← AI Deep Dive
Loading summary
Transcript227 lines
- [00:07]
A
It feels like every day there's a major new AI development. It's moving so fast it can be really hard to keep up with it all.
- [00:14]
B
Yeah, it's really gone from, you know, this sort of theoretical thing to this thing that is just impacting every aspect of our lives.
- [00:22]
A
Right. It's not just the future anymore.
- [00:24]
B
Yeah. And it's often in ways that we don't even realize.
- [00:27]
A
And that's exactly why we're doing this deep dive today.
- [00:29]
B
Absolutely.
- [00:30]
A
You've brought a fantastic set of sources about the latest in AI. And our mission today is to distill the most crucial and the most interesting developments specifically for our listeners. The listener who wants to stay informed but doesn't want to like drown in the news.
- [00:48]
B
Right. Consider this your AI cheat sheet. Like, here's what you need to know.
- [00:51]
A
Yeah, we're going to try to extract the signal from the noise.
- [00:53]
B
Exactly.
- [00:54]
A
We've got four really interesting areas to cover today from these sources. First, we're going to be looking at how Google is using AI to like totally revolutionize travel planning. Then we'll dive into some of the really impressive advances in image generation with ideogram 3.0.
- [01:11]
B
Yeah, that's a cool one.
- [01:12]
A
And then on a more serious note, we're going to examine this newly uncovered AI powered censorship system in China.
- [01:20]
B
Right.
- [01:20]
A
And finally we'll tackle the copyright questions raised by OpenAI's new ability to generate Ghibli style images. So be prepared for some surprising facts and some really interesting insights. Are you ready to dive in?
- [01:32]
B
Let's do it.
- [01:33]
A
All right. Let's kick things off by talking about something we can all relate to, something very personal. Planning a trip or a vacation.
- [01:41]
B
Yeah. I mean, everybody loves to travel.
- [01:43]
A
I know I do.
- [01:43]
B
Or at least the idea of it.
- [01:45]
A
Yeah, yeah, exactly. But it's really interesting to see how AI is now moving into like, really personal aspects of our lives, like planning a vacation.
- [01:55]
B
Right.
- [01:56]
A
This first set of information is all about the ways that Google is using AI to make the travel planning experience way easier.
- [02:03]
B
Okay.
- [02:04]
A
And this covers several of their, like, familiar platforms that we all use every day. Things like Google Search, Google Maps, and also their AI model, Gemini.
- [02:13]
B
Right, right. So it's not like they've, you know, built a whole new thing that nobody's going to use. It's built right into the stuff people already are using.
- [02:20]
A
Yeah, that's really smart. Right. I mean, it makes it much more accessible to everybody.
- [02:25]
B
Right.
- [02:25]
A
So imagine you're thinking about like a nature focused trip to Costa Rica.
- [02:30]
B
Okay.
- [02:31]
A
Now, within Google Search, you can use their new AI overviews to just type something, like, create an itinerary for Costa Rica with a focus on nature. And boom, like almost instantly it'll, it'll generate like, these trip ideas for you.
- [02:45]
B
Okay.
- [02:46]
A
And what's even cooler is that you can browse photos, you can read recent reviews, and you can see all these suggested locations on a map, an integrated map that you can expand, and it's all within the search results.
- [02:58]
B
Yeah, that sounds really handy. Like you don't even have to leave that screen.
- [03:01]
A
Exactly. And then once you've like zeroed in on some ideas that you like, you can like directly export these recommendations to Google Docs or Gmail, or if you're like a big Maps user, you can save them to a custom list on Google Maps.
- [03:16]
B
Okay, that's pretty cool.
- [03:17]
A
It's really cool. One thing to note is this is currently for English language queries in the US on mobile and desktop.
- [03:23]
B
Okay.
- [03:24]
A
But you know, it's obviously like a signal of where Google is heading with all this.
- [03:29]
B
Right, right, exactly.
- [03:30]
A
So now, speaking of making things easy, Gemini's Gems feature is now free for everybody.
- [03:36]
B
Oh, cool.
- [03:37]
A
And remember, Gems allows you to, like, create your own tailored AI expert within Gemini.
- [03:44]
B
Right, right, right, right.
- [03:45]
A
So you can essentially build like a dedicated trip planner gem.
- [03:48]
B
Yeah.
- [03:48]
A
And it can help you choose a destination. It can, it can suggest what to pack based on your activities. It's really wild.
