Main Edition
US
FR
China has limitless data to feed AI Chatbots. Kevin Frayer/Getty
Can anyone compete with China’s Artificial Intelligence super-system? Sleepy government bureaucracies the world over are finally waking up to the hard reality that they have virtually no chance. China is galloping ahead. Only last month, it unveiled its latest rival to San Francisco’s ChatGPT: the Moss bot, and this month it plans to release another. The UK lags far behind.
Tony Blair thinks Britain should put itself on an economic war footing and pour national resources into the creation of an AI framework that might compete with China’s. But it’s hard to see how that is possible — or even desirable.
In large part, this is because AI needs data to work. Lots and lots of data. By feeding huge amounts of information to AIs, deep learning trains them to find correlations between data points that can produce a desired outcome. As deep learning improves the AI, it requires more data, which creates more learning, which requires more data.
While many nations might struggle to cope with AI’s insatiable demand for data, China is in no short supply. Since the nation fully came online around the turn of the millennium, it has been steadily carving out a surveillance state by acquiring endless amounts of data on its population. This initiative has roots in China’s One Child Policy: this impetus for controlling the population in aggregate — that is, on a demographic level — devolved into a need to control the population on an individual level.
This became fully apparent in 1997, when China introduced its first laws addressing “cyber crimes”, and continued into the early 2000s as the CCP began building the Great Firewall to control what its citizens could access online. Its guiding principle was expressed in an aphorism of former-Chinese leader Deng Xiaoping: “If you open the window for fresh air, you have to expect some flies to blow in.” The Great Firewall was a way of keeping out the flies.
China has always had a broad definition of “flies”. In 2017, the Chinese region of Xinjiang, home to the Uighur minority, rolled out the country’s first iris database containing the biometric identification of 30 million people. This was part of a larger effort known as part of a wider Strike Hard Campaign, an effort to bring the Uighur population under control, an effort to bring the Uighur population under control by using anti-terror tactics, rhetoric and surveillance.
Join the discussion
Join like minded readers that support our journalism by becoming a paid subscriber
To join the discussion in the comments, become a paid subscriber.
Join like minded readers that support our journalism, read unlimited articles and enjoy other subscriber-only benefits.
SubscribeAt risk of doxing myself I work I work in the realm of the fourth industrial revolution, although overwhelmingly in the industrial application of this technology and not the social and political applications of it. The consequences of the emergent technologies are enormous, and although I’d rather not put time lines to the development and consequences of deployed technologies, it is clear their impacts will be significant across all domains of life.
AI is not my computing and software specialty but it is beginning to surprise: both in terms of it’s development but also how stupid it can be at times (unless it’s lulling us into a false sense of security). What bothers me is how innovation typically works, which is often about how existing technologies are merged to create something new.
It isn’t necessarily difficult to predict that advanced machinery/robotics + AI renders vast quantities of jobs redundant. As soon as this is cheaper than the cheapest labour in a particular location, it will become the chosen approach for corporate profiteering over globalisation. It could also have benefits for supply chains (and therefore environmental impacts) by allowing factories and plant to be near the source of raw materials. Science fiction writers and futorologists have been predicting this for decades.
The use of AI as a social control system is a little less discussed but not unheard of. However it is clearly the innovation of merging social media, public record and AI. Introducing other 4th IR concepts like digital currencies (central bank or otherwise) it isn’t exactly hard to see where social credit systems come into play. Considering the prior point regarding robots taking people’s jobs, it becomes essential (at least from a political point of view).
The welfare state, which was vastly expanded following skilled jobs being off-shored would need to expand again. In a world where an individual citizen/consumer has become decoupled from production, how exactly will their spending power be governed? As such, I see a certain inevitability as the consequence of automating jobs away will demand this form of response.
Yes, my outlook is dystopian, Huxley would be proud. The more “utopian” vision for an automated society, such as Jacque Fresco’s removes too much power from government, supranational and corporate interests and I see no evidence that those groups would ever surrender control. We will once again see a merger of corporate and political interest: one of Tony Blair’s favourite things.
