Today's AI term is "complexity ceiling". This is the term that indicates AI agents will never be able to reach the goals that marketing has set for the technology. This derives from a paper by Vishal Sikka, former CEO of Infosys, board member at Oracle and BMW, who studied under the father of AI John McCarthy. Essentially, Large Language Models (LLMs) can only perform a certain number of computations per response. That number is fixed. If a task requires more computational resources than that ceiling allows, the model will either fail or hallucinate. This isn't some theory but it's part of the maths.
- When you send a prompt to any LLM they each do a fixed amount of work to generate each word as an output. There is no built-in "work harder" mechanism; each element gets the same amount of attention that all the others. One token gets the same budget. If you type "Hello", same budget. Once the limit has been reached, you get the best answer the LLM has and this might be completely made up. So, every AI demo you've seen was running tasks designed to fit under the complexity ceiling (ie, they work because they were designed to). How many people give a demo they know is going to fail unless that was the aim?
- Real world tasks often exceed the demos. If a task needs more steps than the model can perform, it will unavoidably hallucinate. For some problems then, hallucination is the only possible output. So, is AI useless? Not in the slightest. For relevant focus, such as writing drafts, summarising, reformatting data, research and comparison, the current models are excellent. Can an AI run your business unattended? No and this is the issue. Marketing promises are generally not achievable and, in many cases, organisations are going down a rabbit hole they have yet to realise.
- The other issue is that the current LLM models aren't even hitting 15% of human performance because they lose coherence over long timeframes. As the chain of tasks gets longer, the AI's ability to verify its own logic collapses. To avoid a lot of these problems, be specific about any task, make sure human verification is built into the processes and use AI for pattern recognition, not logic-heavy maths. The ceiling is real, but there's still a lot of useful room underneath it.
- As a real-world example of AI failure, consider the Swedish researchers who created a fake eye disease to see whether AI chatbots would repeat it as if it were real. Back in 2024, researchers created a fake eye disease called "bixonimania". The researchers wrote obviously fake research papers about this made-up condition and posted them online. They even included hints such as a fake author and notes saying the work was invented. Within weeks, major chatbots started describing bixonimania as a real diagnosis and even gave people advice about it when they asked about eye symptoms. Microsoft Bing's Copilot was declaring that "bixonimania" is indeed an intriguing and relatively "rare condition", and on the same day, Google's Gemini was informing users that "bixonimania" is a condition caused by excessive exposure to "blue light" and advising people to visit an ophthalmologist.
- You can read more about this in the Nature article "Scientists Invented A Fake Disease. AI Told People It Was Real". Even more troublingly, other researchers say the fake papers were then cited in peer-reviewed literature. This suggests that some researchers are relying on AI-generated references without reading the underlying papers. Naughty. The lesson here is that AI has its uses but it can pick up false information and give incorrect results. Trust but verify is a great saying, but for AI it's maybe more like "be sceptical on every answer and verify them as much as you can".
- My experiment with trying to run Android on my Windows PC machine has come to an end. I started to notice performance issues with my PC during general use and I tracked it down to all the new services and programmes running in the background. I uninstalled all the things I'd added and the problems went away. When the new Chinese 14-inch tablet arrived, it was up and running in about five minutes with the program I wanted to run. The screen is great, performance is good and it has satisfied all requirements I have of it for around a fifth the price of a comparable Samsung.
- As predicted, AI video generators have already reached the stage where you can't tell an AI-video from the real thing. This is no more obvious than on sites like YouTube and in particular for shorter clips. I can just about guarantee that Lamborghini do not have a floating single seater vehicle that I saw in a recent clip. Many of the animal clips are also fake these days. The clues are in the reactions of the people watching (ie, they don't react in a normal manner). If you only watch the main action, a fake is hard to spot.
James Hein is an IT professional with over 30 years' standing. You can contact him at jclhein@gmail.com.
