> Reason #3a: Work with the model biases, not against
Another note on model biases is that you should lean into them. The tricky part with this is the only way to figure out a model's defaults is to have actual usage and careful monitoring (or have evals that let you spot it).
Instead of forcing the model to behave in ways it ignores, adapt your prompts and post-processing to embrace its defaults. You'll save tokens and get better results.
If the model keeps hallucinating some JSON fields, maybe you should support (or even encourage) those fields instead of trying to prompt the model against them.
Ha-has are perhaps tonally inappropriate, but when you look at the facts it seems unlikely. What we've seen in the last few years is fairly unlikely to continue forever. That's rarely how trends go. If anything if we actually look at the trend lines the improvements between model generations are becoming smaller, and the models are getting larger and more expensive to train.
A perhaps bigger concern is how flimsy the industry itself is. When investors start asking where their returns are at, it's not going to be pretty. The likes of OpenAI and Anthropic are deep in the red, absolutely hemorrhaging money, and they're especially exposed since a big part of their income is from API-deals with VC-funded startups that in turn also have scarlet balance sheets.
Unless we have another miraculous breakthrough that makes these models drastically cheaper to train and operate, or we see massive increases in adoption from people willing to accept significantly higher subscription fees, I just don't see how this is going to end the way the AI optimists think it will.
We're likely looking at something similar to the dot com bubble. It's not that the technology isn't viable or that it's not going to make big waves eventually, it's just that the world needs to catch up to it. Everything people were dreaming of during the dot com bubble did eventually come true, just 15 years later when the logistics had caught up, smartphones had been invented, and the web wasn't just for nerds anymore.
I guess the argument of AI optimists is that these breakthroughs are likely to happen given the recent history. Deep learning was rediscovered like, what, 15 years ago? "Attention is all you need" is 8 years old. So it's easy to assume that something is boiling deep down that will show impressive results 5-10 years down the line.
Scientific breakthroughs happen, but they're notoriously difficult to make happen on command or on a schedule. Taking them for granted or as inevitable seems quite detached from reality.
True, but given how many breakthroughs we had in AI recently, for text, sound, images and video the odds of new breakthroughs happening are probably higher than otherwise.
We have no idea how many of them we need till AGI or at least replacing software engineers though.
That's mostly just a few discoveries finding multiple applications. That's fairly common after a large breakthrough, and what you see is typically a flurry of activity and then things die down as the breakthrough gets figured out.
It's "a few discoveries finding multiple applications" plus throwing as much data and compute as possible at those applications, a process that seems to be increasingly struggling uphill in the last year or so.
When there's something new and shiny progress is made fast until we reach the inevitable ceiling. AI has unsolved issues. The bubble will eventually pop and the damages will be astounding. I will sign the HA HA HAs. People are delusional.
Currently scaling limitations where gain outweighs the cost efficiency. True long term reasoning. Hallucinations and pattern matched reasoning over structural reasoning. Reasoning for novel tasks. Hitting a data wall so lack of training data. Stale knowledge. Biased knowledge. Oh and let's not forget about all the security related issues nobody likes to talk about.
Yeah something I should have covered more is a lot of my frustration comes from the _tooling_ around formatting. Prettier is quite slow, people may not have it setup right, etc.
It doesn’t answer your question but I came across this yesterday doing some research on the R1000 (which is why I came across the OP). You might find it interesting:
Ada Compiler Validation Summary Report:
Rational Environment
https://misinforeview.hks.harvard.edu/article/do-language-mo...
reply