At risk of doxing myself I work I work in the realm of the fourth industrial revolution, although overwhelmingly in the industrial application of this technology and not the social and political applications of it. The consequences of the emergent technologies are enormous, and although I’d rather not put time lines to the development and consequences of deployed technologies, it is clear their impacts will be significant across all domains of life.
AI is not my computing and software specialty but it is beginning to surprise: both in terms of it’s development but also how stupid it can be at times (unless it’s lulling us into a false sense of security). What bothers me is how innovation typically works, which is often about how existing technologies are merged to create something new.
It isn’t necessarily difficult to predict that advanced machinery/robotics + AI renders vast quantities of jobs redundant. As soon as this is cheaper than the cheapest labour in a particular location, it will become the chosen approach for corporate profiteering over globalisation. It could also have benefits for supply chains (and therefore environmental impacts) by allowing factories and plant to be near the source of raw materials. Science fiction writers and futorologists have been predicting this for decades.
The use of AI as a social control system is a little less discussed but not unheard of. However it is clearly the innovation of merging social media, public record and AI. Introducing other 4th IR concepts like digital currencies (central bank or otherwise) it isn’t exactly hard to see where social credit systems come into play. Considering the prior point regarding robots taking people’s jobs, it becomes essential (at least from a political point of view).
The welfare state, which was vastly expanded following skilled jobs being off-shored would need to expand again. In a world where an individual citizen/consumer has become decoupled from production, how exactly will their spending power be governed? As such, I see a certain inevitability as the consequence of automating jobs away will demand this form of response.
Yes, my outlook is dystopian, Huxley would be proud. The more “utopian” vision for an automated society, such as Jacque Fresco’s removes too much power from government, supranational and corporate interests and I see no evidence that those groups would ever surrender control. We will once again see a merger of corporate and political interest: one of Tony Blair’s favourite things.
Yeah nice essay, too bad it’s totally wrong. I got nothing against journalists and writers opining on technical subjects but they should at least try to find someone with knowledge to check their thesis before they embark on writing it. No Kai-Fu Lee doesn’t count, as his agenda these days is to make China look powerful first and be correct second.
To train LLMs like ChatGPT or “MossBot” you need lots of data, but it’s the sort of data you get from the public internet, including possibly books/magazines/newspapers/etc. It isn’t surveillance data of the sort this article is talking about. You got a billion iris images in a database? Good for you, that’ll be super helpful if you want to train an AI that can generate perfect looking random iris images and not much else. You got a billion swabs? Great, now you can generate fake swabs.
Getting the picture here? The reason you need a lot of data scraped from the internet to train a modern AI is because you want that AI to generate the sorts of things you find on the internet: answers to questions, news articles, photos, poetry, code.
So data isn’t the new oil, it’s not like anything even close to oil. Oil is fungible, one barrel is much like another. Data is not, which is why forced analogies like “The Saudi Arabia of data” are the mark of punditry, not expertise (not that you need much to know this stuff!).
Could the British government train a giant LLM if it wanted to? Uh, yes? DeepMind is based in London and has done exactly that, it’s just a matter of offering those people 2x what they currently get paid and then giving them the time and money they need to collect lots of web crawls, book scans and so on + a few tens of millions of dollars worth of hardware from NVIDIA. But why would it? Maybe people like Blair think there’s something strategic about all this but there isn’t. What does he even mean by an AI Framework? A British TensorFlow? Surely not.
I’m a member of the tech community, read AI research papers regularly, take part in AI discussions regularly and have never heard anyone say this. What would that even mean? AI development is the opposite of communist, the most advanced AIs are all being trained by private corporations and the most talked-about AI right now (ChatGPT) is by a well funded startup. There’s nothing I can think of that’s less communist than a startup.
Yeah great observation, I was thinking the same thing as I was reading the article. The Chinese data referred to is constrained, controlled and influenced by policy. MossBot probably gives great answers to questions posed by the CCP.
I totally concur with the points you make, but I would not be quite so dismissive of surveillance data, etc. On its own, surveillance data such as iris recognition, may be of limited value beyond those directly involved in surveillance, but when combined with other data about the individuals its potential uses (and of course abuses) are much greater in scope.
