by Brendan Cooney
While the technology loosely referred to as “artificial intelligence” (AI) has been increasingly part of our lives for the last decade, recent developments in image/video/audio generative AI and large language models (LLMs) have ushered in a new age consisting of rapid development, huge capital investments, and fierce competition between firms competing to develop and sell AI.
In recent weeks, many prominent AI researchers have begun to speak out against the rapid development of AI technology, warning that the pace of development far outpaces our ability to avert possible disaster scenarios. These warnings are amazing to hear coming from folks like Geoffrey Hinton, considered by many to be the godfather of AI research. But it was perhaps inevitable that even those most familiar with the development of the technology would be some of the first to realize that a threshold had been crossed. The stated goal of most AI researchers and firms is to achieve what they call Artificial General Intelligence, or AGI, an intelligence equal to or surpassing human intelligence. It was always only a matter of time before the development of AI research reached a threshold at which the prospect of further development caused alarm and dissention in the field. Now that capitalists and capitalist states have gotten a glimpse of the potential power of the new AI, the race is on, with every large player in the tech world jumping into the ring to develop new LLMs, new uses of chatbots, new chips for AI, new data sets for training AI, etc.
Most AI researchers agree that the new wave of generative AI does not display AGI. But the new wave of generative AI (sophisticated machine learning neural networks that can generate images from text prompts, large language models that can generate impressive text from prompts, etc.) represents a qualitatively new development in the technology. Without venturing into the debate about how these machines generate their output (Is it merely predictive algorithms? Is there some sort of primitive reasoning going on? Is it a black box?) it is clear that the output of these machines can easily be mistaken for the product of human thought. Image generating AI like Stable Diffusion can generate impressive images from simple text prompts, images that can easily pass for human generated images. LLMs like OpenAI can produce text and have conversations in a remarkably human-like fashion. This capacity to closely mimic human thinking creates the opportunity for a great deal of dystopian uses, from sophisticated disinformation campaigns, to opportunistic intellectual theft, to new forms of cybercrime, to labor displacement, etc.
What are the main aspects that determine the character of AI today and its direction and speed of growth?
Data
AI today is based on machine learning in which large sets of training data serve as inputs into artificial neural networks. Machines teach themselves on this training data, learning to predict text, images, etc. from prompts. For many years in which researchers sought improvements in machine learning by tweaking the models the machines learned by, development was slow. In the past few years, these basic models have become adequate enough that researchers began to understand that they could get better results by just scaling up the amount of training data fed into the machines. The bigger the learning data set, the more “intelligent” the AI.
Thus, the major advances behind phenomenon like ChatGPT and Bing are fueled by two things: data and computation power. In other words, AI today needs huge amounts of data and large data processing power. Both the need for data and the need for computing power shape the current development of the field.
The internet is full of data, much of which is useful for training AI, like the LLM used by ChatGPT. However, now that the development of AI has become a battleground in capitalist competition, we are seeing attempts to limit the ability of AI to use the internet as a free source of training data. Legacy media, Reddit, owners of images, Twitter, and others have suggested that AI developers like OpenAI should pay a licensing fee to train on their data.
Also, generative AI has the ability to train on and emulate the unique, personal styles of visual artists, musicians, writers, etc. This has created the possibility for AI to replace specific creative individuals. Already, artists are having to compete in the market with AI versions of themselves. Many worry about an existential threat to creative professions, a class of work that is already notoriously precarious. There have been attempts to defend against the theft of personal artistic identity through various methods, such as digital water marks that taint art so that it can’t be used in training data. But this sort of response seems like the sort of challenge that AI developers could easily overcome with enough time and effort. And it doesn’t address the larger issue of how this technology will undermine creative fields in general, not just the personal careers of well-known creatives.
Elon Musk on toilet. MHI, generated by Stable Diffusion.
