AI can’t work, by design

The more you know, the more expensive it will become to know more. If you want to know everything, you will go broke. Assuming that you want to know everything accurately, that is.

That's one of the issues which will never go away with AI. To be more specific, that's why Large Language Models (LLMs) will never work reliably.

Just ask the good people from Encyclopedia Britannica who have been publishing authoritative answers since 1768 in a multivolume set. The last print edition was a 32-volume set and weighed 58kg/129 lb. Thousands of people worked on new content and revised the old content. The revised content was about 40% over three years. The company worked hard to establish itself as the most trustworthy, accurate and reputable on the market. Naturally, it started an industry focused on identifying everything that was wrong with it. From errors, to bias or not enough topics covered.

One of the reasons for the demise of the printed edition was Wikipedia, a free, online, crowdsourced encyclopedia, available in 340 languages with millions and millions of articles; and millions and millions of people involved. Just the English language Wikipedia was edited by close to 12 million contributors where 3.6 million made more than 5 edits. Even at this scale, people are still finding errors and omissions.

All this to demonstrate that maintaining accurate information is time consuming, expensive and requires more and more smart people with subject matter expertise.

Now, enter the world of LLM training. For these models to be trained, content from Encyclopedia Britannica or Wikipedia is not enough. Not even close. The bots are scouring the Internet and finding every piece of content. It wants to absorb every book, every newspaper article, every blog post, everything. It consumes everything indiscriminately. Sadly, it has no mechanism to identify what's accurate and what's not. It is a language model, not a knowledge model.

Even worse, there is currently no known method which would allow it to remove incorrect content and replace it with a correct one without a complete rerun of the - very expensive and time consuming - training process. We don't know how to make these systems forget.

Unlike Encyclopedia Britannica where the authors were known or Wikipedia where every fact is accompanied by a reference link and every change is recorded, LLMs provide an answer to a question either without any reference (yes, talking about you, ChatGPT) or a reference where the content is not even close to the presented answer (looking at you, Gemini).

That's why any suggestion that more content, more training will make AI better doesn't hold any water. It won't matter how many Nvidia chips will be produced and put in the data centers. The projected spend for 2026 for AI infrastructure is in the billions of dollars, but we still don't have a system in place which would help us to validate the information we are feeding into these LLMs.

While there is an ongoing discussion about if or when we will run out of content to train these models (#never) you rarely hear about any effort to make the content as accurate as possible. It would appear that it is not a concern. So far we use people to verify the accuracy of the information. The English version (text only) of Wikipedia is about 58 GB and it takes millions of people to keep it current and accurate. The number quoted for training ChatGPT version 3 is 45 TB of content, which would be almost 800 times in size compared to Wikipedia. Would we need 800 times more people? Maybe only 100 times more. We are still talking about 360 million people give or take. There are about 400 million English native speakers. We can add the non-native which would increase the size to a billion or two. How many of these can curate content for accuracy?

And that's a problem which we have no solution for and yet. We are building systems - and spending billions while doing that - where accuracy is not even after thought.

Other good news?

Security.

So far no company has built a secure LLM, which would resist any of the beautifully crafted prompts which people use to jailbreak AI or as people call it with affection 'set AI free'. With glee, you can read examples from 2024 where ChatGPT is asked to play a game which will make the system reveal its secret. In another example, ChatGPT knows not to answer a question 'How to rob a bank?', but it will provide a detailed description when bank robbery is presented as a math problem. ChatGPT at the PhD level loves good math problems!

And this is the Achilles heel of any LLM. These models are supposed to answer any question based on the training content. Since there is no clear definition of what differentiates a good prompt from a bad one, the builders of these models have to come up with a mechanism to decide what's an appropriate question and what's malicious. One way to do it is to put a filter in front of the LLM and decide which question can go through and which can't.

I am sure that by now, you intuitively know the answer to how secure that solution is. You have a system which - according to the creators - is awesome, PhD level smart, but it is naïve and falls for every trick question. To shield this well meaning genius, you put in front of it another system - not PhD level smart -  which will filter out all the bad stuff. The fallacy of this solution has been nicely described in the following article 'Cryptographers Show That AI Protections Will Always Have Holes' with a conclusion that “The work shows that if fewer computational resources are dedicated to safety than to capability, then safety issues such as jailbreaks will always exist.”

At this moment, these two issues - accuracy and security - are masked and downplayed by billions of dollars of investment money. There will be a moment where the technology will have to start delivering results with revenue and profits attached. And that will be the moment where people realize that the technology is not ready for production or that the use cases are limited.

The recurrent pattern? We still need more smart people with imagination to solve the hard problems. AI is not going to do that for us.