Link dump for May 22nd 2023
Links: AI & language; AI & science; AI & jobs; AI & jobs, again; Deepfakes in court
Following are five interesting links I found over the past week.
AI and the Limits of Language
This short essay argues that, in spite of its universal use, language is not a great way to convey information:
Language is a very low-bandwidth method for transmitting information: isolated words or sentences, shorn of context, convey little. Moreover, because of the sheer number of homonyms and pronouns, many sentences are deeply ambiguous: does “the box was in the pen” refer to an ink pen or a playpen? As Chomsky and his acolytes have pointed out for decades, language is just not a clear and unambiguous vehicle for clear communication.
The piece goes on to argue that, since language is limited in the amount of information it can convey, it follows that large language models, which are of course based on language, can’t convey much more information or understanding:
While some balk at using the term “understanding” in this context or calling LLMs “intelligent,” it isn’t clear what semantic gatekeeping is buying anyone these days. But critics are right to accuse these systems of being engaged in a kind of mimicry. This is because LLMs’ understanding of language, while impressive, is shallow. This kind of shallow understanding is familiar; classrooms are filled with jargon-spouting students who don’t know what they’re talking about — effectively engaged in a mimicry of their professors or the texts they are reading. This is just part of life; we often don’t know how little we know, especially when it comes to knowledge acquired from language.
LLMs have acquired this kind of shallow understanding about everything. A system like GPT-3 is trained by masking the future words in a sentence or passage and forcing the machine to guess what word is most likely, then being corrected for bad guesses. The system eventually gets proficient at guessing the most likely words, making them an effective predictive system.
This past weekend I was experimenting with ChatGPT’s plugins. Specifically I was using the Wolfram Alpha plugin. I asked it: Using Wolfram Alpha, can I generate code which draws a klein bottle?1
ChatGPT + WolframAlpha dutifully returned the code and the resulting plot:
I then asked Can you explain why this is a klein bottle? This doesn’t look like a “bottle” to me. So I wanted to know why ChatGPT + WolframAlpha says that it is a bottle.
I provide its response in a long footnote below2. The reason that I recount this story is that it gets to the heart of what the essay I link to above argues. The information provided appears to be accurate. It’s certainly coherent and well-written. But it is not evidence of understanding. Merely conveying words doesn’t impart or imply understanding. A toddler babbles all day long, but it is not clear how much of that babbling is understanding, and how much of it is parroting. So, too, it seems to be with LLMs.
In some sense, I think this is ok. I ask ChatGPT why a plot is a Klein bottle, and it provides me a response. It’s not ChatGPT’s responsibility to know whether I understand its output, or even whether its output is accurate. These things are my responsibility. Of course, it’s also true that at least some people who use LLMs like ChatGPT don’t understand the distinction between LLM output and understanding. But this is true of virtually all technology: most people who use a given technology don’t understand how it works at a deep level. It may be true that people erroneously assume that LLMs are necessarily factual and accurate, but people’s misapprehension of a given technology is hardly the fault of the technology.
How is AI impacting science?
AlphaFold, developed by Alphabet subsidiary DeepMind, is a database of protein folding, generated by AI. An interesting question that arises is whether the AI technology used by DeepMind to investigate protein folding can be generalized to other areas of inquiry.
In any event, Michael Nielsen gave a speech recently, which he posted to his site. It’s hard to extract a quote from the speech, and the speech is a bit far afield from my knowledge, so I will offer the preamble he presents, as an indication of his arguments:
The biggest success of AI in sicence so far is the AlphaFold 2 system. This is a deep learning system which has made large strides on a fundamental scientific problem: how to predict the 3-dimensional structure of a protein from the sequence of amino acides making up that protein. This breakthrough has helped set off an ongoing deep learning revolution in molecular biology. While obviously of interest to molecular biologists, I believe this is of much broader interest for science as a whole, as a concrete prototype for how artificial intelligence may impact discovery. In this short survey talk I briefly discuss questions including: how can such systems be validated? Can they be used to identify general principles that human scientists can learn from? And what should we expect a good theory or explanation to provide anyway? The focus of the talk is extant results and the near term, not the longer-term future. However, I hope it may help in grounidng an understanding of that longer-term future, and of the benefits and risks of AI systems.
How scalable is your job?