Sure, but then you’re using AI to process surveillance data, not using that data to create AI. LLMs and other modern AI tech can certainly be used to build a dystopia, no doubt about it, but China doesn’t seem to have any particular advantage in building such tech beyond a desire to do so and lots of smart tech-savvy citizens.
Fair point!
Fair point!
Sure, but then you’re using AI to process surveillance data, not using that data to create AI. LLMs and other modern AI tech can certainly be used to build a dystopia, no doubt about it, but China doesn’t seem to have any particular advantage in building such tech beyond a desire to do so and lots of smart tech-savvy citizens.
Blair was a big proponent of the knowledge based economy when he was PM. He wasn’t quite so big on explaining what it was though.
Yeah great observation, I was thinking the same thing as I was reading the article. The Chinese data referred to is constrained, controlled and influenced by policy. MossBot probably gives great answers to questions posed by the CCP.
I totally concur with the points you make, but I would not be quite so dismissive of surveillance data, etc. On its own, surveillance data such as iris recognition, may be of limited value beyond those directly involved in surveillance, but when combined with other data about the individuals its potential uses (and of course abuses) are much greater in scope.
Blair was a big proponent of the knowledge based economy when he was PM. He wasn’t quite so big on explaining what it was though.
Yeah nice essay, too bad it’s totally wrong. I got nothing against journalists and writers opining on technical subjects but they should at least try to find someone with knowledge to check their thesis before they embark on writing it. No Kai-Fu Lee doesn’t count, as his agenda these days is to make China look powerful first and be correct second.
To train LLMs like ChatGPT or “MossBot” you need lots of data, but it’s the sort of data you get from the public internet, including possibly books/magazines/newspapers/etc. It isn’t surveillance data of the sort this article is talking about. You got a billion iris images in a database? Good for you, that’ll be super helpful if you want to train an AI that can generate perfect looking random iris images and not much else. You got a billion swabs? Great, now you can generate fake swabs.
Getting the picture here? The reason you need a lot of data scraped from the internet to train a modern AI is because you want that AI to generate the sorts of things you find on the internet: answers to questions, news articles, photos, poetry, code.
So data isn’t the new oil, it’s not like anything even close to oil. Oil is fungible, one barrel is much like another. Data is not, which is why forced analogies like “The Saudi Arabia of data” are the mark of punditry, not expertise (not that you need much to know this stuff!).
Could the British government train a giant LLM if it wanted to? Uh, yes? DeepMind is based in London and has done exactly that, it’s just a matter of offering those people 2x what they currently get paid and then giving them the time and money they need to collect lots of web crawls, book scans and so on + a few tens of millions of dollars worth of hardware from NVIDIA. But why would it? Maybe people like Blair think there’s something strategic about all this but there isn’t. What does he even mean by an AI Framework? A British TensorFlow? Surely not.
I’m a member of the tech community, read AI research papers regularly, take part in AI discussions regularly and have never heard anyone say this. What would that even mean? AI development is the opposite of communist, the most advanced AIs are all being trained by private corporations and the most talked-about AI right now (ChatGPT) is by a well funded startup. There’s nothing I can think of that’s less communist than a startup.
Feminists ridicule the verbiage of us males as “manplanation”. I wonder how long it will be before “botplanation” enters the vocabulary. Sometimes, chatGPT raises its hands in the air and admits its just a piece of software, but more often, its responses are a mixture a fact and bu115h1t. As with any interlocutor, human or otherwise, who breezily answer a question when they don’t really know what they are talking about, one quickly learns to ignore them.
So if I instructed chatGPT to “Explain why President Xi’s Zero Covid Policy was such a monumental disaster.” In about two seconds I get a very good, balanced response with four headings: Economics, Human Rights, Long-term harm and Lack of transparency. Great! What do you suppose Moss would do if I asked it for the same information? So the two tools have a distinct purpose. And god help us if a latter-day Dominic Cummings bases policy on the strnegth of a chat with a Bot.
An ex-colleague, a professor of Artificial Intelligence, once told me, researchers only call it artificial intelligence when we don’t really understand why it works. As soon as we DO understand why it works, we start calling it software.
