Everybody in the Silicon Valley is crazy about chatbots lately.
ChatGPT has taken the world by storm with its ability to communicate effortlessly about anything - up to the point that it can generate content that can pass the final exam for some business and law schools.
In the meantime, some folks (including myself) believe that the current wave of AI, albeit surprisingly advanced in some applications, is still in the stochastic parrots phase rather than real artificial intelligence. That's just because of how these models are trained (build a network with billions of units, train it with the whole Internet, and see what comes out of it). Sure, they are very good at spotting and replicate statistical patterns, but they still need input from human moderation in the form of explicit, old-fashioned if-else-like rules to prevent their answers from going completely wild/immoral - and it's even possible to jailbreak them to bypass those restrictions.
Does Big Tech even care about AI?
Anyways, the fact that these models are still in a stochastic parrot stage doesn't seem to matter much for Big Tech. After all, their only purpose is to make their investors rich in the shortest possible time. Not to build a truly intelligent machine with human-like ethics and reasoning skills. That would be stuff that requires actual long-term R&D investments, and VCs nowadays aren't that patient - especially after many of them lost big money in the crypto scam, and they are craving for quick ways to recover from their losses.
No, Big Tech really doesn't care about AI.
Had they really cared about progress in AI, they wouldn't have fired or repurposed entire AI ethical teams for producing research that showed the problems with the AI strategy pursued by the top management.
And they wouldn't have spent the past decade iterating again and again over the same technology (neural networks) to grab all the possible low-hanging fruits, even when most of the academia has doubts about the fact that human-specific skills and behaviour (like ethics and logical reasoning) can really emerge from today's neural networks alone. Sure, they are cheap to scale, you add more hidden layers and feed them with more quality data and their performance improves like magic, and they are good at imitating the human brain in a specific set of activities (mostly those based on associative memory that is reinforced through repeated exposure to examples, mostly from sensorial input/output). But human intelligence isn't only about associative memory that learns to classify and predict things through repeated exposure. Otherwise we would only need the visual neurons in our occipital cortex to process optical signals from our eyes, and all the rest of the brain would be redundant. Unfortunately, research in the field has become dominated by hammer experts who think that the whole world is made of nails.
The past decade has seen many other fields in AI (like Bayesian techniques and symbolic reasoning) languish, as most of the AI conferences became dominated by a handful of large companies which are narrowly-focused on advancing research in the applications that are profitable for their business, rather than pursuing high-hanging fruits that would benefit the field as a whole. And, of course, since their business strategy is often at odds with what a wise and gradual development of such a sensitive field would require, discussions about ethical AI are often held at the fringe of the field, far from the places and the people that are actually forging the technology.
That, of course, is also very deliberate: in a highly competitive market you want to take everybody by surprise by showing people that you are one step ahead. That can only happen if you throw something as a beta to the world, and grab all the headlines before others do. AI ethical teams, peer reviews, or even just giving people the time to let the changes sink in and think over the possible consequences, are all things that stand in the way of being the first at everything, and make investors as happy as a bunch of kids in a candy store. Eventually, the development of technologies that can drastically change our societies is only up to the technical teams in charge of the implementation, the interaction with the scientific community revolves only around topics such as "is it better to use this or that model architecture for this class of problems?", and everything else is just an obstacle in the way of the business' vision. The strategy is akin to delivering to market racing cars without seat belts and airbags: of course somebody is likely to get hurt, but at least you'll have shown your investors that you can build a fast car before your competitors.
ChatGPT caused a "Code Red" at Google, not because of concerns about generative AI being suddenly deployed to everyone without much of a supervision or an ethical framework everybody agrees on. Their only concern was "how the hell did OpenAI manage to roll this out before us, and which corners can we cut to deploy a competitor as soon as possible and show our investors that we're still in the game?".
If companies like Google and Microsoft are going all big on AI lately, it's definitely because of reasons other than genuine interest in advancing the field academically.
All you need is hype
As mentioned earlier, many VCs have been heavily burned by the collapse of the crypto scam. The recent big layoffs in tech are, at least in part, motivated by investors tightening their pursue after losing a lot of money in a bad gamble.
