Today, I’m going to have a short post. Recently the new OpenAI GPT-3 ChatGPT came out. It’s absurd, it mimics lifelike conversation so closely that its uncanny. But this isn’t the only model that came out. There’s also Meta’s new AI Cicero, which can play complex multifaceted games like Diplomacy. It’s absurd and so cool - it looks like to some extent, most information tasks are going to be somewhat automated.
Currently with a lot of these tools the limitation is the interface: powerful models with a lot of predictive capabilities, who are however locked up in obtuse python packages or user unfriendly research repos. But the rise of new AI startups like DeepGram, or platforms like HuggingSpaces, the future appears to be a lot more user friendly, and consequently more powerful as these models work in concert. We could even see in the future something like GPT-3’s code editing capabilities be used to improve the code of itself, while another AI model that reads from it manages the compute for the model, thereby creating a recursive system that could exponentially improve itself to be better. Search itself could be revolutionized too by having an “answer engine” replace the current “search engine”. Instead of having a menu of results to choose from, you could get an answer to your question immediately, and a curated link to select. It could mean the death or rebirth of companies like Apple.
There is currently a space tho for a more dumb user friendly library of AI models. Hugging Spaces is too technical, and a lot of papers describe a model which takes quite a few setup steps to do. Maybe some startup will create a library like this, and in the future 2084, like how you check out a book, you could checkout or buy a specific model from an online platform for your exact needs.