AI is changing the nature of work:
Some celebrities are creating "synthetic" versions of themselves and charging people for private conversations. You can now build similar tools and draw on the insights and intellect of living and dead scientists and philosophers. AI enables experts to productize their knowledge in new ways. James Altucher recently created a rough "tool" that lets people "tap into" his way of thinking about new ideas. To what extent can this be done with your expertise?
At first blush, this sounds like something that only celebrities3 have to worry about. But as Dror Poleg notes in his piece, AI technology allows many more people than just celebrities to radically scale their output and their reach. It’s important to consider how AI can increase your scalability.
Will AI fix work?
Microsoft released a report recently about how AI will make our work lives more efficient. We’re all familiar with the modern knowledge workers’ lament:
We’re all carrying digital debt: the inflow of data, emails, meetings, and notifications has outpaced humans’ ability to process it all. And the pace of work is only intensifying. Everything feels important, so we spend our workdays trying to get out of the red. Nearly 2 in 3 people (64%) say they struggle with haing the time and energy to do their job—and those people are 3.5x more likely to also struggle with innovation and strategic thinking. And nearly 2 in 3 leaders (60%) are already feeling the effects, saying that a lack of innovation or breakthrough ideas on their teams is a concern. There are only so many minutes in the day—and every minute we spend managing this digital debt is a minute not spent on the creative work that leads to innovation. In a world where creativity is the new productivity, digital debt is more than an inconvenience—it’s impacting business.
We can easily imagine a future world of work in which AI agents autonomously handle a lot of this digital administrivia for us. What I am unclear about though, is whether having AI free us from digital drudgery means that we’ll recoup more time. What if work demands merely expand to fill the new time?
Deepfakes in court
The law is generally a static thing. Technology is much more dynamic. Many obvious implications follow from this, the most basic of which is that technology often advances at a rate far beyond that of the law, and when these two things conflict, hilarity ensues.
And, as the more acerbic among us, which is to say, yours truly, has predicted, we are seeing this with the rise of deepfakes. It is now trivially easy to copy someone’s voice or to create a photographic or video depiction of the person out of whole cloth.
If, like me, yours is a somewhat skeptical view of our legal system, you can see where this is going: pretty much no audio, video or photographic evidence submitted in court as part of a legal proceeding can be considered to be genuine. Sure, prosecutors or defense lawyers will assert that the evidence submitted proves (or disproves) the arguments the lawyers are advancing. But, again: creating convincing, yet fake, depictions of people doing bad things has become trivially easy.
Anyway, here’s an article from NPR about this very issue.
Klein bottles are a mathematical structure. Cliff Stoll creates 3D versions of them.
[ChatGPT’s response to my prompt Can you explain why this is a klein bottle?]: A Klein bottle is a non-orientable surface, which means it has no inside or outside. It's a two-dimensional manifold against which a system for determining a normal vector cannot be consistently defined.
The Klein bottle is closely related to the Möbius strip. If you were to take two Möbius strips and attach them along their edges, you would get a Klein bottle. However, this is impossible to do in three dimensions without the surface intersecting itself, which is why Klein bottles can't exist in our three-dimensional space without self-intersection.
The code provided generates a 3D representation of a Klein bottle using parametric equations. Here's a breakdown of the code:
ParametricPlot3Dis a function that generates a three-dimensional plot of a surface defined by a parametric equation.The three functions inside the
{}brackets define the x, y, and z coordinates of points on the surface in terms of two parametersuandv. These functions are designed to generate the shape of a Klein bottle.{u, 0, 2*Pi}and{v, 0, 2*Pi}define the ranges of the parametersuandv. They are both allowed to vary from 0 to 2π, which means every point on the surface is included in the plot.
The resulting plot is a visualization of a Klein bottle in three dimensions. However, it's important to note that this is not a "true" Klein bottle, because a true Klein bottle cannot exist in three-dimensional space without intersecting itself. The plot shows a commonly used representation where the surface intersects itself, but you should keep in mind that a real Klein bottle would not have this intersection.
Incidentally, the 1980s actress Justine Bateman has a very interesting Twitter thread about how AI will affect actors’ work. I think she’s correct about where AI and entertainment is headed. I think she’s less correct about the 2023 Writers Guild of America strike.