Now that there aren't big profits left to make in crypto speculation, they need a new hype wave to get rich, preferably without waiting too long. And who cares if everybody is still far from building anything resembling a human (or even a more primitive) form of true intelligence. If it looks intelligent, then it is intelligent, and you can market it as AI. Something like ChatGPT was the rabbit that was pulled out of the cylinder at the right time.
But there's an even darker reason behind the current hype. That has to do with the real long-term mission of these tech giants.
The path to becoming the Alpha and the Omega of the Web
Both Google and Microsoft are focused on a specific application of conversational AI: search assistants. Turn their search engine experience into something more human, where you get answers by conversing with a bot, instead of scrolling through an endless list of links to other websites. This is not a coincidence.
Their real mission, at least for the past decade, has been to build a platform that isn't "just" a gateway to the Web. It should be THE Web. Google's initial mission was to minimize the time that users spent on their website - you find what you're looking for as fast as possible, you click on it, and then you're out. Their initial success metric was simple: provide relevant results for the user at the top of the page, so users don't have to spend time scrolling and clicking around. The shorter the time the user spends on the website before clicking on what they're looking for, the better.
Over the past 10 years, that mission has been flipped on its head: now Google wants to be the beginning and the end of your Internet journey.
If you can get all the answers you want directly from your search engine, then you will never need to open any another website. All the efforts on Google's side (from going all-in with voice assistants, to adding related questions/answers on the search results, to more and more information provided in inline boxes on the search page, to scraping and showing lyrics directly in the search engine, to showing rates for flights and hotels directly in the results) have gone in that direction. From Google's perspective, a user that spends more time on the website (because all the information that they need is already there) is much more profitable than a user that uses them just as a gateway to the Web - even if that's supposed to be the whole purpose of a search engine.
Silicon Valley isn't that explicit about it (well, except for Elon Musk of course), but the dream of the giants is still to build THE ultimate platform (the Chinese way) where your Internet experience starts and ends. "Everything" platforms like Tencent and WeChat are actually a source of inspiration in the valley. AI bots are just a nice shortcut to get to that end. Why would you want to visit the websites of several outlets to stay up-to-date with the news about an event, when you can just ask a question to your search assistant, and it will provide you with everything it digested on that topic from hundreds or thousands of articles? The AI already read them all for you, so you won't have any of them!
Have all the cakes and eat them
What kind of world would this be on the long run? I've given it a thought lately, and I think that it's world that will eventually leave everybody worse off on the long run - including the tech giants themselves.
Think of it for a moment. Models like ChatGPT exist thanks to the Web. Thanks to a network of billions of websites that they can crawl and scrape, and that content is eventually pre-digested into something that can be fed to humongous models.
The relation is also very asymmetric: the whole business model of these platforms is based on scraping the Web, but they'll do their best to prevent you from scraping THEIR platforms (and, if you manage to do so, they will often take you to court).
If, for most of the users, the Internet journey starts and ends on the search engine, and there's no way of making the information from the search engine "trickle down" to other websites and platforms, then all the other websites will starve off.
News outlets today complain for Google News putting a summary of their articles in their feed (yet with the original link still attached)? Then imagine a world where you just ask a bot about the news on a specific topic, and the bot answers you with what it digested from other news outlets - no links attached.
Not only: since you spend a lot of time on the search engine, the search engine, unlike the Guardian or the NYT, also knows a lot about you. It can provide you with a personalized experience based on the content you're most likely interested in, from the ideological perspective you're more likely to lean on, thereby amplifying the creation of ideological bubbles. What's worse is that, unlike articles written by journalists or other specialists (where the same information is publicly available to everyone for scrutiny), customized answers from a search assistant aren't subject to any form of external scrutiny, nor are accountable for accuracy, since they are just passing around content digested from somewhere else.
At some point, what will be the point of investing time and money into building your website or platform, when people already have the "everything app"? Publishing e.g. a blog, or some technical documentation, or a news outlet, or building an e-commerce platform with an open API, means cooking some food that these large models can feast on. They will digest all of your content, summarize it, spit it out to users, and you may rarely see a single organic visit.
If other websites eventually die off, these AI models will have less and less content to be trained on, unless the companies behind them also take on the organic business of the websites that they are replacing instead of just scraping them (something not very likely to happen, given their resistance to scaling in fields that require hiring more humans).
Eventually, the performance of their models will degrade and they will start to provide outdated information.
