The Issue with AI

This is my newsletter issue that covers AI. But I also mean it the other way.

This issue contains about 1,800 words. It features 26 links that lead to pages with more words and medium-length videos. What I have picked for this issue comes from many months of reading about AI. I spent the time to read each piece, think about them, forget about them, then go back to them to see if they are valuable. I then picked the ones that I thought are worth your attention.

So if you are considering using a language model to summarize this material, please don't. Because that instinct is the part of the problem. Here is why:

AI of today is more buzz than technology

The field of artificial intelligence is amazing. For many decades, devoted scientists worked hard to tackle crazy complicated math and philosophy problems to advance computational intelligence. They worked on many competing and at times contradictory approaches to unlock the AI technologies that we use today. We should recognize them and thank them for their efforts.

Watch this 17-minute video with patience and attention to appreciate the level of sophistication these scientists have been working at for decades. Please see for yourself that there was significant AI work many years before ChatGPT.

But today, we talk less about the fantastic scientific endeavors that made today's AI systems possible and more about the buzzy side of AI.

Do you realize that we call everything AI today? Detecting what is in an image used to be called computer vision, now it's AI. Creating layer-based computational models for processing data used to be called neural networks, now it's AI. Processing data for predictions and insights used to be called machine learning/deep learning, now it's... Heck I saw an AI-powered muffin the other day.

Why do we do that? Because the companies and venture capitalists who have poured millions into an important but consequential step in computer science wants to pump this progress into a frenzy. People in power needs a reason to raise and spend more. Someone needs to pay. That's us.

Intelligence in "AI" isn't the same intelligence in humans

The intelligence in generative artificial intelligence isn't like human intelligence, cat intelligence or even fish intelligence. These models do not think. Instead, they predict likely outputs based on inputs and parameters through intricate, complex mathematical models.

There is a wonderful term for this, coined in 2021: Stochastic parrots

To be able to parrot at a level comparable to a human speaking, these models need tons of data and processing power. However, even the state-of-the-art algorithms for processing these data do not create new reliable data. With each query, we get calculated guesses about the most proper response to the topic at hand. The models present the most plausible, common, and sometimes completely false information because they know how to parrot quite well.

That kind of intelligence isn't good for us

There is no doubt that the current generation of generative AI technologies are beyond laypeople's imagination. I mean, some technologies in Black Mirror are available to download. With a rigorously crafted AI toolchain, you can almost work in OS1 - the operating system in the movie Her (They even had Samantha's voice...)

Even without that level of sophistication, LLMs can be useful. As of this writing, they are quite good at carrying out simple tasks that involve low-risk data manipulation. They quickly create empty, soulless blog posts; they produce reasonable boilerplate code; they create images that can be used stand alone or patched into other existing images.

If you are an experienced writer, coder or graphic designer, that means that a lot of the mundane boring tasks can now be done with a computer program, instantly! Yet, these features are pumped to those who have no idea about the task being automated by the AI.

There is a cringe thread on reddit about an aspiring AI user who lost his programming work because "AI ate it". He was sold the vision (read: bullshit) that AI could code for him without him having to learn coding. But he lost his work with one pass, why, because he wasn't using version control. In essence, he was trying to run a marathon without knowing anything about hydration, shoes and breathing.

Having a computer program quickly write unit test cases for thousands of lines of code sound appealing, but that takes away the experience one gains while they are learning to write these unit tests. The thing you create when you write a unit test is just a unit test, but doing that teaches you about program logic, clean code, dev environments, maintenance and even power plays between developers, influence of architects, product management... Stochastic parrots do not value these by themselves, they don't learn from these valuable moments by themselves; we humans do.

There are very informed, strong opinions and research about how such automation may impact our cognitive abilities in the long run. Here is a table summarizing possible effects from one of those papers. Here is a recent paper on how Gen AI tools nudge us away from problem-solving and critical thinking. If you are not comfortable with academic writing, here is a Forbes article covering the same topic. Some effects are already here with us today, such as our willingness to be influenced by AI and to perpetuate harmful stereotypes.

