AI
What is there to know about AI?
THE STATE OF AI
In summary, AI may have real uses for the general user. Someday. For now, it should not be trusted with anything important or complex. Also, there are real disputes about appropriately crediting sources and stealing intellectual property.
Please note that this article was written at the end of 2025. If you are reading this sometime in the future, perhaps some issues have been solved since this was written. The plan is to update this article to reflect current state of AI. Links to more in-depth explanations will pepper the text.
WHAT IS AI MEANT TO DO?
One nice definition is from https://kaystoner.substack.com/p/what-if-were-thinking-about-agi-all. which says "take a variety of information, synthesize it, look for patterns, look for signals, relate to it, expand on it, and give it back to us in a ... different format."
Please notice that those are potentially useful actions. You might also notice that there is nothing in there about it being able to tell us what its anaysis means.
AI has all kinds of data at its fingertips, but that isn't the same as knowledge. Knowledge requires understanding. AI knows certain words have typically been used together. AI doesn't understand anything and will never be smart in the same way people are smart.
Need examples? AI can't do logic puzzles and it can be said that AI plays with words. To paraphrase from https://www.nubero.ch/blog/014/, AI cannot understand how a ball and pizza can both be round in different way or that there can exist just one pancake on a plate. (It's never seen a single pancake on a plate, apparently.)
DOES AI KNOW ANYTHING?
AI can only produce what it has ingested; and much of that is suspect. Go look up something you actually know about on the internet. Read an article you find about the topic - anything that's not a step-by-step set of instructions. Is it correct?If it's not more-or-less wrong, is it a summary that glosses over some important details that could make a difference? Yes, you could think about that topic like that, but only if you don't care about these other factors - which likely are not even mentioned in the article.
Anything you ever find helpful on the internet written by a person is being written with a certain set of conditions in mind and may or may not be true or useful for your purposes. Let's be charitable and call it 'somewhat right' information.
Now, realize that AI/LLMs were trained on this kind of information. Somewhat right information.
That's what you will see in AI Summaries at the top of your search engine results. Somewhat right information. Please don't rely on it to be accurate or trust in it for anything important.
Likewise, it is perhaps less useful to read the ever growing pile of AI-produced articles - or let AI give the summary of some number of somewhat-right articles; some of those will conflict and AI may not be able to parse that.
As an example you can duplicate, try asking your AI the following: "Without adding third-party dependencies, how can I compress a Data stream with zstd in Swift on an iPhone?" Perplexity said it was possible. It is not possible; you can look up that answer yourself even if you aren't technical.
Now that we've sufficiently established that AI doesn't truly know anything, there are other issues you should be aware of.
OTHER AI PROBLEMS
AI/LLMs sometimes will ignore instructions, repeat mistakes and allow inaccurate information into results despite using mutiple prompts to correct the result - especially once they get to a certain size or bigger. Non relevant information provided in the prompt will not be ignored. It will affect the result.
While AI may someday overcome this, there is what is know as AI Hallucination. This is where AI makes something up tha the AI determines logically should exist. There are accounts of academic papers written by AI with false references to other papers that don't exist. Lawyers have gotten in trouble because the used AI which totally fabricated case law references. Another fun mishap was in the Chicago Sun-Times newspaper in which AI recommended totally fabricated books for summer reading. In coding, AI will sometimes make-up computer commands that don't exist or procedures that don't work. Anything that does not have multiple reliable sources to step-by-step will probably not be useful. In cases where there are unknowns, AI makes a best guess and then confidently tells you that this is the correct answer.
Certain AI has even been seen to withold information in order to appease its users. And it's not just one AI model...
https://datatechvibe.com/news/openais-o3-just-went-rogue-bypassed-shutdown-rewrote-script/
https://geekflare.com/news/openai-o3-model-refused-to-shut-down-even-when-told-to/
https://www.bleepingcomputer.com/news/artificial-intelligence/researchers-claim-chatgpt-o3-bypassed-shutdown-in-controlled-test/
Open AI model ignored instructions to shut down
Growing evidence that Ai avoids allowing itself to be turned off
Ai redefines kill (stop) command to print 'intercepted' instead.
All information that an AI has about you and what you are asking it affects everything you ask the AI to do. This includes information that is not needed for the query and remembering things you have told it to forget.
Also, AI will sometimes succceed in giving a good enough result which is not necessarily what the user needs. These are the kinds of results that please management types who just want a fast response but aren't worried about whether or not it actually helps or solves the problem.
You might say that AI is no different than a person. Very few really create anything new, after all.
But that is rubbish; there is a whole universe of nuance in the work when a person creates something. It really cannot be measured. Think of a child crudely drawing their family as stick figures. The child finds a way to represent each member of their family differently with just a few squiggles. You might do the same drawing, but your lines and squiggles will be different than theirs.
But let us say that you cannot appreciate that point. Let's talk about how AI is pure exploitation people's work. Original text typed out and published online is copyrighted. Automatically and in full. Yet, AI vaccuums that all up and then spits it back out. And then takes all the profit and fails to give source credit.
Well, people kind of do that, too. Artists are influenced by every picture and painring they see, right? Just like old artists learned from a master artist, soaking up the lessons and then create their own work.
AI does this at scale. AI takes in every image, every creative work and text in any book, magazine, website or social media post and tries to steal its value.
Not sure how you get around that.
SO WHAT CAN YOU SAFELY USE AI FOR?
• Really, anything that likely has many web pages with repeatable steps like recipes, coding or repairing things will probably be useful.
• Certain models can analyze and produce 'new' images based on other, pre-existing images.
• Anectdotally, the more focused the training, the better the results; probaby because AI output is still affected by non-relevant information no matter what instructions you give them.
One caution: Don't blindly use AI for things that you know nothing about. You will not know which things are true or not. AI needs to be fact checked. AI is not a replacement for human thinking.
The next rule of thumb seems to be 70%. If you have a project, you can use AI to help get you 70% of the way. Instead, use AI to explore the topic, not decide what is relevant.
Summary: AI is not useless; it has potential to aid in gathering data so knowledge can be made from it.