Robot War part 3: The Race
In "Robot War part 1" 1 , I talked about the oddly uncomfortable feeling of watching how someone 'mistreated' a robot. I don't even really have the language to describe it - if it was a person instead of a robot, then I'm spoilt for choice - "abusing", "bullying", "picking on", "antagonising", "tormenting", "persecuting"; all of these would fit. But none of them seem appropriate when you're talking about a 'soulless' robot.
I also wondered about what would happen when AI gets smart enough to watch YouTube videos and 'robots' actually see this kind of behaviour – and I added a related XKCD comic about fighting with machines (because there's always a related XKCD comic.)
This post is more about our broader relationship with 'machine intelligence' as we figure out what living with them looks like.
Right now, we've arguably got bigger issues to deal with - how we coexist with other people who might be from different countries, different colour skin, different religious backgrounds, different sexualities, different ideas about gender and identity, different physical or neural abilities etc. etc. I don't think you can fairly say that as a society - let alone as a species - we really treat all people like people, so perhaps its pointless to worry about whether we should be behaving in a kinder and more respectful manner towards software. (Meanwhile, humanity has apparently cracked the problem] of how to get intelligent creatures to breed in captivity, so that we can efficiently - if cruelly - kill them.)
That might be correct - to be honest, it probably is. But I still think there's a problem on the horizon, and its coming this way - fast.
In "Robot War part 2, I talked about a news story from early last year when Google engineer Blake Lemoine, working on the LaMDA large language model/chatbot, went public with his theory/belief that the software was sentient.
One of the things I found most interesting about the story was the way we treat, for want of a better word, 'non-people'; I think we find it hard to think of things like AI systems having 'sentience' or 'feelings' or 'wanting' things - but at the same time we have no problem with ideas like "Google doesn't like something" or "Apple wants something"; we recognise a company or corporation as a legal person, but (and I think this was the main point that Blake Lemoine was making) we aren't really prepared to think about 'machine intelligence' in the same way.
Maybe we should, so that we're ready for the day when 'machine intelligence' advances enough to say it has a soul/consciousness/sentience. (Or at least, enough to reject the null hypothesis that it definitely doesn't...) But perhaps its not really about the machines - perhaps its better for our souls that we acknowledge that kindness and respect should be our default position - and maybe that isn't just about machines, but also people, animals, maybe even the planet we live on...
I also talked a little about the long-standing fear of machines 'replacing' humans, and the idea that what might be more important than worrying about "them" (ie. how 'they' replace/take over 'us') might be to focus on "us" - that is, thinking about the way the tools we create go on to affect our own development. (For example, we might talk about things like cybernetics and the idea of having a computer implanted in your brain, but for decades we have outsourced a job that our brains used to do - remembering phone numbers - first to the computers in our pockets, and then to the computers in the cloud. My children, both born in a world where internet-connected devices are increasingly omnipresent, simply aren't going to develop mentally/cognitively/intellectually in the same way that I did in a pre-internet age.)
The race is on...
So... Here we are; March 2023, and in the last few months a lot has happened in the world of AI. I think the best place to start and set the context is this update from Ben Thompson;
This sounds hyperbolic, but I feel like I had the most surprising and mind-blowing computer experience of my life today. [...] Over the last 24 hours, though, I’ve come to believe that the entire focus on facts — including my Update yesterday — is missing the point.
This came on the back of four other Bing/AI-related posts; "New Bing, and an Interview with Kevin Scott and Sam Altman About the Microsoft-OpenAI Partnership" on 8th Feb (free to access) talks about the 'new Bing' announcement the day before; a new AI-chatbot-driven search engine. The next day, Google had their own announcement, which was covered in "Google and Microsoft’s Events, Monetizable Panic, Paradigms and Hardware" - 9th Feb (also free to access);
The event was, frankly, bad: the vast majority of content was a rehash of past Google I/Os, and one of the presenters even managed to forget to have a phone at hand for a demo; the new features that were announced would be available “in the coming months.” The slides seemed out of sync with presenters, as if they had just been prepared a few hours beforehand, and, well, it sure seems like the event and Sundar Pichai’s blog post introducing Bard (which you can’t yet use) were nothing more than a response to Microsoft’s Bing announcement.
The post is all about Microsoft and Google, search engines and business models, mobile versus PC. (All topics that I'm generally very interested in - but I'm skipping over here to focus on something bigger.)