Greed.

There is so much greed in the recent AI zeitgeist.

Companies that are investing in AI, especially those who are training models, need a lot of compute power, which implies electricity use. That greed is so intense that they are willing to reopen catastrophic disaster sites for gain.

The capabilities of the models are very carefully controlled by the companies running those models. In a sense, this is good. We don't want a regurgitation machine to talk to elderly, distressed, curious but naive people about topics that could harm them. But those limits seem to be arbitrary and quite unstable.

And crossing those limits can lead your loved ones to take their lives. And worse, these companies could be so evil that they claim their recklessness with harmful technology is a democratic right. (I'm sorry to say this but the Great America will see more of these...)

Companies know about these limits and exploit them to keep you in their userbase at all costs. Wouldn't it be great if we had a tool that exposed LLM-generated content, so that we could get more genuine writing? Of course it would be, that's why OpenAI won't give us that tool. All of the questionable things that are fed to these models and all of the questionable things that these models generate go through real humans, and that is devastating.

And here is the most concerning part. Geoffrey Hinton, the Godfather of AI, articulated these dangers at a societal level in the clearest terms. He left his supremely senior and influential role at Google due to their (and the industry in general) recklessness. Here is the detailed account.

If you are not seeing the greed here, you need to look closer. If you are seeing it and still using AI extensively without these considerations, you are a part of the AI hype problem. If you are comfortable about your stance, I am assuming that you are also a part of other problems in the world.

Cassie said it best

On Monday I saw a LinkedIn post from Cassie Kozyrkov. She summarized the current capabilities of Gen AI systems perfectly:

• AI only sees the pattern, not the purpose.
• AI only sees what you wrote, not what you thought.
• AI only sees the data trail, not the human story.
• AI only sees what was implemented, not what was considered.
• AI only sees the final decision, not bolts of inspiration.
• AI only sees what worked before, not what will work next.

So what do we do?

Design Implications: Generative AI is like plastics

AI technologies, especially LLMs are like new materials that we can use in our digital systems. I liken their potential to the invention of synthetic plastics.

Wide availability of plastics revolutionized industrial design at the beginning of the 20th century. Up until that point, the designers had to consider the nature of the material as a primary factor in their designs. With plastics, the designers could relax the material constraints considerably and choose to focus on form first and foremost. This relaxed constraint opened up amazing new possibilities and unlocked new realms for industrial design.

But years have shown us that plastics weren't the greatest thing that has happened to us. They did change the world, for sure, but not all for the better. We got serious pollution and deadly health issues due to the enthusiastic and greedy use of plastics. We will get serious pollution, social inequalities and cognitive issues if we overuse Gen AI technologies.

Like plastics, we have a choice.

• We can choose to be mindful about when to use LLMs or not. Want to tap into the vast common thread about a topic in an interactive way? Use an LLM. Want to learn about something specific that is based on facts? Use an encyclopedia, use a search engine.
• We can supplement our work with LLM tools as statistical tools. We should do the work that enhances us, and we can leave the low-value task to the computer. Don't forget that just because an LLM produces an output doesn't mean that it did a proper job.
• We can contain and limit AI tool usage by customizing the AI workflow, instead of blindly accepting the workflows pushed to us by the main corporate entities. Designers at Pentagram did this for a government site and the results were promising. Unfortunately, the page has been Musked recently, so please check this blog and Pentagram's process page.
• Spend time understanding the ecosystem from multiple perspectives. I found a good start here at AI Commons. I would love to read/participate to other forums that you know.

Use AI in moderation, responsibly, and have a great week.

If you are curious about the history of plastics and its impact on design and society, and if your first instinct is to ask ChatGPT (or alike), maybe you should scroll up and read again 😑