This is a download from BFM 89.9, the Business Station. BFM 89.9, the Business Station. My name is Rich Bradbury. This is Matt Spleynd.
I am not the Matt Spleynd. However, ancient philosophers and mathematicians employed scribes to record their works and observations. Oral traditions have passed wisdom from generation to generation for generations. Where do you record knowledge in an age where millions of pages of information are generated every single day?
For the sake of future generations of NFT key holders, Matt Armitage wants to know. I can't just store all of that information on thumb drives. Hey, Richard. I know that you're not the Matt Spleynd.
Maybe we should have a title for you like no Terrias Ricardus or something like that. We'll come up with something. Sure, we can pull all of that information onto a thumb drive. We can put the internet on a thumb drive.
I'm trying to wrap my brain. That's something that I've watched during this never-ending pandemic that used a plot device of a family that didn't understand what the internet was. They were conned into thinking that they had possession of the hard drive for the server that was the internet. They had to be glowing red button on top.
It'll come back to me eventually. But if you're talking about things like social media posts, photos, the data that most of us generate at work, sure, of course, you can archive all of that information. We have talked about observations before and the efforts by some organisations to preserve as many different data formats and protocols as possible so that when the apes do inherit the earth, they'll be able to pluck that thumb drive of humanity's knowledge out of the statue of the independence is outstretched hand along the beach and they'll get a blueprint of how to open and understand all those files. But when it comes to that hard information, the science and the technology, the engineering, the architecture and the medicine.
In other words, Matt, anything that requires a lot of maths? Yeah, the bit that I don't get. It's not as simple as just opening that file. So it's great that I can have a repository of every kitten video ever made.
Future Matt will be able to start the world's most awesome digital cat museum. But what about all those scientific theories? Because they're not as simple as hitting play on a video player. Someone in the future will still need to understand that information.
So you could argue that in a post-apocalyptic future where the main priorities are reinventing the wheel and getting Netflix up and running again, we might not need to understand things like nanotechnology or string theory. Let me just run something by you. Have you ever imagined that we might not need to prepare for post-apocalypse? No, because that's absurd.
But even assuming that your premise, far fetched as it is, could possibly be correct, there will still be a need to create systems that both preserve and understand the information that we're creating today. So look at a subject like artificial intelligence. I could read about artificial intelligence. I could be reading books, papers, going through news items all day, every day.
And I'd still fall behind. There's so much information that's being generated. So one of the things that we've noted over the years and that it's only very recently that we've seen more concerted efforts to correct is linking scientists and researchers in different fields together, bringing chemists and physicists together, for example. We talked about marine and plant biologists teaming up in the search for new antibiotics.
And that's just the difficulty of getting collaborations for people within the same field, let alone outside the science. So often when we see these things, they tend to come together by happenstance rather than by planning. You mean like the bit-reading cocktos in Sydney? Yeah, that great story that we featured a few weeks ago.
So the amazed resident in Sydney films, the birds flipping bin lids, a local scientist sees it, realizes that there's an evolutionary aspect to what is otherwise just today's cute TikTok clip. And so science has contacts with an animal behavioral research team in Germany. And they're interested enough that they want to launch a study. So of course, you know, a lot of breakthroughs are planned and built on the work created by previous generations of diligent researchers.
But what about all of those outliers like the research on the cocktos? You know, it took decades to prove that something like the Higgs boson actually existed. So what if that work, those theorems had simply faded into complete obscurity? And you know, you wonder as well what other nuggets of inspiration are lying around in the cold science journals, admittedly, you know, it's more likely to be the key to a more efficiently pouring juice box than the solution to climate change.
But you can see what I mean with this. Yeah. But surely, you know, archival and indexing projects like Google Scholar, Google Books should prevent that information from being lost. Yeah.
But as I said earlier, you know, it's not about that information being backed up. It's about being able to find the relevant information across link it. So I can do that with pictures of puppies or lollipops because the AI has been trained to identify and to retrieve those images. And it's the same with text.
I could search for any reference in a book or paper to marshmallows, for example. But when you start talking about scientific theories and especially mathematics, those retrieval engines will need to do more than just pull up some papers. The limitations that most systems have at the moment is that the AI will actually create too many irrelevant correlations. So humans have to check the references that the computers pull up.
