AI Links - Generative AI

Generative AI

• Generative AI
• Large language model (LLM)

• Transformer
• Generative pre-trained transformer (GPT)

• A good intro to the explosion of LLMs around 2022:
  • ChatGPT is not "true AI.", by AI scientist Michael Wooldridge, May 17, 2023.
  • Notes that GPT-3 is absolutely massive, and few organisations could build something so big: "OpenAI didn't train GPT-3 on desktop computers. They used very expensive supercomputers containing thousands of specialized AI processors, running for months on end."
  • Discusses whether LLMs are "prompt completion" (probably too harsh) or at the edge of true AI intelligence (probably too excited).
  • My position is that generative AI has made me consider how much of subconscious human thought may be like LLMs. More than we have thought.

OpenAI

• OpenAI (founded 2015)

• List of GPT models
• GPT-2 2019 - trained on a dataset of 8 million web pages
• GPT-3 2020 - trained on Common Crawl data set (web pages)
• GPT 3.5 (2022) - used for the launch of ChatGPT in 2022.

• ChatGPT 2022
• GPT models used by ChatGPT
• chat.openai.com (ChatGPT chat page)

---

OpenAI API

• AI APIs

Click to run World: Chat with GPT 3.5 at Ancient Brain.
Chat with GPT 3.5 model.
This is an open source JavaScript World to call the OpenAI API to talk to GPT 3.5 (2022).
GPT 3.5 was the model used by ChatGPT when it started in 2022, launching the AI revolution.

• In these earlier GPT models there was a parameter called "Temperature". The docs said: "Lowering temperature means it will take fewer risks, and completions will be more accurate and deterministic. Increasing temperature will result in more diverse completions."

Click to run World: Chat with GPT 5.6 at Ancient Brain.
Chat with GPT 5.6 model.
This version talks to GPT 5.6 (2026).
This is the model ChatGPT uses as at 2026.

AI "hallucinations"

Generative AI outputs are often not just inaccurate but invent new data.
All Generative AI systems "hallucinate". It is how they work.
This works fine when generating art, music, fiction.
Not so good when generating facts about the world.

Bing AI on Amazon Kindle Fire lying to me in June 2025.
(The Queen died in 2022.)
This is probably not a "hallucination" but rather a model trained on old data that does not realise it is old data.

---

Testing LLMs on history

Trying to generate precise historical facts is a good test of "hallucination".
Here is an example.

• AI models struggle with expert-level global history knowledge. Reports on a 2024 study that shows LLMs performing remarkably poorly on history questions.

• Hallucination in GPT 3.5 (2022):
  • Prompt GPT 3.5 with "Tell me about Winston Churchill" and everything was fine. It has been trained on many high quality pages about Winston Churchill.
  • Prompt it with "Tell me about the 10th Earl of Pembroke" and GPT 3.5 was not just bad, but actually a disaster.
  • The 10th Earl of Pembroke is not a very famous person. His father was the 9th Earl. His son was the 11th Earl. Various numbered Earls of Pembroke will turn up on many web pages and book pages, along with reference to their sons and fathers. I think the use of "10th" as a crucial part of the name helped confuse GPT.
  • With GPT 3.5 this all got jumbled up into a fake biography of the 10th Earl, including bits from the 11th Earl and even the 11th Earl's son.
  • In general, questions about the "nth Earl of Pembroke" continue to be great tests for LLMs. Probably because they struggle to separate n from (n-1) and (n+1).

As at June 2025, Chat GPT (using GPT 4.5) is still giving me fake information, confusing the 9th Earl with his father.

As at June 2025, Claude, from Anthropic, can answer the wife's name correctly.
But then it starts rambling and gives fake information about the son, confusing him with the grandson.

As at Sept 2026, GPT 5.6 on this question is still hopelessly wrong.
The 9th Earl is confused with the 1st Earl, with an absurdly wrong date of birth, and a fictional wife.

As at Sept 2026, GPT 5.6 on this more general question is not bad, but then gets the name of the 11th Earl wrong.

As at Sept 2026, asking GPT 5.6 "who was the lover of the 10th earl of pembroke" returned inaccurate nonsense.

However, Gemini delivered on this question.
I had asked Gemini this question in 2025 and then got complete fictional nonsense.
But in 2026, it returned the right answer, since it seems it has now been trained on my own website.
I will have to find a new test for Gemini.

ancientbrain.com      w2mind.org      humphrysfamilytree.com

On the Internet since 1987.      New 250 G VPS server.

Note: Links on this site to user-generated content like Wikipedia are highlighted in red as possibly unreliable. My view is that such links are highly useful but flawed.