The next update - "The Google CEO Question, Steve Ballmer and Peacetime CEOs, About That Bard Mistake", posted on 13th Feb (subscriber only) is again about the business; whether the AI-driven shifts that seem to be on the horizon warrant Google appointing a 'wartime' CEO as - very suddenly - the assumptions about Google's continuing dominance of the (massive) Search industry are looking questionable for the first time in... well, ever. It also talked about a mistake in Google Bard's demo about a space telescope which got a lot of attention - more attention than the fact that both Google and Bing 'traditional' search gives the same error.
The next day, ("New Bing Errors, User Preferences and Company Reputations, Section 230 and LLMs" - posted on 14th Feb (subscriber only) went into more details about mistakes made by the AI-powered Bing, as well as Google. It also touched on the 'legal person' issue; a platform like Facebook is has legal protection (from Section 230) from being held liable for posts made on its platform by "information content providers" (ie. people - at least, for now), and Google is not liable for information in its search results because that information comes from a distinct "information content provider". But for a Large Language Model (like AI-Bing/Sydney, Google Bard, ChatGPT), although trained on content from "information content providers", appears to be an "information content provider" itself - meaning that the company who 'owns' it is liable for its content. And this raises all kinds of issues around copyright, intellectual property etc. etc. - without necessarily needing to get into philosophical questions like "are they sentient?"
Don't get me wrong - I think all of this stuff is fascinating. But I'm also aware that its fascinating to people like me, who are interested in technology (I'm a geek), the tech industry (I'm a geek with a particular interest in the history and development of communication technology - which has been largely shaped by how its been commercialised) and the advertising industry that has largely funded it (I work in the advertising industry.)
But at the same time, I think there's something bigger going on right now. When Blake Lemoine was blowing the whistle last June about the possibility that Google LaMDA was sentient, nobody outside of Google had access to LaMDA; the story was about what someone inside Google thought about Google's secrets; it wasn't much of a surprise that the guy who went public didn't have a job there for long. https://www.theverge.com/2022/7/22/23274958/google-ai-engineer-blake-lemoine-chatbot-lamda-2-sentience;
So, it’s regrettable that despite lengthy engagement on this topic, Blake still chose to persistently violate clear employment and data security policies that include the need to safeguard product information. We will continue our careful development of language models, and we wish Blake well.
(Side note- companies always talk about themselves in the first person plural; an acknowledgement that 'the company' isn't really a single entity but a bunch of people. Maybe chatbots - neural networks - should do the same?)
When ChatGPT made a similar large language model available to the public in late 2022, it was deliberately limited; it doesn't have access to new information, so its effectively ignorant about the world after September 2021 (presumably thats when the training dataset was compiled and the model was built.) It also isn't allowed 'to express political opinions or engage in political activism' (source: BBC)
Nevertheless, when it was released it exploded; 100 million users in two months. (For context, Twitter, Facebook, WhatsApp and Instagram took years to get that many users; TikTok took 9 months - with a massive marketing push behind it.) Microsoft - who already had investment in OpenAI (the company behind ChatGPT) - apparently moved quickly to extend their partnership. And Google - by any accounts a leader in the development of AI (if you're learning how to actually build a neural network then you're very likely to be using Google's Tensorflow framework) seem to be at risk of getting left behind in the PR war. (Also Ben Thompson; "it’s not the greatest look for Google to have their “coming soon” announcements controlled by Microsoft’s PR schedule"
So; this is the tech war of the moment; a race to be the first to develop... something. Not just AI/Machine Learning/Deep Learning - but a product that will be a commercial success.
Why does that matter? Well, the thing about a (commercially) successful tech/data product is that it creates a virtuous cycle. Xerox built the first commercial computer with a GUI; Apple and Microsoft built the first commercially successful computers with GUI. For many reasons, Xerox haven't really been relevant in the field of computer innovation for decades - but not being able to come up with and develop original, innovative, valuable ideas isn't one of them…
Google was the best search engine, which meant that more people used it. That meant it had the best data on search terms, what people clicked on, what clicks led to people coming back and searching again - which meant it was the best data for improving how the search engine worked. Once that flywheel started spinning, nobody else had a chance of beating them in the Search game.
At least, not without a complete paradigm shift in what the "Search" product was; users searching for things, while advertisers pay to put their message in front of people searching for particular things (ie. wanting to be found enough to throw money at the problem.) What this new wave of technology does is - potentially - completely changes the paradigm. If what users want isn't to find a web page but to find information, then a search engine thats better at giving them the information they were looking for is likely to capture more searches; meaning more data.
