The Large Language Centipede

First gather a corpus of text. Let's call it the entire public internet.

Next train a Large Language Model (LLM) against this corpus and you will get a system like Grok, Gemini or ChatGPT which approximates Artificial Intelligence (AI).

Now sell this LLM to people who use it to generate content they later post to the internet.

---

The moment you make available the first generation of your LLM, you will immediately begin developing the second generation of your LLM in order to remain competitive in the AI arms race.

---

Gather a new corpus of text. A fresh snapshot of the entire public internet.

Train your new LLM against this new corpus, knowing that it contains some non-zero percentage of LLM generated content produced by users of the first generation of your product.

Now sell v2 of your LLM to more people who use it to generate more content they later post to the internet.

---

You can see where this is going.

---

Each new generation of your LLM is trained on a higher percentage of LLM generated inputs as the last.

This leads to Macro self-training where large language models are inadvertently affecting the quality of successive generations by feeding their own outputs (and the outputs of other models) into the inputs of the newer generation of models.

However inaccurate LLMs are today, we should expect them to become more deranged and hallucinatory over time. Not less.

---

When developing the nuclear bomb, some scientists worried it might ignite the atmosphere and end the human race.

This concern did not stop them from detonating the most destructive device ever constructed.

If you think legitimate concerns regarding the safety of AI will be heeded, you don't understand the macroeconomics. These multi-billion dollar models are going to choke on their own excrement spraying trash across the entire internet and crash the global economy full speed into the brick wall that is The Ouroboros Problem.

Comments

[Closely]

Nice!Recommended this app!