When I was student, in the immediate run up to the Iraq war, our Artificial Intelligence and Artificial Neural Networks professor announced he wouldn’t be available for a couple of weeks due to meetings with the government. A friend of mine quipped they’d developed an expert system to decide whether to invade or not. In the next lecture after his return, he made a point about expert systems being used in the decision process for invading Iraq. I do wonder if he heard my spit take.
When I was student, in the immediate run up to the Iraq war, our Artificial Intelligence and Artificial Neural Networks professor announced he wouldn’t be available for a couple of weeks due to meetings with the government. A friend of mine quipped they’d developed an expert system to decide whether to invade or not. In the next lecture after his return, he made a point about expert systems being used in the decision process for invading Iraq. I do wonder if he heard my spit take.
Feminists ridicule the verbiage of us males as “manplanation”. I wonder how long it will be before “botplanation” enters the vocabulary. Sometimes, chatGPT raises its hands in the air and admits its just a piece of software, but more often, its responses are a mixture a fact and bu115h1t. As with any interlocutor, human or otherwise, who breezily answer a question when they don’t really know what they are talking about, one quickly learns to ignore them.
So if I instructed chatGPT to “Explain why President Xi’s Zero Covid Policy was such a monumental disaster.” In about two seconds I get a very good, balanced response with four headings: Economics, Human Rights, Long-term harm and Lack of transparency. Great! What do you suppose Moss would do if I asked it for the same information? So the two tools have a distinct purpose. And god help us if a latter-day Dominic Cummings bases policy on the strnegth of a chat with a Bot.
An ex-colleague, a professor of Artificial Intelligence, once told me, researchers only call it artificial intelligence when we don’t really understand why it works. As soon as we DO understand why it works, we start calling it software.
Social credit scores coming to a postcode near you? Seriously though, the UK government can’t even produce a database of NHS medical records. How on earth would they manage a project like this? It won’t stop them trying though. Management consultants are already rubbing their hands at the prospect of the coming bonanza.
An excellent observation about the pitiful use of IT within the NHS. I had dealings with IT people drafted into our health service over the last two decades of my NHS career (up to 2016) and let’s just say the type of people recruited were pretty third rate. Why? Quite simply because anyone with any real IT talent could earn far more money in the private sector.
On the national level, i attended conferences in the early 2000s about the introduction of a UK database for medical records. I expect such conferences are still being attended two decades later, with the same blather and the same costs in attending.
To try to bridge the talent gap, the NHS employs IT Consultants at eye-watering rates. They come in, do their thing (with the tech available at the time) and disappear, leaving the system they’ve helped introduce to become a “legacy” within the space of a few short years since no-one remaining in the organisation understands it, nor they can change it. More efficient tech overtakes the legacy system and none of the disparate systems “talk” to each other. Repeat ad infinitum.
If this were to be the type of template for AI at the state level, it’d be a huge waste of time and money. However… i think what’s being suggested is something of a different order. In the NHS, the systems rely on overstretched staff inputting data (e.g. drug regime, changes to regime, drug administered, if not why not etc.) which simply can’t happen automatically. Errors creep in all the time which stymies the system. In an environment where data is routinely collated via an automation process (as with smartphones), there’s a different type of potential. There’s no reason in theory why the UK couldn’t replicate what’s happening in China, except that state control and citizen consent are of a different order. If this were to happen by stealth, i.e. without citizen consent (and it may already be happening) then we’re in a completely new ballpark. This article is very welcome as a warning of the double-edged sword that’s hanging over us all. We’re all Damocles now.
You are correct, it is already happening by stealth e.g. centralised ID database for access to all government services, BOE preparing to introduce a CBDC, handover of power to the WHO to control the UK population in the event of a new pandemic (see Unherd’s excellent piece today on this subject). None of this to my knowledge has even been debated in Parliament, let alone put to the British people.
Back on the subject of the NHS, my brother-in-law had a sinecure for many years working for one of the major consultancies, going round the country managing IT projects for NHS trusts. Exactly as you described, finish project, move on, leave the poorly equipped trusts to deal with the fallout, system falls into oblivion within a few years.