For Google, thats an opportunity for incremental improvement of their search product - their business model means that they still need the step in the middle (the search results page), because thats the media space that their business is built on selling. But for a competitor - one with deep pockets, excellent tech resources and a strong desire to be seen as a leader in cloud computing/machine learning - its an opportunity to completely rethink what 'search' could mean. Because they can afford to break things that Google can't.
A race to... where?
The thing is, I'm pretty sure that this is a race. Google have a head start (at least, in the development of deep learning technology and a £175 billion Search industry) - Microsoft look like they could be set to catch up, and maybe even overtake Google; if Bing + ChatGPT + OpenAI technology can do a 'better' job.
But the thing about machine learning that makes it dificult isn't so much collecting masses of data (read: text, images, videos etc.) but having that data clearly and correctly categorised; the training data that tells the model whats 'good' and whats 'bad', which helps it progress towards its goal.
So the question is, if this is a race then where is the finish line? What is the ultimate goal? Is it about being the first to start spinning a sustainable/profitable business flywheel, where data processing delivers a service that generates revenues? Bearing in mind that, just because Google's Search generates revenues from advertisers, whatever consumer service takes its place might well generate its revenues in a different way; Google's head start in advertising might turn out to be a strategy tax in whatever the next phase looks like.
Suppose that the advertising model just doesn't work when it comes to a next-generation search engine; that we're actually looking at something more like a virtual assistant than a 'media' platform. Perhaps the 'winner' will be something that ends up tied to a particular product; a Google assistant that is exclusive to those who buy Google devices (not just the Android operating system), an Apple assistant that is exclusive to Apple devices - where does that leave Amazon and Microsoft, both of which have tried (and failed) to own their own phone platforms? Could a new 'assistant' app be good enough to break into the mature smartphone market? (Would Google/Apple allow it on their platforms? Are open-source operating systems good enough for it not to matter?) Might new partnerships be needed? (If Google went 'exclusive', where would that leave Samsung - for example? Might someone like Samsung end up building the 'exclusive' hardware platform for a Microsoft/Amazon software assistant?)
Or maybe its a more direct relationship with the consumer- how good would a next-generation search/assistant need to be to justify actually paying for it? And how much might people actually be willing to pay? Would it be in the order of a high-end cable TV subscription - say, £60 a month? Probably not... What about an SVOD subscription - something like £10-20 a month? Perhaps... what about something like a mobile game subscription - say, the £3 a month I pay for a persistent Minecraft world that my kids can play in? In the advertising industry, there's value in scale; reaching a large number of customers makes the product better. It seems like the same might be true of AI services - more users = more data - but perhaps the economy of the computing hardware it needs to run on change the equation. Maybe - like 'quality' journalism online - the difference between what paying customers will bear vs. what advertisers will fund shifts the balance of the economic model?
What does it mean if a sustainable economic model for an AI service relies on more revenue per user than the advertising industry will bear? "True" AI assistants for those who can afford it; Alexa for the rest?
For comparison, Google's ARPU in 2019 was estimated at $256 for the US and $137 per year, globally. In other words, if Google was paid for by users, they would need to pay about $21 a month (in the US) for Google to make the same sort of revenues - bearing in mind that fewer people would use it if they had to pay, it wouldn't cost much less to run, so either that price would quickly start to go up (so, the number of users would go down further), or Google's profit margin would quickly drop...
For now, the stakes are high; a $175 billion Search market dominated by Google, and - unthinkably just a few months ago - looking like it could genuinely be threatened by Bing. I don't know what the implications of a fundamental shift in the search engine market really means - but even a fraction of a $175 billion market moving is going to have enormous consequences, for the media industry if nothing else. Maybe this is the end of the advertiser-funded era of Search; maybe Google will just figure out a way to evolve their Search service. Maybe Microsoft will eschew the advertising model and roll it into their Microsoft subscription product. Maybe someone else will come along and offer something else entirely (a GPT-powered Siri could be massively more useful than its current iteration...) Then there are the second-order effects; who wants their writing to feed a large language model that they might be in direct competition with? (How do you stop it - other than take your writing off the open web?) Or the flip side; what does LLM-centric SEO work look like? Where will the $175 billion of advertising money go if Google becomes a less effective advertising channel?
The race is on...