Compute
When it comes to computing power, the more computations one runs, the more resources one consumes. Similar to crypto mining, training and using LLMs consumes a lot of electricity and water. ChatGPT consumed as much electricity as 175,000 people in January of 2023. Hence, the development of this technology has only been possible for large companies with a lot of capital to invest. The hype around generative AI is such that the big players in the field are seeing a lot of investment, spending a lot on scaling up development, and offering big salaries to the best engineers. In many respects, it feels like similar moments in tech history when hype sustained the boom times long before the technology was profitable.
Most significant advances in the power of generative AI have come from scaling up the amount of training data and computing power used. Thus, every improvement in the model requires greater inputs of data and energy. Researchers in the field are beginning to see diminishing returns on this scaling. Minor improvements in the reliability of a LLM require large increases in data and computation, with attending monetary and environmental costs. These diminishing returns are an important aspect of the current evolution of the industry. It may even be the case that OpenAI’s rush to put ChatGPT on the market is a tactic to compensate for these diminishing returns by creating hype to attract investors. Perhaps OpenAI realized they were encountering a barrier and rushed to dump their product on the market to distract investors from the fact that their approach to scaling up was hitting the law of diminishing returns.
There are several approaches being proposed to this new problem of diminishing returns on scaling. One is to prune data sets so that they take less time to compute. By getting rid of redundant or useless data, the theory goes, one can save computing resources. Another idea is to use large AI models to train smaller models. This creates smaller, more efficient models that can do the same work but require fewer resources. It remains to be seen how successful these methods will be at solving the problem of diminishing returns from scaling.
Open Source and Moats
Attempts to make smaller, more efficient AI models have created another issue with immediate ramification for competition and proliferation. In recent months, we have seen the proliferation of open-source AI models. Most of these came with the leak of Meta’s AI model named Llama. (Meta released the model open source but did not release the weights. Those were leaked shortly after Meta released Llama.) The release of Llama ushered in a frenzy of experimentation, resulting in many people being able to get AI LLMs and other generative AI models running on home computers. Thus far, these open source Llama models aren’t quite as powerful or fast as large models like GPT4, but they are very close. This has created a problem for the big players in the industry like Google and OpenAI, who could lose out big if the millions they invested in developing LLMs goes down the drain because people can use open source LLMs for free or cheaply. These fears were summed up in early May in a leaked internal memo from an anonymous author at Google, entitled “We Have No Moat and Neither Does OpenAI”. The author argues that with the proliferation of open source AI, the big players no longer have a “moat” by which to retain their market dominance.
It is not clear how this will play out. What we do know is that the valuable things to control in this industry are data and computing power (including access to chips and storage capacity). Google likely owns more data than any other company in the world. They will have a moat for the foreseeable future. And big developments in AI will likely still require more computing power than individuals or small firms can afford, so there will likely continue to be a place for big tech. However, it is not hard to imagine that the spectacle of OpenAi’s chief executive Sam Altman appearing before Congress on May 16 to urge Congress to regulate the industry, including granting licenses for companies to develop AI, was a direct response to the problem of the proliferation of cheap open source alternatives to ChatGPT. If OpenAi can work with Congress to craft legislation in a way that assures their market dominance (for instance, by issuing expensive licenses for AI developers so that small-time open source developers are kept out of the market), and also gives the world the impression that the company takes AI safety seriously, that could be a big win for OpenAI.
Regardless of how this competition plays out, it seems, for now, that access to or ownership of machine learning algorithms will not be a significant moat. Instead, competition over access to and ownership of data, chips, storage and computing power will be the main issues.
Alignment
The hype around AI is so big because AI has the potential to significantly disrupt almost every aspect of human life. Most technologies are use-specific. Coal makes energy, cars move people, etc. AI is potentially more disruptive a technology than even coal or cars because it is not use-specific. AI can potentially be applied to any activity one can think of, or that the AI thinks of. It is thus very powerful and very dangerous.
