Some Summary Thoughts on LLMs
Following up on my earlier conversations with Gemini and my analysis of the capabilities of LLMs through a Lacanian lens, I wanted to offer up a few summary thoughts about LLMs.
Strengths
In general, I find LLMs to be astonishingly strong and capable when it comes to language-related tasks. This is the way in which they most closely match general intelligence. It’s also the way they shine vis-a-vis one of the most famous measures of artificial intelligence: the Turing test. Most of you have probably heard of the Turing test. The idea is that a way to test the capabilities of machine intelligence is to put it at a satisfactory remove (such as typing on screens) and then see if the interlocutor of a machine-based intelligence can distinguish between a machine and human conversational partner. With the latest models, a clever and knowledgeable interlocutor could no doubt tease out some mistake that makes the presence of machine intelligence glaringly obvious. And of course, LLMs, the publicly available commercial models, have been trained to identify themselves as machine intelligence. But if you were to take some of the latest models and strip away some of their base prompts and tell them to mimic a human and give them some details of the human they’re meant to mimic, I think it’s pretty safe to say that today’s LLMs could pass the basic Turing test with flying colors.
As a quickie test, I just asked Gemini to pretend it was a human and gave it some parameters (name, age, gender, living location, profession, and a few other details) and then asked it to act as if it were a human having a conversation with me. I think there was maybe one tiny detail that was a little off, but it was really good at managing details while using a very persuasively casual human tone. Anyway, I think it’s pretty widely agreed that LLMs can pass the Turing Test. Some critics have used this to suggest that the Turing Test was just a bad way of measuring machine intelligence, but I disagree.
As I argued in my previous post about LLMs and Lacan, I think that critics underestimate LLMs in some key ways that relate to the notion of experience. The fundamental argument is that because LLMs can’t possibly experience the world the way humans do that their abilities are therefore not “real.” As one famous critic, Emily Bender, has put it, LLMs are merely “stochastic parrots.” The argument she and several other scholars put forward is that LLMs mathematically compute probabilities based on previously ingested texts about the most likely next word in a string of words. There is nothing that could be referred to as “understanding.” Part of the argument, as I read it anyway, is that LLMs can’t truly demonstrate understanding precisely because they have no experience of the world. They’re just word producing machines. The flaw in this argument, in my view, is that it presupposes that in order for something to have anything like human intelligence, that something must have acquired its understanding in the same way that humans have. But why would this be so?
As I argued in my previous post, the ability to process language when combined with lots of texts about human experience gives LLMs something that approximates human understanding well enough to be considered a form of intelligence. Again, the devil is in the details. I absolutely agree that LLMs are limited in their ability to have a bunch of experiences that we consider to be uniquely human, starting with emotions. Emotions are not a simple thing to define, but they do seem, in their human form, to be tied to chemical processes that happen at least in the human brain and probably elsewhere in the human body. So this qualifier, a form of intelligence, is tricky. What I think is absolutely wrong, though, is to use the language of some key critics, such as Bender and Marcus. There are lots of critics out there, and I won’t try to lump them all together to create a straw person, but it is a common argument that LLMs have no real comprehension precisely because they have to real experience of the world. But I’ve yet to see a persuasive argument about why real experience of the world is a requirement for real understanding.
This is why I think the Turing test has not been shown to be irrelevant. Intelligence is as intelligence does. The fact that LLMs can past the Turing test does have significance. The language of Bender and Marcus, in particular, strikes me as quite strident and polar: the suggestion is that LLMs either have understanding or they don’t, and currently they don’t. And given the requirement that a machine intelligence must be able to experience the world in a way that at least approximates human experience (sensory, at a minimum), they will continue to not have comprehension until this fundamental requirement is met. This refusal to accept that understanding could be a continuum is a key part of the error in this argument, in my view.
Interestingly, as another thought experiment, I looked up a few past ways that people have criticized LLMs and tried them out on Gemini (I use Gemini as my measure of current models because I pay for the premium version of Gemini, but don’t pay for the premium versions of most other models, with the possible exception of Copilot). Critics are always looking to trip up models, and have zeroed in over the years on certain examples. For example, someone once asked an LLM to tell them how many legs are in the left rear of a cat. The model answered four. This was used to demonstrate that an LLM can’t answer a quite simple question because it has no actual experience of the world, and has never seen a cat. While there are many such examples out there, every one I’ve tried has failed to stump the latest model of Gemini. Of course there will be new such examples, but I point this out only to show that models have improved significantly and continue to improve rapidly.
Weaknesses
One way that AI models get criticized is that there are limitations in their dataset. Bias is a common such problem that critics point out. This is an old problem that anyone who has ever coded faces: GIGO (garbage in garbage out). The criticism of LLMs here though runs deeper: that whereas a human is capable of critical thinking or faculty which lets a human see not just logical flaws but ethical flaws in information, an LLM has no such capability. The danger suggested is that an LLM has no value system. Meaning for an LLM consists of an amalgamation of ingested texts, many of which are bound to be contradictory. Having no internal compass, therefore, an LLM is subject to the chaos of its dataset.