You are correct, it is already happening by stealth e.g. centralised ID database for access to all government services, BOE preparing to introduce a CBDC, handover of power to the WHO to control the UK population in the event of a new pandemic (see Unherd’s excellent piece today on this subject). None of this to my knowledge has even been debated in Parliament, let alone put to the British people.
Back on the subject of the NHS, my brother-in-law had a sinecure for many years working for one of the major consultancies, going round the country managing IT projects for NHS trusts. Exactly as you described, finish project, move on, leave the poorly equipped trusts to deal with the fallout, system falls into oblivion within a few years.
An excellent observation about the pitiful use of IT within the NHS. I had dealings with IT people drafted into our health service over the last two decades of my NHS career (up to 2016) and let’s just say the type of people recruited were pretty third rate. Why? Quite simply because anyone with any real IT talent could earn far more money in the private sector.
On the national level, i attended conferences in the early 2000s about the introduction of a UK database for medical records. I expect such conferences are still being attended two decades later, with the same blather and the same costs in attending.
To try to bridge the talent gap, the NHS employs IT Consultants at eye-watering rates. They come in, do their thing (with the tech available at the time) and disappear, leaving the system they’ve helped introduce to become a “legacy” within the space of a few short years since no-one remaining in the organisation understands it, nor they can change it. More efficient tech overtakes the legacy system and none of the disparate systems “talk” to each other. Repeat ad infinitum.
If this were to be the type of template for AI at the state level, it’d be a huge waste of time and money. However… i think what’s being suggested is something of a different order. In the NHS, the systems rely on overstretched staff inputting data (e.g. drug regime, changes to regime, drug administered, if not why not etc.) which simply can’t happen automatically. Errors creep in all the time which stymies the system. In an environment where data is routinely collated via an automation process (as with smartphones), there’s a different type of potential. There’s no reason in theory why the UK couldn’t replicate what’s happening in China, except that state control and citizen consent are of a different order. If this were to happen by stealth, i.e. without citizen consent (and it may already be happening) then we’re in a completely new ballpark. This article is very welcome as a warning of the double-edged sword that’s hanging over us all. We’re all Damocles now.
Social credit scores coming to a postcode near you? Seriously though, the UK government can’t even produce a database of NHS medical records. How on earth would they manage a project like this? It won’t stop them trying though. Management consultants are already rubbing their hands at the prospect of the coming bonanza.
Not all data has the same value. China certainly is monitoring and recording more conversations than any other entity (except maybe the NSA?) and this provides better training for AI based on LLMs like Chat GPT. Similarly for number of images of people and thus people & image recognizing AIs.
But as numerous others more knowledgeable than me have demonstrated, the current deep learning models have little true understanding and are far from a general AI. See Gary Marcus’ excellent Substack on AI
https://garymarcus.substack.com/p/smells-a-little-bit-like-ai-winter
It is not clear that the current dominant models will be the path towards AGI and that achieving AGI is just about more processors and more data.
Not all data has the same value. China certainly is monitoring and recording more conversations than any other entity (except maybe the NSA?) and this provides better training for AI based on LLMs like Chat GPT. Similarly for number of images of people and thus people & image recognizing AIs.
But as numerous others more knowledgeable than me have demonstrated, the current deep learning models have little true understanding and are far from a general AI. See Gary Marcus’ excellent Substack on AI
https://garymarcus.substack.com/p/smells-a-little-bit-like-ai-winter
It is not clear that the current dominant models will be the path towards AGI and that achieving AGI is just about more processors and more data.
I have no idea what chatbot CBT is?
Just pray you never need to find out.
With an ordinary browser, you type in some keywords and it gives you a list of web pages that contain the keywords. With a chatbot, such as chat GPT, you can (a) ask a question and you get an essay-style answer or (b) engage in a dialogue with the chatbot. the essay-style responses are produced remarkably quickly and are reasonably well structured. but for me, the big problem is that the chatbot only occasionally admits that it is out of its depth. It can produce nonsensical answers, sometimes becuase it has not be trained on the appropriate data.
No
No