Anon Taranis writes that LLMs are, in large part, a failure and that all AI companies have hit a wall that they’ll never surmount.
I believe it is always good to read both sides of debates. If I find myself overly positive or negative about a topic or technology, I seek out those that disagree. Hopefully, this has helped me to have a more balanced opinion. I don’t know.
There are many parts of this post I agree with and a few I do not. Here is one I agree with:
A good example is transformers used to assist in programming, or to generate code from scratch. This has convinced many non-programmers that they can program, but the results are consistently disastrous, because it still requires genuine expertise to spot the hallucinations. Plausible hallucinations in code often result in really horrible bugs, security holes, etc., and can be incredibly difficult to find and fix. My own suspicion is that this might get you close to what you think is finished, but actually getting over the line to real production code still requires real engineering, and it’s a horrible liability to have to maintain a codebase that nobody on the team actually authored.
I agree with the above and wrote something similar recently. This is what I wrote:
In my experience, LLMs are very good at helping me with my job but they aren’t very good (yet) at doing my job. Most code written by agents (meaning, LLM tools that have a bit more autonomy to do more than just suggest code updates) takes nearly the same amount of work to fix than it would have been if you wrote it yourself. It also has the added drawback of the programmer not being intimately familiar with the codebase. Which, in the longterm, could be a real issue. But perhaps this will be improved upon and go away and we’ll never need to see code again? I’m not sure.
Taranis post largely makes the case that the underlying technology of modern LLMs (transformers, primarily) are a dead end. That the initial progress made by throwing enormous amounts of compute at the problem has stalled and won’t improve even if, somehow, you could throw ten times as much compute at it again.
I’m not an expert enough in this field to know if that is true or not. But, I think progress may come from areas yet unexplored. It is really hard to say that something will never make progress. Never say never, they say. So while it could be true that the current technology for LLMs has already peaked, we may very well see an entirely different approach emerge soon enough.