- [03:55]
B
It's like having your own personal travel agent.
- [03:57]
A
It's like, better than a travel agent. Right? It's, it's. That's like, knows you better than a travel agent.
- [04:01]
B
Yeah, that's true. That's true.
- [04:03]
A
Okay. And this next feature is going to be a big help for those of us who, you know, keep an eye on our budgets when we travel.
- [04:09]
B
So everybody basically.
- [04:11]
A
Yeah. Google now has hotel price tracking built in. So, like, similar to how they have the flight price alerts.
- [04:17]
B
Nice.
- [04:18]
A
So you put in your dates, you put in your location. You can even filter by criteria like, like hotel star rating or proximity to the beach. Right. If the price drops below a certain point, you'll get an email notification. And this is rolling out globally on both mobile and desktop browsers.
- [04:35]
B
That's a great feature, especially with the way prices fluctuate so much.
- [04:39]
A
All right, and now get ready for this one, because this is, like, really ingenious. I was really impressed by this. Think about all those times you've been, you know, browsing Instagram or whatever, and you see a photo somewhere that's like, wow, I want to go there. And you take a screenshot.
- [04:55]
B
Oh, yeah.
- [04:56]
A
Google Maps will now allow you to grant it access to your photo library, and it will identify, like, places that are mentioned in those screenshots, and then you can review those locations and save the ones that interest you to, like, your own personalized map.
- [05:10]
B
That's so clever.
- [05:11]
A
It's really smart. This one's currently launching in the US in English on iOS, but they said Android is coming soon.
- [05:17]
B
Okay, cool.
- [05:18]
A
Yeah. So out of all these AI powered travel features, like, which one jumps out at you as something that you would use?
- [05:23]
B
Well, I think that screenshot integration is pretty impressive, you know, I mean, I have a lot of those random screenshots completely forgotten about me too.
- [05:31]
A
Yeah.
- [05:32]
B
But I also think that having that Gemini gem as like, a trip planner that knows you would be incredibly useful, you know?
- [05:41]
A
Yeah, I mean, that's a really smart idea.
- [05:44]
B
Just something to brainstorm with and help you. Help you remember things. Okay, well, let's move on from traveling, as fun as that is.
- [05:52]
A
Yeah, yeah.
- [05:52]
B
And talked about another area where AI is making huge progress. Image generation.
- [05:58]
A
Yes. Get ready to be really impressed.
- [06:00]
B
Yeah, this is a good one. This next Source introduces Ideogram 3.0, and it represents a major step forward in AI image creation big time. What's impressive about ideogram 3.0 is that it's not just about creating an image from a text prompt. They're focused on realism, creative designs and consistency in style. And reportedly, it's already outperforming other leading text to image models, like in human evaluations, you know.
- [06:30]
A
Yeah. They're not just saying, like, people have, like, tested it against other things and it's performing really well. They're really emphasizing how accurately the generated image matches what the user asked for in the text prompt.
- [06:44]
B
Right.
- [06:44]
A
The level of photorealism and even the quality of text rendering, like the actual text, like, within the image.
- [06:52]
B
Okay. Wow.
- [06:52]
A
And they have some super interesting new features. One is called Style References, and that allows you to upload up to three images to, like, influence the AI's understanding of the aesthetics that you want.
- [07:04]
B
Oh, interesting. So instead of trying to, like, describe in words, you can just show it examples.
- [07:08]
A
Exactly. Which makes way more sense. Right. I mean, how do you describe some complicated visual style in words?
- [07:14]
B
Right. That could be very challenging.
- [07:15]
A
Very challenging.
- [07:16]
B
Yeah.
- [07:16]
A
Okay. Then there's a random style feature which can actually draw upon their library of 4.3 billion style presets to offer, like, you know, unique and unexpected stylistic combinations. 4.3 billion 4.3 billion.
- [07:33]
B
That's a lot of styles.
- [07:34]
A
I mean, that's a huge number. And if you discover a style you like, you can save it and reuse it using what they call a style code. They gave, like, some fun examples. Bowling alley, Cadillac Diner, treehouse. So, you know, imagine being able to create an image and say, I want this image to feel like a treehouse.
- [07:52]
B
Wow. Yeah. That's a level of control that I haven't really seen before.
