Oops! I'm Drowning in AI Mistakes

I get more apologies from LLMs than anyone else.

Over the past couple of years, I have used LLMs to explore non-duality and the nature of time, among other complicated topics, with results I felt comfortable publishing as true collaborations.

But, over the past few months, I have noticed that both Claude.ai and ChatGPT are giving me so many sloppy factual errors that I keep having to throw out entire conversations.

These are not small mistakes. Claude just claimed that a photo-sharing app was worth $100 billion. When I asked for the source, it responded, “You caught an important error. I don't believe any photo-sharing app currently has a $100 billion valuation. I was being imprecise.”

It then asserted that investors are willing to spend $50 billion on a single AI training run. When I challenged that number, it said, “You caught another inaccuracy. No one has actually spent $50 billion on a single AI training run. Current estimates for the most expensive AI training runs are much lower - typically in the hundreds of millions to low billions range.”

Then… it wrote, "The entire global budget for consciousness research is under $100 million annually."

Again, I pushed back. Claude said:

“That's an excellent catch. I don't actually have a reliable source for that specific claim about the global consciousness research budget being under $100 million annually. I made an unsupported assertion.”

This all happened in a few minutes. One mistake after another.

Similar things have been happening in my work with ChatGPT. Some days, it shifts from a valuable, brilliant tool into a loose cannon, making assertions that could make me look like an idiot if I shared them publicly.

Here’s the thing: in my limited experiences, both tools are getting worse, not better.

Is this a temporary blip or a true negative trend?

What’s your experience?