But hype won’t be enough to sustain the industry. Investors and clients want to know that the AI they are buying or investing in is good at what it does. Thus, one of the main issues facing AI today involves the quality of AI performance.
Clearly, chatbots like ChatGPT are truly remarkable in their ability to produce language. It is a jaw-dropping achievement. But these chatbots are currently far better at generating language than they are at making true statements. While chatbots can pass the bar exam, they struggle with basic logic. They are horrible at math. They do poorly when executing a series of tasks. They often make up facts and stubbornly resist being corrected. For better or worse these defects are called “hallucinations” in the industry and the media (not “lies” or “stupidity”, etc.). And it is almost a requirement that anyone doing a piece of journalism about a chatbot must first talk with a chatbot into the wee hours of the morning until they get the bot to say something psychotic.
In addition to hallucinations, it is often difficult to get generative AI to produce the specific results one wants. One has to try multiple prompts and to find ways to tweak the results. The issue of aligning what humans want with what AI produces is called the problem of “alignment,” and it is one of the most talked about and researched issues at the moment. (Sometimes people use the term “alignment” to refer to the issue of how to prompt a machine to do what you want, but sometimes people use “alignment” to refer to the more existential question of how to get machines to obey human intentions rather than going rogue and taking actions that could harm people.)
Hallucination and alignment may sound similar, but they are different issues. The hallucination issue reflects the fact that although LLMs are great at language, they don’t understand the world and don’t reason well. The alignment issue refers to the fact that the AI could do what humans want if we had a more precise way of controlling it.
There are a lot of researchers working on both these issues. The proposed solutions are too complex to get into here, but one thing that they have in common is that they don’t seem to rely merely on scaling up the data used to train LLMs. Rather they seem to be about creating “world models,” “knowledge graphs” or “vector embedding”–models of the real world which work in conjunction with an LLM so that the language model has a body of knowledge to interact with. It is not inconceivable that these problems could be solved with the right approach. The industry has achieved a great deal already and there doesn’t seem to be any force in sight capable of stopping them from their attempts to get to AGI.
The Dangers
Chickens at a Trump Rally. MHI, generated by Stable Diffusion.
Science fiction writers have been writing about the dangers of AI since the late 1800s, so most of us can easily imagine a wide variety of dystopian problems that could arise from the AI industry’s quest for AGI. Setting aside the ultimate fear of a future AI becoming super-intelligent and going rogue, many other serious problems are likely to soon arise from the current level of AI technology: out of control disinformation, the erosion of truth, labor displacement, the end of privacy, mass surveillance, robotic warfare, cybercrime run amok, and racially biased AI making decisions about finance and policing in ways that exacerbate social inequalities. Like human intelligence, generative AI can be put to a wide variety of uses. It will likely be applied to all sorts of dystopian uses, including uses that have yet to be dreamed of. This section discusses a few of these possible problematic applications of AI.
Labor Displacement
Theoretically, if AI was able to operate with AGI, it could replace any human in any field. The field of robotics is developing quickly in tandem with AGI, with fast moving research into robotic vision and movement. So it is not much of a stretch of the imagination to imagine a future of software and hardware AI workers with the potential to replace humans.
But what about the current crop of ChatGPT-level AI? Is it ready to disrupt the labor market? Already, LLMs have become a major issue in the strike by the Writers Guild of America. The union has demanded that their contract stipulate that AI can never be used at any stage in the writing process. It may be that some well-organized sectors, like Hollywood writers, are able to organize to keep their jobs safe from AI. But writing in general is a type of work spread-out over all sorts of industries. Much writing work is freelance work of a precarious nature. So while Hollywood writers might have luck in preserving their jobs, there is a lot of other writing work that will not be able to resist automation—copy writing, writers of web content, translators, editors, etc. These sorts of low-hanging fruit could easily be targeted by firms looking to replace workers with AI.
