OpenAi's promise though, isn't the models they release as is, it's how easy it is to fine-tune them for your needs.

So if you wanted to create a dataset on Conspiracies, you'd only need to compile conspiracy texts, and train overnight to a couple days on a personal computer, as opposed to having to compile thousands of varied texts and gain for weeks or months on a 'supercomputer'.

Fine-tuning an OpenAI model is like teaching a person a new word. OpenAI training a model from scratch is like having to give a cochlear implant to someone in order to teach them what sound is so you can try to teach them spoken language.

It's only as good as the data it is trained on, so that is to be expected. Compiling and formatting data is the bulk of the work when training models.

OpenAi would have to dedicate manhours to prepping datasets specifically on conspiracy theories.

And even then, 'AI' is barely more complex than autocompletion. 'This word comes after that word 43% of the time in association with some subject.'

OpenAi's promise though, isn't the models they release as is, it's how easy it is to fine-tune them for your needs. So if you wanted to create a data set on Conspiracies, you'd only need to compile conspiracy texts, and train overnight tona couple days on a personal computer, as opposed to having to compile thousands of varied texts and gain for weeks or months on a 'supercomputer'.

Fine-tuning an OpenAI model is like teaching person a new word. OpenAI training a model is like having to give a cochlear implant to someone in order to teach them what sound is so you can try to teach them spoken language.