September 2024 Diary - Thoughts on AI from an AI Product Manager
If you didn’t already know, unfortunately my main source of income does not come from the sparse blogs I write on Substack, but in fact comes from working for a large technology company as an AI/ML Product Manager.
Unfortunately, again, that title doesn’t mean much to people who work outside of tech. The short version is that I come up with and implement ideas for Artificial Intelligence & Machine Learning (AI/ML) products in a given space (in my case, customer experience and marketing). Basically, I think about AI capabilities all day and try and figure out where I can use them for my company.
Thanks for reading On Second Thought! Subscribe for free to receive new posts and support my work.
I don’t typically use this space to talk about my job because honestly, I like to use this to explore other areas of my life but recent developments in AI in the past year have made me want to share what I know along with some of my personal thoughts as someone who works extremely closely with this technology.
A very quick crash course for those who might be newer to this whole topic: Artificial intelligence (AI) is a technology with human-like problem-solving capabilities. This type of technology is not new, we have been working on it for decades already. But where things have gotten crazy in the past couple years is with the rise of a subsection of AI called Generative AI. Generative AI (GenAI) refers to newer technology that can generate (create) high-quality text, images, and other content based on the data the technology was trained on. If you want to dive any deeper than that, I have linked some research papers that explain how this technology actually works in more detail.
Why should you care? Prior to the rise of GenAI, this technology was carefully cultivated by researchers, scientists, and businesses. It was near impossible for an average person to interact with it and it was calibrated towards very narrow purposes. GenAI has democratized this technology and this is the first time it’s being used towards general applications rather than a very specific use cases. You can actually hold this tech in the palm of your hand and use it to create something new which is pretty groundbreaking.
Personally, I feel very conflicted about this technology. I do believe it can unlock speed in new achievements in a variety of fields but I believe that will come with heavy, heavy costs at the personal and sociological level. So here’s my breakdown of the good, the bad, and the ugly and why you can’t just simply ignore this technology.
The Good
Most GenAI solutions are powered by the use of Large Language Models (LLMs). LLMs are data models that are trained on an incomprehensible amount of unstructured data that makes them applicable towards a variety of goals. From this point on, every time I use the term GenAI, know that they cannot operate without an LLM or similar large model (also called foundation models) supporting underneath. LLMs are where I think a lot of the good part of this technology stems from. Because these models are trained on so much data, they can draw connections and conclusions where humans have previously been limited. For example, an LLM can be trained on every piece of medical research and can project diagnoses quicker than a human doctor. From a business perspective, an LLM can consume all of your business information and you can get insights much faster than individually assessing every report by hand and synthesizing the information.
The Bad
Like any technology, GenAI is not perfect. Although this technology is geared towards general applications, as you try to use it for more things, it tends to make more mistakes. You may have heard that this technology ‘hallucinates’ which essentially means it will make up new information that has no basis in truth or reality. These hallucinations and mistakes do not occur consistently and ‘why’ a model chooses to do what it does remains frustratingly opaque. Although this is perhaps the most human thing about it, it makes correcting these issues very difficult. In order to use it for a specific purpose, you have to train an LLM or a smaller model to do specific tasks which requires a lot of specific data and a lot of set up.
Furthermore, because LLMs require such large amounts of data, they are trained on any data their overlords can get their hands on - this includes the entire public internet. YOUR data in case that wasn’t clear. We already knew data privacy was an issue even from general internet use and social media. And while you may not be concerned with random Reddit comments, if you’re an artist, this means any art you’ve shared is feeding these models. This blog is probably in a few of those models already. Your content is actually being used to train these models. So not only does this mean that the model can produce art and content as if it were a specific person (clearly an issue for the people who make a living off of creating), it means your data is being directly used towards training a product on behalf of a company. You might think, well my data is being used elsewhere so who cares? The difference is that you signed up for instagram or whatever and signed their terms and conditions. You never got such an agreement from an LLM purveyor did you? It is the Wild West of AI.
Laws are already too slow for technology that has the potential to be downright dangerous. There have already been instances of people using this technology to create fake nude photos of classmates and multiply downright false or damaging information and there is no way to persecute people for this or to convey any real consequences. This is not a unique problem for this technology, but because this is a technology that everyone now has access to, the amount of damage that can be done is potentially larger than previous products.
The Ugly
This section is where I’ll highlight how I personally thing this technology will impact society. Unfortunately, I do not believe this technology will be a net good. I do think it has the power to do good, but I think the negative impacts will be far greater. Essentially, my ‘ugly’ thoughts about this technology can be broken down into two categories: 1/ The means of this tech may not justify the ends. 2/ Long term impacts on society
Let’s start with the first one. This technology demands a huge amount of resources. One query to ChatGPT uses as much energy as it takes to power a lightbulb for approximately 20 minutes. Larger queries use even more energy and there is already a huge race to buy up land and resources to develop means of extracting more energy in order to train the LLMs and power this technology on an ongoing basis. This technology will have broad impacts on climate change, energy allocation, and land allocation. All of those impacts will affect you and I in the form of rising temperatures, expensive energy prices, and expensive housing and land prices. As if those categories were not all distressful enough already.
