind of underplayed what GPT-4 could do
00:02:58.400 | by presenting this outline of Africa.
00:03:00.720 | Now, I am sure that some of you are thinking,
00:03:02.900 | that's not that interesting, not that impressive.
00:03:05.120 | But check this out.
00:03:06.420 | In an indirect kind of way, GPT-4 knows where it was made.
00:03:10.540 | It was able to construct a map of the semiconductor supply chain.
00:03:15.140 | It not only knows about the design, manufacturing materials, equipment,
00:03:18.980 | and tools that go into the hardware that helps make GPT-4,
00:03:22.740 | it also knows about the design of the hardware that helps make GPT-4.
00:03:23.300 | It also knows the locations of where this is all done.
00:03:26.220 | And as the authors later say,
00:03:28.020 | looking to the future, if frontier models beyond GPT-4 continue to advance in capabilities,
00:03:33.840 | the geographic knowledge and planning abilities present in the current model
00:03:37.960 | may later evolve to represent a significant risk through misuse or misalignment.
00:03:43.960 | On a much less important note, did you notice how I could do this demo
00:03:47.880 | without that sidebar of all my previous chats?
00:03:50.860 | That's because OpenAI have brought in this new button.
00:03:53.280 | You can see the chat here where you can hide the chats.
00:03:55.800 | And as a bonus, some of you may not know that you can now share a link of the chats that you've already done
00:04:01.580 | just by clicking that button to the left.
00:04:03.440 | And as it says, messages you send after creating your link won't be shared.
00:04:07.340 | So if you carry on the conversation, people won't be able to see.
00:04:09.900 | But anyone with the URL will be able to view the shared chats.
00:04:13.880 | But before we move on from OpenAI and ChatGPT, I did find this table really quite interesting.
00:04:19.380 | It gives the daily average number of visits to each of these sites,
00:04:23.260 | along with the visit duration.
00:04:25.140 | And there's two things that strike me from this table.
00:04:27.720 | The first is how much more popular ChatGPT is compared to Google's BARD.
00:04:31.940 | It's got about 15 times the number of visitors who stay for about twice as long.
00:04:36.480 | But look at the dark horse on the right, Character AI.
00:04:39.820 | I've talked about them a couple of times before.
00:04:42.320 | And while their daily average visit total isn't too crazy, look at the visit duration.
00:04:47.700 | In terms of grabbing people's attention and keeping it, they are truly a dark horse.
00:04:52.440 | Next, I want to talk about the number of visits to each of these sites.
00:04:53.240 | I want to briefly dip into augmented reality.
00:04:56.120 | We are going to be creating our own worlds and living in them.
00:04:59.800 | Some people, like in this video, might choose to live their lives as if they were in an animation.
00:05:04.800 | Others might see augmented reality as a way of augmenting their intelligence or memory live.
00:05:10.920 | My prediction would be that wearables that resemble things like Google Glass might flop.
00:05:19.680 | But something like an always-on app on your phone mediated through Gmail
00:05:23.220 | could become really popular or even enforced in certain workplace settings.
00:05:28.680 | All of this reminded me of a recent video about conducting a video interview with help
00:05:33.300 | from GPT 3.5.
00:05:34.740 | What about your development areas?
00:05:37.840 | What do you have identified as your greatest and biggest improvement areas?
00:05:43.020 | And what have you done to improve them so far?
00:05:46.100 | I would say my greatest development area is my communication skills.
00:05:52.540 | I work on improving my communication skills.
00:05:53.200 | I'm improving my ability to clearly convey my thoughts and ideas to others.
00:05:58.320 | Of course at the moment this is only really viable with GPT 3.5 because of inference speed.
00:06:03.640 | But OpenAI are aggressively planning a cheaper and faster GPT 4.
00:06:08.360 | I wouldn't be surprised if video interviewers soon require you to take out any headphones.
00:06:13.580 | Although I guess with Maxine 3D you could maintain eye contact with the camera while
00:06:18.320 | you're actually reading off a GPT 4 teleprompter.
00:06:21.860 | Anyway what about gaming?
00:06:23.180 | This is Nvidia's Neuralangelo where you can take a 2D video and turn it into a detailed
00:06:29.220 | 3D landscape with high fidelity.
00:06:31.720 | My first thought turned into imagining the kind of things you could then bring into games
00:06:36.080 | using Unreal Engine 5.
00:06:38.260 | This is a recently trailered horror game, link in the description, but don't worry I'm
00:06:42.140 | only going to show you 2 or 3 seconds of it.
00:06:44.240 | It's getting to the point where it's quite hard to believe that this is a game, but it
00:06:47.500 | is.
00:06:48.500 | And on games don't forget this, look at the realism that can now be achieved in terms
00:06:52.500 | of skin technology.
00:06:53.160 | The texture and movement.
00:06:54.640 | For the final bit of AI news that you might have missed, I want to focus on AI drug discovery.
00:06:59.980 | I think there's no question that there is a before and after in drug discovery and one
00:07:08.140 | of them is AI.
00:07:09.660 | Alan Espuruguzic is the director of the University of Toronto's Acceleration Consortium which
00:07:14.420 | in April 2023 received a $200 million grant to build an AI powered self-driving lab.
00:07:20.760 | The Acceleration Consortium has already been using AI.
00:07:23.140 | To help discover molecules that have potential drug-like traits that can be used to develop
00:07:27.480 | life-saving treatments.
00:07:28.480 | Developing a drug can be up to a decade and this is just the discovery piece.
00:07:33.240 | So that process let's say takes a year or two and we compress it to 45 days in that
00:07:37.960 | case and then 30 days recently.
00:07:40.920 | In January 2023 the Acceleration Consortium used an AI powered protein structure database
00:07:46.060 | called AlphaFold to design and synthesize a possible liver cancer drug in just 30 days.
00:07:51.120 | Within two weeks we can formulate a new drug.
00:07:53.120 | We can create a new drug as well as some people have done it in years.
00:07:55.920 | Suddenly AI has surpassed any human created algorithm.
00:07:59.180 | AI what allows us to do is lower the bar of what you need to do certain things and therefore
00:08:03.560 | more and more people will have access to it.
00:08:05.460 | In general unleashing more innovation in the planet.
00:08:08.100 | Same token, someone with nefarious intentions could unleash very dangerous, deadly chemicals
00:08:14.800 | on the world.
00:08:15.800 | Absolutely.
00:08:16.800 | I am an optimist but I'm also aware of these pitfalls that very soon will face us.
00:08:22.200 | And videos like that are not just for the people.
00:08:23.100 | They are for the people.
00:08:24.100 | And that is why I agree with Sam Altman when he says a much faster rate of change is his
00:08:27.840 | single highest confidence prediction about what a world with AGI in it will be like.
00:08:32.940 | I follow AI news full time and can barely keep up so I can only imagine what the situation
00:08:39.040 | will be like when we get full AGI.
00:08:41.920 | But until the very last moment that it's humanly possible to keep up with the news I will try.
00:08:46.780 | So thank you so much for watching to the end and have a wonderful day.