Other low-hanging fruit are jobs like customer service agents, accountants and bookkeepers, programmers and people who read medical scans like MRIs and EKGs. The livelihoods of artists and photographers have been put immediately in danger, and it is likely that AI generated music will further decimate a music industry already struggling in the era of streaming music.
The ability of AI to replace a lot of this work does not mean that it will immediately do so. Firms will have to consider how to handle the issues of hallucination and alignment before they adopt chatbots whole-hog for, say, their customer service departments. Furthermore, the cost of scaling up AI to take over a big job like a customer service call/chat center could be prohibitive. When call center workers in India are cheap and reliable, why bother replacing them with data-hungry computers? Like any other labor-saving technology, AI will be adopted once it is cheaper than, and as reliable as, human labor.
Warfare
The integration of AI into warfare is already underway, with AI running autonomous drones, gathering surveillance, etc. Recent developments in LLMs have created AI that can analyze a battlefield and suggest a course of action to a human operator. In a demo video of new, proposed military AI from Palantir, a human operator asks a LLM AI to provide an assessment of a battlefield–the number of enemy tanks, the amount of supplies available, etc. The AI then displays a menu of options for battle which the human operator selects from to initiate a course of action, similar to how we now choose driving routes from the suggestions provided by Google Maps.
Theoretically, this integration of AI into decision making processes allows for a faster and more accurate analysis of battlefield data than humans alone are capable of. But by speeding up this process of data collection and analysis the pace of warfare could quickly escalate. In the example from Palantir, the human operator acts as a guardrail against autonomous AI decision making. The AI does all of the analysis and suggests actions, but by allowing a human to make the final call it appears as if humans were still in charge. But is this the case? We do not know how or why this AI makes the decisions it makes. We can make guesses, but we don’t always know. If AI suggests a best course of action, how likely are human operators in the heat of battle to not follow the suggestions of AI and decide on their own course of action? What sort of personal liability would that entail? And if all the operator knows is what the AI has decided they should know, then the operator is making decisions within a very narrow window determined by the prior decisions made by the AI.
Furthermore, every time a human operator has to take the time to read over the AI’s analysis and select a course of action, this slows down the potential speed of military action. In a hypothetical battle between two forces both equipped with AI like Palantir’s AIP software, the speed of decision making could be a decisive factor in who wins and who loses. This creates an incentive to cede ultimate decision making to the AI.
A central aspect of modern militaries is the principle of deterrence. In a world of nuclear proliferation, states practice a degree of restraint in warfare. States avoid rapid escalation because it could trigger a global nuclear holocaust. Human restraint plays a vital role in this delicate and dangerous game. If AI creates an incentive to shrink the role of human judgement, ceding more and more control to AI, this could quickly devolve into very dangerous situations.
Cyber Security
The Turing test known as a “CAPTCHA” is a standard technique used by websites to prevent bots from accessing web pages reserved only for humans. (CAPTCHA stands for “Completely Automated Public Turing Test to tell Computers and Humans Apart.”) It is one of many aspects of internet security that relies on some method distinguishing between humans and machines. The CAPTCHA is based on the premise that bots are bad at identifying objects in pictures, while all humans can do this easily.
But did you know that every time you complete a CAPTCHA you are training AI to recognize images in pictures? In the early days of CAPTCHA many of these tests involved typing in a word that appeared warped or was written on top of another word. Human answers to these questions were used to train Google’s AI to read books as part of its ambitious project to scan all of the books in existence. Once that project was completed CAPTCHAs began to focus on more general image recognition with the results teaching AI how to perform as well as humans at this task. Currently most CAPTCHAs relate to visual phenomena needed to teach self-driving cars how to navigate streets–crosswalks, motorcycles, bridges, etc. AI has now become so good at identifying these images that it scores better than humans on CAPTCHAs.
