Alex

With the AI model struggling to create a 16 hour clock, It’s important to assess the implications of your response. The LLM initially failed to create something outside of its training material. This distinction is very important because it is the difference between reasoning and copying. This shows the future that we have in store for us with LLM-based AI which I’ll touch on later. You then cajoled it through prompting to sort of create the thing outside of its training material. But is it accurate even after cajoling it multiple times? No. For example, the hour hand is about 80% of the way to 14 but the minute hand is pointing at 2. To be logically consistent, the minute hand should be around 11. What did the LLM do to finally give you what you wanted? The 99.99% of training material that it was trained on acts as a huge "gravity well" pulling it toward drawing a 12 hour clock. Your cajoling through negative prompting, "No, stop it. I want a 16 hour clock, not a 12 hour clock" pulls it into a "16" token and "linear sequence" token away from its training material of a "12 hour clock". This has many ramifications to real-world results. One, this shows that LLMs really struggle to go beyond their training material. And the hyperparameters that allow them to do this (ex: top-p, top-k, temperature) have horrible side effects - hallucinations. This flaw has been in LLMs since they were first created and will remain forever. It is an immutable side-effect of the technology. Two, this effect is easy to spot in a flawed 16 hour clock. It is very difficult to spot in flawed code or a flawed legal brief. Three, Google had a symbiotic relationship with content creators. A User searched, Google provided results driving the User to a Content Creator's website. User won, Google one, Content creator won. AI Companies have a parasitic relationship with Content Creators. AI Companies steal all of the content and host it themselves. In the short term, Users and AI Companies win, Content Creators lose. Some people hand-wave this problem away by saying that LLMs will get so good they won't need Content Creators anymore. But your response shows that AI Companies will always need Content Creators. LLMs can program well because sites like Stack Overflow taught them how to program in a Problem->Solution format. Today, Stack Overflow is effectively dead. LLMs cannot learn by reading the documentation and coming up with novel solutions. We only think they can because they've been trained on actual thinking. How will LLMs stay up to date when there's no more human-generated content to train on?