And those humans may have been able to find the same or similar information just as easily without the assistance of the machines. And I'm sure, you know, there could be legal implications too. I mean, I don't know how much of an issue that would be, though we have heard of instances where books or papers have been digitized without the copyright owners express permission. So certainly those human researchers might have to spend some time double checking all those attributions.
What might be more an issue, I think, is language, the fact that papers are published in different languages. So an article called Humans Can't Be the Soul Keepers of Scientific Knowledge by the physicist and the editor-in-chief of Nature Reviews Physics, Julia Georgescuk. The article was on Wired. This served as the inspiration for the topic today.
And she quotes the old physics joke that everything has been discovered and it was published in a Russian science journal back in the 1960s. So in that sense, you know, recording information in natural languages becomes a very inefficient way to record it if the purpose is to spread that information and knowledge for others to build on. So essentially, this is a we need a smarter AI story. Well, I can see where you'd get that idea because that's what most of my stories are, but no, not necessarily.
So to be really useful, those knowledge, storage and retrieval systems have to have that ability to contextualize. So they need some understanding of what the text, the jargon, the formula in the equations actually mean, especially with jargon, because the same words can have different meanings across different disciplines. So we can't wait and make the assumption that we'll suddenly have machines capable of making that insight. If we do get to that point, you know, that's great.
But in the meantime, they are both existing and developing systems that will allow the machines that we do have to understand that information in a limited capacity. So actually, this is another one of those humans serving the machines type topics with the suggestion that scientists record their research as computer code rather than in languages like English. So you're suggesting that scientists record their research in machine readable formats, essentially in computer code? Well, thankfully, it's not my suggestion because then everyone would just ignore it.
It's more than a suggestion. It's actually happening. It has been happening for some time. And it's another of those areas where we see a lot of convergence.
So a couple of weeks ago, I think MSP 179, we were talking about AI and language processing, and we talked about codecs, which is open AI system to auto generate code. And the ability of that platform to translate that code into multiple different programming languages. So with this convergence of the technology, you start looking at something that is more akin to a universal translator for scientific knowledge. As long as the information has been inputted in a compatible form, it can then be generated in a format or language that the person searching for it can understand.
And that's potentially a really great leveler. You know, English has often been the default international language for scientific knowledge. In the future, that might be one less worry for a busy scientist, proficiency in a foreign language in order to publish their work across borders. And of course, to help make that knowledge universally available.
You mentioned that this has been happening for a while, but systems like codecs are new. Yeah, so we're back to what I said about coding. Some mathematicians have been recording their work, their proofs and theorems in a language called lean. So we talked about codecs being an assistant.
It guides the user to create code and it makes suggestions to correct it as well. So similarly, lean is a place that you can record, but it's also an assistant, in this case, a proof assistant. So when you input theorem, the system can help to check whether your hypothesis is true or false because it can cross check it against the databases of existing proofs. So it can see where in the theorem the logic breaks down or may need further improvement.
But don't you think this is like adding another layer of complexity asking scientists to jump through yet another? Well, I think we're kind of looking at it back to front. So yes, it may be a big ask to get a Nobel Laureate in her 50s or 60s with a research team to run and an academic role to fulfill to start to learn this new programming language. But the next generation of scientists can learn this language that they share with the machines from the outset.
The Zeno project at Imperial College London is aiming to record the entire undergraduate math syllabus in lean. So if language is like lean throughout the school year throughout school systems, then the scientists of the future will have this kind of machine lingua franca, that common language, which it's likely will be more straightforward to learn and understand than complex natural languages like English and Mandarin, which are likely to be the default scientific languages of the future. Now, doesn't that mean we're making an assumption that all the scientific disciplines are the same, that all this knowledge can be recorded in exactly the same way? Well, that is a good point.
And for these tools to be useful, they have to be expansive and not productive. So that's why I use those elements like GPT-3 and codecs. So if you only had to learn a bunch of rules and guidelines, then science would be, you know, really simple. Julia Jodgescu uses the example of quantum mechanics, which has a clear mathematical description.
