Musing 102
The last few weeks of AI have been interesting. I’ve started ramping up my usage of local AI (first through ollama then through LM Studio) and I find my thoughts coalescing on where we’re going.
1) LLMs are not AGI.
LLMs, even as good as they are, do not truly reason. If they did, hallucinations would never have been a long term problem.
On top of this, the research is increasingly suggesting that there are architectural limitations to LLMs that outside of other groundbreaking research, means that while LLMs are extremely powerful tools for a lot of tasks, they aren’t the savior of mankind people want them to be.
If LLM’s can’t really reach AGI, what’s going on, why are they so useful? LLMs take advantage of the fact that there is simply too much to know and master. Especially with the ‘novel’ research being done, what is often the case is that such ‘novelty’ is not “here’s this completely unknown area with no human contribution”, it’s “here’s where humans have put together 3 separate datapoints, and the LLM can connect those dots into a triangle”. And let’s be clear, this does not diminish the progress being made. Being able to brute force so many paths of combining elements is no small feat, and when the LLMs are correct, it’s still legitimate progress being made.
And to address one thing - we also don’t know if there are other workarounds to these limitations. I think the harnessing of tools to LLMs and AI agents is a pointer to how we can navigate around some of the limitations that LLMs have.
2) LLMs being ‘true’ AGI is irrelevant to most of the other concerns around AI.
The simple reality is that LLMs being the “must be AGI” is marketing hype to capitalize on nerd money. But even setting aside those concerns, LLMs are a good tool for a lot of things. Are they perfect? No.
Personally, the more I work with LLMs, the more I think their best use case is “creat more specialized tooling for in-house work more easily”. This gets us the best of both worlds. It minimizes the chance for hallucinations to mess up data analytics (as more and more stories are coming out where this happens), but still lets us take advantage of code being cheaper to produce, while also not allowing public facing software to become legacy by default.
And I think the fact that LLMs often build output to use existing tooling is indicative of this. Why not build your specific tooling directly if you have the ability to do it yourself, and it can be easier for you to do so?
3) LLMs will get rid of some jobs, but the majority of such layoffs (currently) are AI washed
While yes, LLMs will lead to some job losses, there’s also so many other economic factors going on that it is blatantly obvious that most layoffs are not directly AI caused. At best, the majority of CEOs know that blaming them on AI is just an easy way to mask off-shoring, bad business decisions, or other factors and make line go up at the same time.
Now, will that always be the case? I don’t know. I think the early indications (ie - precipitous drops in software quality across the board, companies quietly rehiring many of the people laid off, etc) are that we will not see long term layoffs around core business functionality. Now, with it feeling like the rate of change is increasing, I don’t know if that’ll always be true. It may be - we may see LLMs hit performance limits and harnessing/other improvements show diminishing returns. On the other, we could see LLMs find ways to leverage human knowledge to legitimately enter recursive self-improvement. While I don’t think we’ll see true RSI, we might.