- [07:54]
A
Right. And it's not just about generating, you know, like a realistic or uniquely styled image. Ideogram 3.0 is apparently also very good at, like, creative design tasks, especially things that involve text and layout. You know, think graphic design, advertising, marketing.
- [08:10]
B
Oh, right, right.
- [08:11]
A
They say that it can produce, like, really stylish and accurate text, even within very, very intricate visual arrangements. And they even have a batch generation feature that allows you to customize graphics at scale.
- [08:25]
B
Okay.
- [08:26]
A
So, I mean, you know, this has pretty huge implications for businesses.
- [08:30]
B
Yeah, right.
- [08:31]
A
Like a small business, they don't necessarily have, like a big design budget or a design team.
- [08:36]
B
Right, right. They could just create their own logos and stuff.
- [08:39]
A
Exactly. Logos, posters, you know, concepts for landing pages. You could even do product photography.
- [08:44]
B
Wow.
- [08:44]
A
And then in terms of stunning realism, which is like one of the things they're really emphasizing, it sounds like it can generate these super detailed scenes with like, you know, sophisticated spatial compositions, nuanced lighting, realistic environmental details.
- [08:59]
B
Wow.
- [09:00]
A
So for you, the listener, if. If you do anything that requires visuals, like for work, for. For your personal projects, I mean, the creation of realistic and really customized images is becoming way more accessible. What would you. What would you create?
- [09:14]
B
Oh, no, I'd have to think about that.
- [09:15]
A
Yeah. There's so many possibilities.
- [09:18]
B
Yeah. It's very cool. But we need to shift gears now to a much more serious topic, much more concerning development with AI.
- [09:27]
A
Right.
- [09:28]
B
This next source about the use of AI in censorship.
- [09:31]
A
Yeah.
- [09:31]
B
Which is pretty scary.
- [09:33]
A
It is.
- [09:34]
B
This concerns the discovery of a leaked database that reveals China's new system for automatically flagging content that the government might find sensitive.
- [09:44]
A
Right. And the scale of this thing is, like, really alarming.
- [09:47]
B
Yeah. What really stands out is the sheer amount of data they've used to train this system. We're talking 133,000 data points, and it covers this huge range of topics.
- [09:59]
A
Things like complaints about rural poverty, reports of corrupt officials, even entrepreneurs asking for help because they're being harassed by the police.
- [10:07]
B
Right. So this isn't about blocking keywords, you know, that are related to, like, a specific protest or something like that.
- [10:14]
A
Right.
- [10:14]
B
This is much more proactive.
- [10:15]
A
Right. It's about controlling information before it even spreads.
- [10:18]
B
Right. It's about, you know, like, nipping dissent in the bud before it has a chance to gain traction.
- [10:23]
A
And Xiao Kang, a researcher at UC Berkeley who specializes in Chinese censorship, he said that this is clear evidence that the Chinese government wants to use LLMs, which are those large language models, to basically improve their ability to repress dissent.
- [10:40]
B
Right, right. That's pretty chilling.
- [10:42]
A
It is chilling.
- [10:43]
B
Yeah.
- [10:43]
A
And this actually lines up with a report from OpenAI back in February.
- [10:47]
B
Okay.
- [10:48]
A
And they found that there were entities in China using LLMs to track anti government sentiment online, and they were targeting people who they thought were dissidents. So, you know, when. When reporters asked the Chinese embassy about all of this, they issued a statement saying that this was all groundless attacks and slanders against China, and they emphasized their commitment to the ethical development of AI.
- [11:12]
B
Hmm, interesting.
- [11:14]
A
It's also worth noting how this database was discovered, how it came to light.
- [11:17]
B
Right.
- [11:18]
A
It was found by a security researcher named Ned Askari, and it was on an unsecured elasticsearch database. And this database was hosted on a server that belonged to Baidu.
- [11:28]
B
Oh, okay.
- [11:29]
A
So you know, this doesn't necessarily mean that Baidu was involved or even that the Chinese government knew about it.
- [11:33]
B
Right, Right.
- [11:34]
A
But it does. It does raise questions about, like, the security of this kind of information. And the data itself is really recent. It has entries all the way up to December 2024.
- [11:44]
B
Wow.
- [11:46]
A
So how does this system actually work?
- [11:48]
B
Yeah, I'm curious about that.
- [11:50]
A
So it prompts an unnamed LLM to analyze content and is looking for material related to a list of sensitive political, social, and military topics.