Yet there are reams of information out there that try and explain quantum mechanics. The mathematics itself contains really big ideas that most people's brains really struggle to grasp and get it at all. And she makes the argument that machine systems like this may be better at assisting the next generation of scientists to work through those ideas, because what seems counterintuitive to a human brain may not seem counterintuitive to a machine. And is that where these other AI systems platforms come in?
Well, yeah, it may be that Lean isn't suited to scientific knowledge in every field, but we are getting close to a time where scientists across the spectrum can use various forms of machine systems to speed up their work. We talked earlier this year about AI modelling that's speeding up the development of vaccines by working through like the candidates. So we like to think of machines being a one-stop solution, you know, the ultimate simple fix. But it may be that different systems with different capabilities end up storing different parts of the information, which is a little bit like our own brains.
But if we build those systems to talk to each other, to connect, then it becomes moot to a point because as the user, we still end up with information that's indexed, cross-checked and available to us. And those systems will tend to cascade upwards. So eventually, either all news and information will be available in these universal forms, or systems like codecs and GPT-3 will be able to translate it into languages and formats that the engine is comfortable with. Interesting.
Okay, we're sticking with language and AI today. When we come back, a system for combat fake news in encrypted messaging apps attuned into Matt's plane to here on BFM 89.9, the Business Station. Before Friday materialises, BFM 89.9. BFM 89.9, the Business Station.
This is Matt's plane, I'm Rich Bredry. And we've already seen how machines can be used to record knowledge and assist scientists as they search for that next breakthrough. But for many others, misinformation poses a more immediate threat than scientific innovation. And one of the most virulent means of spreading information has been encrypted private messaging apps.
Yeah, for sure. You know, we've heard about the rumours and the disinformation that spreads on platforms and WhatsApp and Signalon and Telegram. We've heard as well about the use of these platforms by criminal and terrorist groups. And I've long argued that despite those risks, we still need messaging services that are protected from intrusion and surveillance.
And I haven't changed my position on that. But at the same time, we need systems that can check the lies and destruction that are making the rounds on those social platforms and try and reduce some of the harm that that information may cause. But how does that gel with encryption? Either you can see the message or you can't.
Well, yeah, I mean, that's the either-all with encryption. Either you can break it, in which case it's useless, or you can't see anything. So in that sense, I think what we're looking for is not so much a solution. As a compromise, we saw this with Apple recently with its plans to scan the devices of its users for child-borne images, which the company quietly abandoned.
Because the issue is that the machine, the bot can essentially be tasked with looking for anything once it's got access to the phone. Today, it might be child-borne images. And I know this is an absurd example, but what if a crazy dictator banned orange cats and ordered the owners of orange cats arrested? In goes the crawler, looking for evidence of orange cats on everybody's phones.
And of course, I'd be the first to be executed in that case. You can't argue that AI should be able to bypass encryption simply because it isn't sentient. It doesn't know what it is that it's looking for. It's who that information it gathers is passed onto that count and how that information can then be used.
So what's the compromise then? Well, WhatsApp has long had a feature that limits the number of times a particular message can be forwarded. Some of us have run into that by forwarding invitations to a party, for example. Or if you've tried to forward a message claiming that squirrel tails can cure COVID and are also great for dusting your home or something equally absurd.
In any case, unlike on Facebook or Twitter where fact checkers can view the content of the messages, platforms like WhatsApp can only halt the spread based on that frequency that it's forwarded. They can't see what the message is. So the compromise is a system created by a team at George Washington University in Washington, DC. The system is called fuzzy anonymous complaints tally system.
Wait, fact. Yeah, it's another one of those tortured acronyms though. I would also go to great lengths to create. I think it's great.
But you know, it's a combination of that numerical forwarding thing along with a reporting mechanism. It's the anonymous part. So the idea is that users flag and report offensive posts. And when those posts reach a certain threshold, then that message's contents are revealed to the third party that's been tasked with monitoring this disinformation.
But what would that limit be then? Well, the Washington team hasn't set a threshold. So they argue that it would be the policy of the platform that sets it rather than anything that they arbitrarily decide. And they acknowledge that the tools could be used maliciously.