AI isn’t just developing the ability to pass CAPTCHAs. Many people have a difficult time distinguishing between AI generated images, video and audio and human generated content. AI is quickly getting increasingly better at fooling humans. In February of this year, 53% of survey respondents in a ToolTester survey could not successfully discern between human generated text and text generated by GPT3.5. When OpenAI released GPT4 in March the number of people unable to tell the difference jumped to 63.5%. As competition drives AI companies to race to improve their models, it is likely that it will become increasingly difficult for people to spot the difference.
Now that AI can pass Turing tests, now that it can outperform humans on Captchas, we seem to have come to a dangerous threshold for cybersecurity. What is to stop AI bots from overrunning the internet? Any bad-intentioned fool with a beefy laptop and some open-source AI can generate totally original, AI-authored malware, send out AI-composed phishing emails, impersonate voices with AI audio tools, blow through CAPTCHAs, create entire networks of fake social media accounts full of fake photos and videos, all updated regularly with LLM-generated text, launch ransomware attacks, steal data, etc. We are likely entering a period of increasingly sophisticated and ubiquitous cybercrime.
Misinformation
The erosion of cybersecurity inherent in the proliferation of bots that can pass Turing tests and the wide availability of generative AI lead to an inevitable deluge of misinformation. Since the dawn of social media, society has been suffering from the slow disintegration of notions of truth due to the ease at which disinformation can be spread and the lack of accountability for those who create and spread it. The widespread proliferation of generative AI could have truly disastrous consequences when it comes to the ability of all sorts of actors to spread disinformation. Society could lose all grounding in truth.
Amazon is already being flooded with AI generated ebooks. Human authors are already reconsidering publishing their own books because they can be used to train a chatbot which will then write a cheaper book on the same topic. These new AI books are likely riddled with “hallucinations” or even deliberate misinformation.
AI is also being used to write web copy. The AI scrapes the web for content, content written by humans, and then generates its own content based on what it has learned. Similar to authors of books, authors of web content will soon begin to question the purpose of putting content online if it is only to be used to train AI. One of the main reasons for creating online content is to generate ad revenue from web traffic. But if people start to use ChatGPT as an alternative to the internet, then content creators will see less reason to create content.
As more and more of the content on the internet and in books becomes produced by AI, AI will be increasingly trained on AI-written data. When an AI program scrapes the web for a hamburger recipe it will likely encounter hamburger recipes written by AI programs. When AI researches ebooks for answers to questions, it will encounter other ebooks written by AI. The proportion of human generated content will shrink in relation to AI-generated content. The disastrous consequences are not hard to imagine. In an internet already rife with disinformation, full of bizarre feedback loops of propaganda and lies, the erosion of truth will likely be rapidly accelerated by the introduction of the current generation of sophisticated AI.
The technology is now proliferating rapidly in the wild, already being used for the worst purposes. Chatbots are running many social media accounts already, many of them featuring hate speech. The GOP has already begun using AI generated images in its political videos. The Chinese state is producing fake English news videos with deepfake technology.
Capitalist Competition and the Future of AI
Mark and a Monster. MHI, generated by Stable Diffusion.
While it is easy to imagine all manner of dystopian outcomes resulting from the proliferation and rapid development of AI, it is also possible to imagine many socially beneficial uses of the technology. The ability of AI to analyze large amounts of data, to model complex systems, and to outperform humans in certain tasks could be used in scientific research to cure diseases. It could be used to predict complex input and output variables for socialist planning. It could be used to eliminate alienating work.
In order to use AI beneficially, we would need significant guardrails to protect against its deployment for nefarious and destabilizing purposes. But we do not live in a society capable of ensuring AI can be deployed within some sort of bounded, regulated, safe context. Instead, the forces of capitalist competition are propelling the development and proliferation of the technology far faster than the private or public sector have been able to establish, or even conceptualize, appropriate frameworks for guarding against the inevitable problems of AI.
