Published on 2023-08-01 14:25 by Topfi

My Thoughts on Generative Models

This won’t go away

The cat is out of the bag, the genie won’t go back in the bottle, and Pandora’s box has been opened. We have these models. There will never come a day (beyond societal collapse; good luck then) when we won’t have these models. More to the point, this technology will never be as limited, tend to make such noticeable mistakes, or as likely to “hallucinate”, as currently. These models are only going to improve, get more efficient, and increase in capability and accuracy. On top of that, these models will run locally, negating any restrictions or attempts at regulation. In fact, some already do.

We now live in a world with generative models producing images, videos, text, and speech. We cannot act like this isn’t the case, as, just like it was a fool’s errand to try and regulate locally run encryption, so it will be with these models. There will be attempts and different proposed solutions, like watermarking generated vs. human-made content, though all of these will at best postpone the inevitable. Once locally run models are truly competitive with the models currently reliant on large, remote compute clusters, any solution attempting to differentiate human-made content from purely generated output will end up moot. Just as with encryption, centralized institutions may try to enforce what they believe is right, but they will fail. If you need proof, look up the Clipper chip.

Impact on Employment

The capabilities of LLMs when it comes to handling certain knowledge based tasks are already very impressive, though tend to be very unreliable in my limited experience. Provide these solutions with a database or some other source to ground the output and they can appear even more astounding, though extended use still unearths clear deficits. Image generation models are equally impressive. Both also seem to improve at an incredible pace. What impact will this have on the labor market? I dare not make any prediction. Well reasoned arguments have been made that these models will supplant a significant part of current jobs. Others have highlighted parallels in history that lead to far more optimistic outlooks. How similar things have happened before, how new technologies tend to merely assist with and ease workloads, allowing people to focus on more important, less mundane tasks and that these rarely reduce the total number of jobs in an economy.

In the past, I very much agreed with the latter philosophy, looking mainly at how the absolute number of available employment tended to historically increase, even whilst automation/calculators/computers took away tasks once at the core of important professions. Having spent more time reading up on these transitions however, I now see, that even if we merely transition into a world of professionals working “with AI” (read generative models) rather than being replaced by these models entirely, timelines are key.

Coal mining jobs in parts of the United States are being and have been lost, partly in the transition from fossil to renewable energy sources, though to a far larger extend as part of newer solutions such as mountaintop removal, which alongside other recent developments allowing for larger scale mining operations with fewer workers, has in my limited research come up as the most significant contributor in job losses. As this however is purely based on me reading, without any personal or professional experience, I may be wrong in this regard.

Regardless of what has caused the decrease in available employment in the US mining industry, purely looking at the absolute numbers, any loss in jobs there should not be an issue over all, as this happens concurrently with new industries like solar adding thousands of additional jobs to the labor market. The catch is, unfortunately, that those new jobs are often in a different location and may require vastly different skills. As resources for relocation and retraining are limited, people tend to suffer, even when this transition brings with it a net increase in jobs across the wider labor market. Or, to get this back to LLMs and generative models: If these technologies somehow lead to a net increase in labor demand (or the absolute number of jobs roughly stays the same), that may still come at the expense of vast sections of the population, as for one reason or another, they are going to be left behind in the transition.

That is a big “if” though as I feel no historical analogue fully captures the extent which I see in generative models potential impact on human work. Perhaps I am overly dramatic, but the improvements we have seen in generative models across the last five years, combined with future estimates, make even the significant pace of our transition from fossil fuels to renewables appear sluggish by comparison. In response, an often cited solution to either smooth out such a transitionary period or replace lost income from employement all together is the introduction of an univeral basic income (UBI).

Whilst I personally have been a proponent of UBI in the past, I feel it is important to caution against ignoring potential alternate solutions. Due to my preconceived notions that UBI would be beneficial to society, I may be tempted to just view it as a panacea. If you got a hammer, everything tends to seem like a nail and it may very well be that UBI is not the solution I, as well as others, might see it as. I feel that when it comes to the potential future of large sections of the populus, we should be able to evaluate any idea, even ones we tend to few favourably from the outset, in a fair, yet critical light. Just because I like UBI as a concept and feel we should explore it, doesn’t mean decision makers shouldn’t discuss alternate options. And, even if UBI is the ideal solution, we shouldn’t loose sight of one thing:

UBI tends to mean something different, depending on who you ask. For some, such as myself, UBI should supplement existing social and health care safety nets. Others very strongly believe that UBI should replace existing safety nets.

Just talking about UBI as some nebulous concept without having a detailed discussion on what that may entail, how it would be implemented and financed in the long term and what that could mean for people whose, e.g. medical needs may require additional financial assistance beyond what UBI provides, isn’t going to yield results in my eyes. Or, even worse, it leads to flawed results, erroding safety nets vital for large sections of the population. All the nuances of introducing a way to soften or remove the impact generative models may have on the labor market must be thouroughly examined by policy makers before a decision is being made.

