Why I’m Skeptical on AI

Remember "The Cloud?"

These days, it's impossible to escape hearing about advancements in AI. The more stories I hear about AI taking jobs and changing everything forever, the more I'm reminded of similar stories about the cloud. And in case you're still confused: "The Cloud" is a marketing term for "Other People's Computers." It referred to a paradigm shift of more and more companies leveraging paying others to deal with all of the day-to-day costs of running servers, and those companies who sold those services were able to utilize economies of scale to make it so that their customers had (some) financial savings (sometimes) in doing so. But also, lest we forget, that everything was going to be “on the blockchain”, and that went nowhere fast!

I recall back in 2020 when I learned and tried out a game called "AI Dungeon," built on GPT-2, and did a deep dive on how those AI learn and work. I have since come away with an incredibly pessimistic view of "AI" (specifically general-use case models like ChatGPT), and the more I've learned about how they work and seen how they're shoe-horned into everything, the more it’s cemented my opinion. Chatbots and the summarized options you see on top of Google searches are all made with a technology called a Generative Pre-trained Transformer (GPT), which is a type of Large Language Model (a broader term encompassing GPT, as well as the things used along-side it to improve accuracy).

GPTs

GPTs work by creating a simulated neural network, with multiple layers of virtual neurons, and pushing pre-selected training data through it and measuring what comes out. They do so with random connections between the neurons, and random strengths to those connections. They then run these tests millions or billions of times, slowly picking the best performing connections until they have a model that performs the best overall. Not the best in every scenario, but the best average. It's because of this aim to be a jack of all trades that it's a master of none.

Now, don't get me wrong: GPTs can be incredibly effective at certain, specific tasks when they are trained on an incredibly specific data set that is human-verified, such as a recent Harvard-produced model that can analyze biopsy slides and identify cancer with 94% accuracy, beating out humans. In some of the tests, the AI was able to detect cancer that humans couldn't see yet, just by picking up on almost imperceptible differences in how slides look.

All general-use AI is trained hands-off with huge dumps of stolen and unethically scraped data. And because there's just too much data to confirm everything being poured in is valid, true, and unbiased, the end result is impossible to make accurate. Garbage in, garbage out. And now there is an issue with malicious actors (or other AI companies) seeding the internet with data they know will break AI models. This is also the case in image generation, which is why all AI generated images tend to look very similar.

Hallucinations

If you've seen much about AI, you've likely heard that they "hallucinate" and come up with wrong information. The "hallucination" problem is a misnomer, because fundamentally, everything an LLM outputs is a hallucination, it's just people have assigned a label to when it's obviously incorrect. There's still potential for inaccuracy derived from biases in the training data, as well as from prompts being misinterpreted. Bias is almost to be expected in AI where data is being pulled from random forums, Twitter, and Reddit. Especially when they're not able to pull the context of any conversation or account for non-sequiturs, sarcasm, or reference-based humor.

The litany of "AI can program" posts I've seen highlights another problem, the Gell-Mann amnesia effect. It was coined by Jurassic Park author Michael Crichton to describe those who would read a newspaper, find an article about a topic you're knowledgeable and passionate, spot numerous errors, and then turn the page and trust what you read there is accurate. Unless you know what the answer is supposed to be, you can't be certain whether what an AI spits out is accurate.

You're Not Having a Conversation

AIs that carry on conversations do so via a "context window." The program that you interact with online is read-only and is incapable of learning or remembering anything, by design. When you have a chat that continues past a single prompt, it works by feeding in a rolling window of the previous chat history along with your new prompt. This is limited, as the AI gets confused if you dump too much data into it. So if you've ever chatted with an AI and seen it forget things you've said or the context of the conversation, that's why.

Which, if you think about it, makes sense. Most high-end AI models use upwards of 10 thousand artificial, simplified neurons for training. The average cockroach has over 1 million neurons. The "intelligence" of LLMs is fundamentally an illusion based on good training and taking raw output and passing it through filters to spell check and make sure it's writing complete sentences. We won't get better results from this type of AI, and either a fundamentally different model needs to be thought up, or a massive (minimum 10,000x) increase in computing power needs to happen. OpenAI and other researchers have proposed using multiple types of LLMs together to tackle the shortcomings of each, but ultimately they're all still trained on garbage.

The Cost of AI

Generalized AI (ChatGPT and similar image/audio/and video generators) uses more energy per year than the Netherlands, and AI data centers consume disproportionately so much energy from their local grids that it's damaging appliances and equipment to the tune of billions of dollars per year (950 MW of power in Chicago goes to AI alone btw). These stats are predicted to rise to up to 3% of global energy consumption by 2027.

AI is currently in a bubble. Nvidia, Microsoft, Google, ChatGPT, Amazon, etc. have spent a collective $1 trillion on AI, because a rampant tulip mania convinced them that it was the "next big thing" after "the blockchain" failed miserably.  And now they have to justify that expense to shareholders, which is why they're cramming it into everything and creating massive smoke-and-mirror shows to convince the average consumer that it's the key to the future. They fundamentally know it's a dead-end tech, but until someone else invents the next big thing they can copy, they're stuck.

Now with AI!

I've also seen a trend of companies taking algorithms and rebranding them as AI, when in reality, they aren't doing anything different. I used to use a fitness app called BodBot and I paid for a lifetime license ages ago. If you look at the app now, it says "AI workouts," but if you look at their site back before 2020, it bragged about their algorithms.

Overall, AI is a collective noun describing a host of technologies and products. In small, specific use cases, it can be a useful tool for summarizing data and finding patterns or for providing work shortcuts (voice transcription, for instance). But when it's used blindly or without actually thinking through the purpose and ensuring that the limitations of tools are kept in focus, it becomes a nightmare waiting to happen.