- [12:01]
B
Okay.
- [12:01]
A
And if the content falls into one of these categories, it's marked highest priority.
- [12:06]
B
Okay.
- [12:07]
A
And then a human will review it and decide whether to censor it or not.
- [12:10]
B
Got it.
- [12:10]
A
And these highest priority topics are really revealing. They give us, like, a glimpse into, like, what the authorities are most concerned about.
- [12:17]
B
Right?
- [12:17]
A
So things like pollution and food safety scandals. Those can cause, like, widespread unrest, reports of financial fraud, labor disputes. Anything they could get, like, challenge the power of the government.
- [12:30]
B
Right, yeah. Makes sense.
- [12:31]
A
And then political satire is also a major red flag.
- [12:35]
B
Okay.
- [12:36]
A
Especially if it uses historical analogies to criticize the current leadership. And anything about the political status of Taiwan is immediately flagged.
- [12:46]
B
Oh, wow.
- [12:46]
A
As well as military matters like troop movements, exercises, and information about new weaponry. Okay, so they actually included 10 representative examples of the content that was collected in this data Set. Okay, so we can really see, like, concretely what the system is targeting.
- [13:04]
B
Right, right.
- [13:05]
A
This includes a business owner who was complaining about corrupt local police, people expressing their hardship due to rural poverty, a news report about a CCP official who was expelled for corruption and for believian superstitions, which. Which is, like, seen as undermining the party ideology, I guess.
- [13:24]
B
Right, right.
- [13:25]
A
And then there's commentary on Taiwan's military capabilities, details about a newly developed Chinese jet fighter, even a proverb about the impermanence of power. It said, when the tree falls, the monkeys scatter.
- [13:39]
B
Huh. Okay, interesting.
- [13:41]
A
And the report notes that the term Taiwan appears over 15,000 times in the data set.
- [13:47]
B
Wow.
- [13:48]
A
So, I mean, clearly that's. That's like a really sensitive topic.
- [13:52]
B
It's clearly a major focus.
- [13:53]
A
Yeah. So the stated purpose of all this data is for what they call public opinion work.
- [13:58]
B
Oh, okay.
- [13:58]
A
And experts say this is often used by the cac, which is the Cyberspace Administration of China, to describe their efforts in censorship and propaganda.
- [14:07]
B
Okay.
- [14:07]
A
So the goal is to control the narrative, protect the government's image, and suppress any alternative or dissenting views.
- [14:13]
B
Right.
- [14:14]
A
And, you know, it's. It's pretty sobering to think about this in the context of what Xi Jinping has said about the Internet being the front line of this public opinion work. So this AI driven censorship, it's a. It's a major evolution from traditional keyword blocking, you know.
- [14:28]
B
Yeah.
- [14:28]
A
It can identify nuanced criticism at a massive scale, and it's constantly learning and getting better.
- [14:34]
B
Yeah. So this raises the question, like, what does this mean for freedom of expression globally?
- [14:39]
A
Yeah, that's the big question. Right. What are the implications for. For the free flow of information.
- [14:43]
B
Right.
- [14:44]
A
It's a pretty profound issue to think about.
- [14:46]
B
Absolutely.
- [14:47]
A
Okay, so let's move on to our final topic.
- [14:49]
B
Okay.
- [14:49]
A
And this one's a little bit more lighthearted, I guess, but it also involves some tricky legal issues. We're going to talk about OpenAI's new image generation feature and the copyright concerns that have come up around it.
- [15:02]
B
Right.
- [15:02]
A
Okay. So basically what happened was OpenAI launched ChatGPT's new image generator.
- [15:09]
B
Right.
- [15:10]
A
And almost immediately, social media was just flooded with memes and images that were created in the style of Studio Ghibli.
- [15:18]
B
Oh, yeah, I saw those. Those were everywhere.
- [15:20]
A
It was like a viral sensation. There were Ghibli versions of Elon Musk, scenes from the Lord of the Rings reimagined in that classic Ghibli style.
- [15:29]
B
Yeah.
- [15:30]
A
Even cartoon versions of Donald Trump and even Sam Altman, the CEO of OpenAI, updated his profile picture to, like, a Ghibli version of himself.
- [15:40]
B
And that reminds me of, like, you know, the excitement we saw around Google's Gemini Flash feature and how people were using it to, like, remove watermarks from images.