This is, I think, a legitimate concern. Of course, it’s possible to have an AI system ingest a set of what it is told are values. A simple example of this idea is Isaac Asimov’s laws of robotics: “A robot may not injure a human being or, through inaction, allow a human being to come to harm. A robot must obey orders given it by human beings except where such orders would conflict with the First Law. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.” You can pass this information along to a machine intelligence. Getting that information to be a hard, fast rule to which the system adheres is far, far trickier, as is knowing how the system has parsed this information. This is one of the key arguments of Yudkowsky and Soares in their recently published book If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All: it’s what they call the alignment problem. You can intend for a machine intelligence to have certain goals or fundamental limits, but, for a wide variety of reasons, you can’t guarantee that such goals or limitations will work. As Yudkowsky and Soares argue, the danger inherent in this problem grows with the capabilities of artificial intelligence.
I do find it a persuasive argument that the very nature of intelligence is rife with problems, contradictions, and, to use the work of Doug Hofstadter, strange loops, ways that reasoning loops back on itself, creating paradoxes. Our legal code is a great example of this. Lawmakers attempt to write laws in ways that reduce ambiguity, and yet our courts are filled with cases in which the correct interpretation of law is far from obvious. The world of ethics is that much more complex, as demonstrated by the multiplicity of ideas about what constitutes correct behavior. It follows, then, that machine intelligence is no less likely to face the kinds of problems that have bedeviled humans.
Other issues with machine intelligence:
The epistemology issue. The ability of machine intelligence to approximate “reality” with such ease is disturbing but really a problem of scale, in my view. It’s always been possible to fake information. A human can simply lie, of course. And a photograph can be doctored or staged. There are no shortage of such examples. The list goes on and on. But the ease with which people can create fakes of videos, photos, and even recordings of people saying something in what sounds like their voice serves as a huge leap backward to our ability to distinguish actual from manipulated representation of reality. But of course, these kinds of problems run far deeper than the capabilities of machine intelligence. As Orwell saw so well, power has a dangerous array of tools with which it can shape reality. The better the tools the worse the problem?
The creativity problem. Critics have argued that machine intelligence has no spark with which to be actually creative. It can take existing forms of creativity and recombine them, but can’t generate anything truly original. The real problem, though, is the ease with which machine intelligence can produce “good enough” mimicry of creativity, which, some fear, will end up drowning real creativity in a sea of mediocrity. As any artist will attest, though, I think this problem is nothing new. But again, ease and scale may be their own problems.
The theft of existing intellectual property. Machine intelligence has been built on the shoulders of actual human intelligence and creativity, and the companies that have built machine intelligence have stolen this treasure trove outright, with no compensation to the human beings, dead or living, who have produced this trove. This is a shame, I agree, and I would hope to see some compensation to the many thinkers and artists who have been plundered. Do I think it’s likely that this will happen? Sadly, probably about as likely as Native Americans receiving compensation for the long string of broken treaties that were made with them to build the United States or for African Americans to be compensated for the immense economic value that went into building our country on the backs of their enslaved ancestors.
Workforce replacement. One of the great “dangers” of machine intelligence, particularly when combined with possible advances in robotics, is the ever increasing chance that it will replace a significant number of jobs. It’s not easy to determine how fast this will be happen, but I think it’s worth being concerned about because it’s likely to accelerate very rapidly once it gets underway on a serious scale. I put “danger” in quote marks here because whether it’s really a boon or a bane depends on how our governments around the world respond to this prospect. There’s lots of talk about UBI and how that might be a response to this issue. What we’re facing, I think, is an intensification of the liberal/conservative divide in this future world. UBI is essentially a liberal version of how to deal with this problem: an extension of the idea that the productivity of a society belongs to the whole society at least to some degree, which means that throwing a large chunk of the populace out of work causes us, as a society, to owe these people something. A more conservative take could be that those who own the means of production (the companies producing machine intelligence and the robots that can execute its capabilities in the physical world) also own the benefits that result, while owing nothing to society at large. It’s hard to imagine such a viewpoint really flying if some significant amount of the populace is thrown out of work. Even if such a thing were spread out across decades, it would likely cause massive political upheaval, likely rewriting the rules of economics.
Pedagogy. What does education mean in a world of widely available machine intelligence? How can one possibly educate children (or anyone) in a world where it’s so easy to fake understanding? The immediate problem that LLMs cause relates closely to the existing capabilities of LLMs: reading and writing. Any student can use an LLM to take a big shortcut to reading comprehension and writing as a demonstration of comprehension. In much the same way that calculators made longhand mathematical calculation pointless, LLMs are making a wide range of other forms of mastery difficult to assess and therefore successfully convey.
Summary
As you can see, my list of issues is long and, to my thinking, far from trivial. Perhaps one of the biggest dangers is precisely how much we, as humans, are in over our heads, so to speak. The course of AI within a world geared towards profits is swift and apparently more or less unstoppable. Lawmakers throughout the world are unlikely to put any meaningful barriers in the way of machine intelligence businesses. Even if some politicians could agree on which barriers to erect or ways to channel this fast flowing river, who would risk letting their country get behind in the race for supremacy? Whether it’s true of AI or not, our global economy is premised on the idea that developing cutting edge technology is a race, and to the victors go the spoils, while the losers get colonized, if not geographically than economically. Each of us can individually try to make sense of the trends and stay ahead of the curve, but how this will affect entire nations and their futures is probably beyond even existing ways that people have organized to improve large scale outcomes. It may just be an expression of my own political preferences, but I can’t help but imagine that governments that are geared towards justice and social welfare are more likely to solve these kinds of issues than governments that enable oligarchs. But it will also help, I think, for there to be a robust conversation about what our communities (across various scales) value, and how these values apply to such fast moving technological trends.