So someone who doesn't like you could forward all of your messages as fake news. So the system would also see how often that particular user is reporting posts to see if it's somebody who's trying to game the system. It can also place limits on how many times users can report posts in any kind of given timeframe. So basically it sounds like a system to score complaints.
Yeah, I mean, that's really effectively what it is, hence that tally in the name. But it has the advantage that the individuals making the complaints aren't revealing the contents to the moderator until it reaches the threshold where the content is revealed. So the team behind the system, they are open about the shortcomings. For example, what action would the platform take against someone spreading disinformation?
Would they ban the user? And if so, what kind of appeals procedure should be in place with platforms like Facebook and Twitter? It's usually quite obvious because you are posting publicly and their terms and conditions set out what their community standards are. But even within that sphere banning or limiting the activity of users is still extremely contentious.
So how would that work on systems that promote privacy and the protection of encryption? And of course, the system doesn't deal with issues like terrorism and criminality, because those groups will quickly adapt to the system. They'll change the contents that it isn't flagged by those systems. And it also doesn't address the issue of misinformation amongst the converted groups that have no interest in challenging misinformation or disinformation.
It would only be relevant when it leaks to people outside the group. But these are conversations that we do need to have and the conversations that those tech providers should be having. The world of news and information has changed probably irrevocably. But as yet, society hasn't kept pace with those changes.
So where else are we seeing these kind of cat and mouse games between AI and people? We reported earlier in the year or maybe last year about the use of masks in protests around the world before the pandemic to avoid facial recognition systems. And scientists in China have created systems that could still identify mask-wearers. Now this was promoted as a COVID protection measure.
They were teamed with temperature scanners. But you can see that being used post-pandemic to identify people who've tried to mask the bottom half of their face. So this is from a report from a new scientist about measures to use AI to tell people how to cheat facial recognition software simply by applying makeup to full it. This is from a system developed at a university in Israel.
The researchers used an adversarial system to pit two aisles against one another. One system essentially reverse engineers the facial recognition. So it looks for the points that those systems use as the unique identifiers for an individual. And the other looks for ways to alter the appearance of the individual so that they're no longer identified as that same person by the system.
So it's essentially creating a map of changes that someone can take to alter their appearance in the real world. Got it. But how accurate or I guess inaccurate were the results? Well highly inaccurate.
And it's not often you celebrate an inaccurate result. So the facial recognition system the researchers used was accurate around 47% of the time in normal conditions which goes to show you the real world limitations of a lot of commercial recognition systems. But this dropped to only 1.2% when the subjects followed those cues and applied makeup to alter their faces. So almost everyone became not invisible but unidentifiable to those recognition systems.
And where's the difference in wearing a mask or radically altering your appearance with makeup? Well you know we're making a assumption that you have to radically alter your face. For example you know make yourself up to look like a clown covered in bright colors. And that's not the case at all.
So the Israeli research was done using natural tones of makeup. And that's what makes this story interesting because it dovetails with current trends. It's a bit like the convergence that we were talking about in the first part of the show. You can use makeup to create new facial contours to alter the appearance of the shape of your face.
And this has become one of the tick tock makeup trends of the year. And that kind of contouring is the kind of thing that can be used to gain these recognition systems. As the new scientist piece points out it allows you to alter your appearance without drawing attention to yourself which could be something that's increasingly important as countries move to these kind of more authoritarian models of governing and control. Especially using a lot of automated tools and recognition systems.
So this is the idea that makeup could be the next battleground of the privacy debate. And for me that's something that's really fascinating. Makeup and matte. That's what you should call the show.
Not what you expect to hear. Thanks very much for that, matte. Thank you. I think I'd need a lot of makeup to alter my appearance.
Anyway, if you want to see what it actually looks like you can find matte on Instagram and Twitter at Culture Matte. You can also head over to culturepop.com for transcripts of these shows and information about CulturePop. And it's consulting services and if you did miss any part of the show not to get you can download the podcast wherever you normally did it from. I recommend the BFM app.
It's available from the Apple App Store or Google Play. My name is Rich Bradbury. This is BFM 89.9, the Business Station. Thank you for listening to this podcast.
To find more great interviews go to BFM.my or find us on iTunes. BFM 89.9, the Business Station.