There's a lot of hype and anticipation for "AGI", like that is the magic moment when all of our problems will be solved. In reality, there is so much "infrastructure" that we need to build for AI to use before we can start reaping any kind of large benefits (aside from research breakthroughs made by AI). Just giving more intelligence to an AI doesn't increase how much business value it creates. If you have an AI that's as smart as Einstein, or even smarter than Einstein, it doesn't help your business a lot when it's behind a chat interface, with no access to any tools.

The current bottleneck of GPTs in general is their lack of access to tools. (See: Karpathy's thoughts on a "LLM OS"[1]). Giving LLMs access to tools such as long-term memory, searching and browsing relevant documentation, learning from human provided examples are the problems that need to be solved, to have AI agents capable of creating vast amounts of business value and replacing many so-called "bullshit jobs", and assisting with more demanding work.

This is because LLMs don't need to be smarter than humans to replace humans. Their advantage is in the SCALE. The same goes for robots aswell, but that's a whole other discussion.

On this front, the Open Interpreter project is doing great work, with their tools giving LLMs the power to control your operating system.[3]

Some practical examples

If you think about the function of customer support. Imagine that you have 20 customer support agents, all somewhat know how the product they're supporting works, et cetera. Creating an GPT who has basic access to tools, historical discussions, company docs, learning from inquiries where a human had to intervene is enough to answer to most of the "basic inquries" that all those 20 customer support agents receive. The GPT would only need to consult with a human when it is asked of something it can't help with. With a product like this, only a few human customer support agents could provide support for tens of thousands of customers, but those customer support agents would need to know the product inside and out, and understand everything on a very technical level.

On the front of programming, even open-source models such as the Falcon 7B [2] already generates quite good code with a super-low inference cost. All of these models can be fine-tuned further to idiomatically use the frameworks and libraries your code-base uses, and code in a style that matches your code-base. All you need is a senior programmer providing some feedback, create a finetuning feedback loop, and boom, you have an AI that will be on the same level as a junior developer. The only difference? This one works 24/7, and you can multiply them by paying your GPT wrapper provider some extra money. Current LLMs are "smart" enough for inquiries such as "rewrite our whole code-base to idiomatically use Next.js 14 app directory" or "rewrite all our components to use Tailwind CSS instead of stitches", but the only thing lacking is the proper tools to ground and align the LLM with your expected output. Tasks like the previous ones stated would be done in a matter of hours by the LLM consting only the inference costs, versus a week of work for the junior developer, costing a week's worth of salary.

Conclusion

I feel like many people have their focus in the wrong place. Projects such as Open Interpreter[3] show massive promise though.