Past Informational Revolutions

A take I have seen repeated multiple times in response to currently available models is that “this goes to far”. That these models are so capable that humans will inevitable lean on them en massé and anyone who is born after this technology has become common place will invariably grow up to live their life without any solid skills or base knowledge. Now, I cannot look into the future, but unlike with employment, I feel that history can serve as a good indicator what direction we are heading in. Whenever a new repository for knowledge or machine to take over a complex task gets released, there will be those that decry it as being the one thing that will lead to the end of all academic pursuit.

To quote Plato: “If men learn [to read/write], it will implant forgetfulness in their souls; they will cease to exercise memory because they rely on that which is written, calling things to remembrance no longer from within themselves, but by means of external marks.” According to Plato, people would overly rely on books, which would bring about an end to actually learning and retaining knowledge. That, obviously, has not played out. Reading and writing are universally agreed to be beneficial for students and a core part of modern education. Beyond the idea that Plato was incorrect in decrying books 2000+ years ago, let us remember that something similar has happened more recently.

Whether we are talking about calculators or the ability to search for the contents of books online, there are countless modern examples of technology, initially heavily criticized, then adopted as an invaluable tool for educators and their students. This song and dance happens every time a new informational technology shows up. Looking at how it tends to go, I suspect that the best approach here is to provide the next generation with good core knowledge concerning the new technologies ins-and-outs before introduction. That way, students can ideally get the most out of the new technology in the long run. You won’t get much use out of a calculator if you lack an understanding of the concepts you use it for. Since generative models are however likely going to find heavy use in entertainment, it could be hard to impossible for a future students first encounter with these models to be in a controlled, educational setting, similar to how few students first encounter with a personal computer or the internet happens at school.

There is however also another, more imminent perspective. A lot of educators are concerned that these models are already being used by students to receive solutions to excersises without the need to engage in the subject matter at all. That is a fair concern and there have been countless examples of students being caught red-handed. It pains me to say, that I currently see few, if any, ways to mitigate this however. Whilst yes, these models currently can produce a somewhat consistent style[0] that may be distinct enough from a specific students expected output to raise suspicions, this can already be mitigated by requesting a specific style or making a superficial edit to the LLMs output. There are also companies that promise to provide “ChatGPT detection” for educators, but once you start researching and testing these with content you know for certain was created entirely by a human, you will find that many of these are no better than a coin toss. In fact, looking into cases, there are roughly as many cases of students getting caught passing off LLM output as their work, as there are those falsely accussed by these “detectors”. So what to do? Trust nothing you didn’t see a student make?

Currently, students should try to keep notes that showcase they have undergone a process researching the content they have submitted. Google Docs histories showing a timeline of all edits can also be helpful. But that is not going to be a long term solution for a few reasons:

1. There are going to be excersises that a student can finish in a straight forward manner. Writing a fictional story, for example, can be done without significant edits in one sitting.
2. With broad and cheap API access and (soon) local models, adding the ability to simulate edits and research notes is not a question of if, but when.
3. Educators, from Kindergarten to University, are also, ideally, people of trust. The best teachers are both authority figures, but also give students the feeling that they can convide in them. This trust is a delicate balance that could suffer from students being under constant suspicion of utilizing LLMs and may be damaged if a teacher accuses someone who actually did put in the work.

As educators have the responsibility to educate, their concerns are more than reasonable however. So what to do? I am merely musing online, with no background in education whatsoever, so I am certain there are more solutions that aim to adress the concerns I have covered, but a potential solution I have found is to adopt teaching methods that balance being able to monitor a students progress without noticably reducing trust in them. An example of this may be a flipped classroom approach, which could make it easier for teachers to detect whether a student is doing a sufficient amount of work on their own or is overly reliant on external sources (be that copying from books, online encycolopedias or LLMs), whereupon appropriate actions can be taken to ensure that the student becomes familiarized with the subject matter.

[0] Note the recent release of a preprint that talked about a detection model that is able to differentiate between human and ChatGPT content with 99%+ accuracy. However, these findings are limited. The method described utilizes ChatGPTs specific style as a basis for dectection. This can be mitigated using a variety of methods, such as querying for a different style or making manual edits. It also is likely that future models will more easily produce a wider range of output styles and could even be instructed to adopt a students expected writing mannerisms reducing the possibility of detection further. As said on multiple occassions throughout this page, I see no way to permanently differentiate between human and LLM ouput. Even watermarks will at best be a temporary solutions, until local models have caught up.

Written by Topfi