I'm pretty sure that there are smart people at these companies who have reached the same conclusions. And yet they are still pushing at breakneck speed with this strategy, because short-term returns are much more important than the long-term damage inflicted to everyone (including themselves).
Search assistants are a flawed version of what the semantic Web could have been
What infurates me the most, however, is that these businesses aren't inventing anything new.
We already envisioned something similar to what these companies are working on - a "Web of meaning" and connected information that can easily be parsed by machines as well as human eyes. But it was better, less costly to run, open and decentralized.
Actually, somebody already envisioned it more than 20 years ago - I mean, Tim Berners-Lee wrote an article about it already in 2001. It was called "semantic Web" - a.k.a. the "real" Web 3.0.
The idea is based on a curated layer of "meaning" built on top of the Web. A Web page shouldn't be only HTML that can be rendered on a screen. A Web page actually contains information, and, from the perspective of a machine, that information is best digested when provided in a structured format.
Consider this simple HTML snippet for example:
<div> Alan Turing was born in Maida Vale, London, on June 23rd, 1912. </div>
What can a machine infer from this snippet? Well, actually not much - unless it scrapes text from the HTML, discards all the tags that aren't relevant, filters out all the text from menus, ads etc., and it feeds the extracted text to a language model that converts it to a structured representation. It's a snippet that is meant to be rendered by a browser as text that obeys specific style rules. It's a presentation layer for the human eye: the machine is just an intermediary.
What if, however, you were writing HTML like this?
<div vocab="https://schema.org/" typeof="Person"> <span property="name">Alan Turing</span> was born in <span property="birthPlace">Maida Vale, London</span> on <span property="birthDate">June 23rd, 1912</span>. </div>
The HTML would still be perfectly valid and rendered in the browser the same way. However, just adding a schema/vocabulary makes it easy to model an ontology on top of HTML that can be also understood by a machine. Now a machine must no longer rely on complex scrapers and NLP models to get the meaning of a Web page. It just needs to go through the tags, extract the schemas, read the properties and the predicates, and it can already provide you a summary of the page worth of ChatGPT.
This is exactly what these language models, at the end of the day, do: extract meaning out of natural language, and generate content in natural language. But, instead of having grammar and semantics baked into the markup of the Web, they are scraping whatever they can find on the web, and basically "brute-force" the meaning out of them through expensive, huge statistical models. It's like looking for somebody's house in a city that you don't know, and, instead of trying to get the address and putting your hands on a map, you tried to walk all the roads and ring all the doorbells until you find the right one. Or putting thousands of monkeys in a room filled with typewriters, and wait until one of them ends up replicating a Shakespeare sonnet. You want to use the right tools for the right purpose, and stochastic methods just aren't the most efficient solution for all the problems.
Think of it from a moment. In the annotated HTML snippet above, extracting meaning from a page and summarizing it doesn't require a big model with billions of units, trained with billions of documents on machines collectively worth millions of dollars (with all the connected environmental concerns). A simple DOM parser written by a junior developer would suffice to extract all the information you need from any website. Anybody can do it, anybody could build their own search engine, anybody could make their own network of crawlers that extracts clean information from web pages. Sure, Web creators will also have to be in charge of modelling the information in their websites in a structured format. But, if that was the requirement for the content on their websites to be indexed and searchable, would it be much different from the craze they already go through for SEO optimization? Or would it be much different from an Instagram influencer curating their posts with lists of hashtags so they can be easily searched?
So, if this idea was so amazing, if it's been pushed by the inventor of the Web himself, how come we haven't embraced it yet?
Well, exactly because it would empower anybody to build algorithms that could extract meaning from the Web!
Data is the new oil, and large companies don't want a world where anybody can extract it. Such a world would dilute their value proposition. Technologies that lower the barriers to access structured information, like the semantic Web, are existential threats to their business model. In a world where anybody can build a crawler that extracts clean information from the Web, products like Google search would either have a fierce competition (and monopolies hate competition), or no reason to exist at all.
So their strategy can be summarized as it follows:
Keep entry barriers as high as possible, so that the only competitors who can access the market are those who can invest millions/billions on language models - and there aren't many of them around.
Use your position of advantage to scrape the whole Web and feed it to a huge model, while discouraging other people from scraping.
Teach your model how to reproduce/mimic whatever it has learned.
Your model becomes the last Web you will ever need.