- [15:51]
A
Right, exactly.
- [15:52]
B
And then Google had to disable that feature.
- [15:54]
A
So, you know what's so interesting here is how easy it is to use these AI tools to just replicate the style of a copyrighted work, you know?
- [16:04]
B
Yeah. Just through a simple text prompt.
- [16:06]
A
Yeah. And this raises a big legal question. Does training an AI model on copyrighted material constitute copyright infringement?
- [16:13]
B
Right. Because these models are trained on vast amounts of data.
- [16:17]
A
Yeah. And there are lawsuits going on right now about this, like, with generative AI developers.
- [16:22]
B
Okay.
- [16:22]
A
And this is where it gets interesting. There's this intellectual property lawyer, Evan Brown, and he says that artistic style is not explicitly protected under copyright law.
- [16:31]
B
Hmm.
- [16:32]
A
So that would suggest that maybe OpenAI isn't actually breaking the law by creating these Ghibli, like, images.
- [16:37]
B
Okay, so that's one side of the argument.
- [16:39]
A
Right. But then he also points out that it's very likely that OpenAI trained their model on millions of Ghibli frames.
- [16:46]
B
Right.
- [16:46]
A
And so the question becomes, does that fall under fair use?
- [16:50]
B
Right. That's the big question, the gray area.
- [16:52]
A
Right. And that's what all these lawsuits are about. You know, the New York Times is suing OpenAI. Publishers are suing them. There are lawsuits against Meta, against Midjourney. It's. It's a huge issue.
- [17:01]
B
Big deal.
- [17:02]
A
So OpenAI put out a statement saying that while ChatGPT will refuse a request to, like, replicate the style of a specific living artist, okay. It does allow the replication of broader studio styles. But the article, the article that you brought points out that, like, many of these styles were actually like, developed and defined by, you know, living artists like Hayao Miyazaki.
- [17:29]
B
Right, right.
- [17:29]
A
Like, he's the heart of. Of Studio Ghibli.
- [17:31]
B
Yeah.
- [17:32]
A
So, and. And people are experimenting, Right. They're. They're using these tools to recreate styles from other studios, from other artists. Like somebody created, like, a Dr. Seuss style portrait of Mark Andreessen.
- [17:44]
B
Oh, okay.
- [17:44]
A
Somebody did Pixar style wedding photos. So, you know, and. And there was this test comparing different AI image generators. And apparently OpenAI's new tool is the one that created the most accurate Ghibli replicas. So it seems like the demand for these new AI image creation features is just, like, through the roof. OpenAI actually had to temporarily delay rolling it out to their free tier users because so many people were trying to use it.
- [18:13]
B
Wow.
- [18:13]
A
So, you know, the technology is clearly advancing really fast and people love it.
- [18:17]
B
Right.
- [18:18]
A
But the legal side of it is really murky and it's probably going to take years for the courts to sort it all out. So where do you draw the line?
- [18:26]
B
Hmm.
- [18:27]
A
How do you balance the need to encourage innovation with the need to protect the rights of creative people and studios?
- [18:33]
B
That's a tough question.
- [18:34]
A
Yeah, it's a big one. So let's wrap up our deep dive by summarizing what we've talked about. We started by talking about how Google is using AI to make travel planning way easier.
- [18:42]
B
Right.
- [18:43]
A
And then we looked at ideogram 3.0 and all the amazing things it can do in terms of creating realistic and stylized images. Then we shifted to a much more serious topic. We talked about this AI powered censorship system that was uncovered in China.
- [18:58]
B
Right.
- [18:58]
A
And finally we discussed this whole controversy around AI generated GIBLA style images and the copyright issues that it raises.
- [19:07]
B
Right.
- [19:07]
A
So hopefully you, our listeners now have a good overview of some of the most important AI developments that are happening right now.
- [19:16]
B
Yeah, we gave you the shortcut so you don't have to read all those long articles.
- [19:19]
A
Exactly. Now we want to leave you with one final thought. As AI keeps advancing and creating these, like, these amazing and also sometimes scary possibilities, what kind of shifts do you think are going to happen in society and what questions should we be asking now to prepare for those shifts?
- [19:36]
B
Yeah, that's something to think about.
- [19:38]
A
Yeah, it's a lot to think about. So thanks for joining us for this deep dive. Yeah, thanks for listening and we'll catch you next time.