Since the release of ChatGPT last year, there has been a race among large tech firms to release rival LLMs. Google responded to the release of ChatGPT with the rollout of its new chatbot Bard. Google knew that Bard was not ready for primetime (Internal messages showed Google employees referring to Bard as a “pathological liar”), but it rushed to rollout the product anyway because of the hype surrounding ChatGPT. Microsoft launched Bing AI recently, a bot clearly not ready for primetime. But it was released anyway because of the pressure to compete with ChatGPT.
Meta’s release of Llama to the opensource community was a naked attempt to become the industry leader in AI. Meta hopes that if its technology becomes widely proliferated it will become the standard bearer in AI. This is a common strategy used by tech companies to try to outcompete rival capitals by releasing their products for free, hoping to become the dominant player in the field. Google did this with the Android operating system. It also allows a company to crowdsource the development of software, reaping the benefits of this development without paying a dime.
Meta is currently scheming ways to privatize publicly available web-data so that only powerful players in the industry have access to the valuable troves of data needed to train AI. The battle over who has access to training data may prove, in the long run, even more significant than the battle to have created the most popular chatbot. Most data on the internet is publicly available, created for free by billions of users typing social media posts, arguing on Reddit, etc. If private firms are able to claim ownership of this data, giving them the exclusive rights to the use or sale of this data for AI training/scraping, this would give firms like Meta and Google even more powerful positions in the global marketplace.
Meta is also working on its own AI silicon computer chips to compete with the likes of Google and Intel. Computer chips are emerging as a prime asset in the geo-political-economic competition between the U.S. and China. The U.S. has taken actions to restrict China from access to the most advanced chips used in AI in order to give U.S. companies a leg up in the AI race. It is a geo-political competition with far-reaching consequences in the economic, information and military spheres.
All of this competition creates a problem. The race to dominate the market takes priority over safe development. Why did Google rush to release the “pathological liar,” Bard? Because it needed to show investors that it had a product that could compete with ChatGPT. Why did Meta release dangerous opensource LLM technology into the wild? So that it could potentially spread its technology faster than rivals. Why did we train AI to pass CAPTCHAs? So that Google could translate all the books in the world and develop the technology behind self-driving cars.
In such an atmosphere of intense competition, there are little incentives to move slowly and exercise caution. The phenomenon of so many AI researchers publicly calling for a pause in AI development is clearly a sign that the forces of capitalist competition are pushing researchers past their comfort level. Many of these dissenters are people that have worked in the field for decades under the assumption that the dangerous potential problems of AI were only technical problems that could be carefully sorted out over time. Now that the forces of competition are compelling constant releases of new, improved AI, these researchers are suddenly coming forward to raise the alarm.
What seems a technical problem on the surface has become embroiled in a social problem—the mindless and frenetic competition that compels production for profit at the expense of everything else.
What incentive is there to put up guardrails when rivals might release a product with fewer restrictions? Why spend time carefully fine-tuning your model so that it doesn’t produce hate speech, when there are open source models that have no guards against hate speech? What is the incentive? There are companies developing AI that they claim is able to identify AI-generated images and text. But we all know that these will just be used to train the generative AI to pass increasingly discriminating Turing tests, so there doesn’t seem to be much reason to hope that these “solutions” will solve much in the long run.
But what about the ability of states to regulate this technology? There are moves in the EU to provide some safety standards around AI development. And there have been performative gestures from the US government. But one must realize that any regulation is caught up in the geopolitical struggles between nations. China and the US are currently in the midst of a serious AI arms race as both countries pour billions of dollars into research in order to outcompete their rival. Any significant regulation of AI development in the US will give China a leg up in the fight. In this context of fierce geopolitical competition, it seems unlikely that any state will be able to effectively regulate the development of this technology in safe ways.
Though generative AI is a new and flashy technology, its development is caught up in a tale as old as capitalism. The pursuit of profit for its own sake, the privileging of individual interests in socially destructive ways, the reduction of all activity to the rules of competition–these are not universal aspects of social life. They are the universal aspect of the capitalist mode of production. No new technology escapes these pressures.
thanks Brendan.