This brings me to my even more personal second point and my prediction for the future. GenAI is already being used at high rates with businesses clamoring to find ways to use it to increase profits. However, I believe the bigger long term danger to our society comes from this technology being promoted to be used at an individual level. This technology is completely results oriented. The whole purpose of it is to cut out the manual work of creating something and give you only the end result. You can use it to create financial reports without conducting financial analysis yourself. You can use it to plan a trip for you without doing any research on the places you’re interested in. You can use it to create art without learning anything about shapes or colors. On the surface, this can be super helpful and can speed up work for many people. However, I do believe there is something dangerous about removing the process part of the equation.
First of all, removing process removes the learning. Sure, I can use GenAI to write an email for me but I only know if the resulting email is usable because I already learned how to write an email the manual way. For a completely new area, it will be difficult for me to know if the output is good or bad and how to refine it because I won’t know which part of the underlying process needs fixing. For an adult who has already specialized in a few areas, this may not be an issue, but I do think this is an issue for anyone whose frontal lobe hasn’t finished developing. Students are already using this technology to write essays and complete assignments for them, essentially using it to outsource their own thinking. We shouldn’t be surprised by this, the US education system has long been results oriented, but this can have devastating long term impacts on a child’s development. They are already outsourcing the necessary processes of intaking new information, thinking about it, and developing a response.
What is even more insane about this is that even if they do get great results in school using this technology, it will not translate at all to the workplace. Every interview I’ve ever done has put more emphasis on my explanation of my learning and decision making processes rather than the direct outcome. My ability to explain how to approach a problem is what is valuable at my job, not just the end result. People expect you to be able to explain how you came to a certain conclusion. ‘ChatGPT did it for me’ is not good enough. At least not yet.
Which leads me to the greater societal impact. This technology was released into the wild and will now potentially lead to an entire generation of people who literally won’t know how to think for themselves. I’m hard pressed to think of anything scarier.
What Now?
You can’t put the genie back in the bottle. This technology is out there now and there’s no way it just goes away. A popular criticism of AI generated literature that resonates with me is “why would I bother to read something no one bothered to write.*” This brings up both new questions of value and consumption. If this technology can take care of most of the process of creating something, then where does the value of it stem from? When we take something like writing a novel, we place value on that novel because the author had to come up with a narrative and write each word over a period of time. But if GenAI can do it in seconds, how do we define what the value is of the end result?
Consequently, if GenAI does have this ability, and everyone has access to this technology, then there will most likely be an uptick in the amount of content that is being produced. Although quality of content is a concern here, my bigger question is WHO is going to consume all this content. If I can create 10 novels in 10 minutes, but people can’t read 10 novels that fast then who will I be creating for? Content has already been niched down and subdivided within an inch of its life. Is this how we get to hyper-specific content? We can’t even keep up with consuming the amount of content that already exists, and even if we did, it would just further fracture society since literally everyone would be consuming a different version of information.
I won’t deny that this technology is powerful and there is huge potential to use it for breakthroughs and towards positive change. However, I do believe there are many questions that still need to be answered and I’m happy to be a part of that exploration for now. Despite all of my misgivings, society never goes back on technological advancements. Even if you feel the same as I do, huge companies and governments are investing billions into figuring out how to use this. Whether we like it or not, it is going to start to permeate many different industries and parts of our lives. It is in all of our best interests to learn what we can about this technology and simply exercise caution in our own personal use.
*Can’t find the original person who said this so please let me know if there’s someone specific to give credit to)
Other resources:
Explaining LLMs in layman’s terms: https://medium.com/data-science-at-microsoft/how-large-language-models-work-91c362f5b78f
Overview of LLMs: https://arxiv.org/pdf/2304.00612
Overview of LLMs: https://aws.amazon.com/what-is/large-language-model/
Comic is by David Skaufjord (https://www.instagram.com/davidskaufjord?igsh=aXNwNzJib2lscmVl)
August Consumption Report
I’ve developed the following rating system, it’s half Michelin inspired and half of my own making.
5/5 - go out of your way to consume this, it is incredible, we will have endless discussions if you do. I’m obsessed.
4/5 - It is good. If you pick it up, I bet you won’t be disappointed
3/5 - Take it or leave it. It didn’t offend nor impress me
2/5 - It maybe has a couple redeemable qualities. Would not consume it knowing what I know now.
1/5 - Avoid at all costs. Go out of your way to avoid it.
Here are September’s ratings: My media consumption usually starts to suffer from a sharp drop off at this point in the year as the business ramps up but still got to enjoy some gems this month!
Books
The Memory Police by Yoko Ogawa (5/5)
Movies
Smile (3/5)
TV Shows
Emily in Paris S4 (3/5)
Golden Kamuy (4/5)
Kimi ni Todoke (5/5)
