In previous years, this was discussed a lot; now I feel like it's not talked about nearly as much. People claimed LLMs would never reach AGI, but we're getting closer and closer by the day. I wouldn't be surprised if it does reach AGI within the next 12-18 months.
I know I'm going to get a lot of disagreement about what AGI is. I'm going by the OpenAI definition. For the uninformed, this is it: “(a) highly autonomous systems that outperform humans at most economically valuable work.”
I feel as if we're almost there, especially with the release of Astra. I wouldn't be surprised if Astra already can do a lot of what the average white-collar worker does.
Here's my dispute of some common claims about how LLMs won't reach AGI:
- LLM's can't learn anything new/can't edit their own weights.
This is the most solid argument IMO. My counter to this is that, for LLMs to "outperform humans at most economically valuable work.", they don't need to be able to learn incredibly new, complex things. A lot of jobs don't require the worker to learn anything incredibly new or novel after getting the basics.
- LLM's only predict the next word
This is a massive oversimplification of how they actually work, let alone all the emergent behavior that we've seen arise in them. Also, does it really matter "how" something is intelligent if it gets the job done?
- LLM's aren't creative, which is required for many jobs
I do understand this point to a degree, but recent models actually are incredibly creative. I wouldn't blame someone for thinking this if they formed their opinion on AI even 5-6 months ago.
Creativity isn't just shown in art, etc, etc, but in how problems are solved. Current models can take a problem and invent a totally new approach to solve it. You can see this all the time when using AI to develop software.
[link] [comments]