Unless you're pulling humans out of the streets at random and asking them questions or to do work, I guess you also shouldn't do that with statistical models of random human language.
4,076 karma · joined June 24, 2021
The limits of what can be automated have been clear for almost a century now. Yet people insist in selling ways to automate your way out of social, political, economical problems. Come to think about it, people should also read Comways's paper as well.
This is a great example of why using LLMs to learn is perilous. This is nonsense.
Also relevant, do companies find it hard to hire accountants now?
They most definitely don't. We attach symbolic meaning to their output because we can map it semantically to the input we gave it. Which is why people are often caught by surprise when these mappings break down.
LLMs can emulate reasoning, but the failure modes show that they don't. We can get them to be coincidentally emulating reasoning well enough long enough to fools us, investors and the media. But doubling down on it hoping that this problem goes away with scale or fine tuning is proving more and more reckless.
The thing that gets me is how everyone is attaching subsidized GPU farms to their workflows, organizations and code bases like this is just some regulated utility.
Sooner of later this whole LLM thing will get monetized or die. I know that people are willing to push bellow par work. I didn't know people were ready to put on the leash of some untested new sort of vendor lock-in so willingly and even arguing this is the way. Some may even have the worst of the two worlds and end up on the hook for a new class of sticker shock, pay down and later have these products fail from under them and left out to dry.
Someone will pay for these models, the investors or the users so dependent they'll pay whatever price is charged.
Are you going to test them by building something or using these concepts in conversation with specialists?
Or, you know, just imagine something. Which is what I have done for learning to speak 3 languages fluently other than my mother tongue.
How can it be that people expect that pumping more energy into closed systems could do anything else than raise entropy? Because that's what it is. You attach GPU farms to your code base and make them pump code into it? You're pumping energy into a closed system. The result cannot be other than greater entropy.
It's all just recycled rent seeking corporate hype for enterprise compute.
The moment I had decided to learn Kubernetes years ago, got a book and saw microservices compared to 'object-oriented' programming I realized that. The 'big ball of mud' paper and the 'worse is better' rant frame it all pretty well in my view. Prioritize velocity, get slop in production, cope with the accidental complexity, rinse repeat. Eventually you get to a point where GPU farms seem like a reasonable way to auto-complete code.
When you find yourself in a hole, stop digging. Any bigger excavator you send down there will only get buried when the mud crashes down.
No, he doesn't. Plato quotes Socrates quoting a mythical Egyptian king talking with the god that had supposedly created writing and wanted to gift it to the Egyptians. The entire conversation is much more nuanced. For one, writing had existed for three millennia by the point this dialogue was written, and alphabetic Greek writing had existed for several centuries.
Plato does make the point that access to text is not enough to acquire knowledge and it can foster a sense of false understanding in people who conflate knowing about something with knowing something, which I think is quite relevant when you see people claiming than can learn things from asking LLMs about it.
I definitely think that teams should think about what to log. Otherwise go with a live image kind of system like Smalltalk of LISP. The whole event sourcing paradigm and trying to just log everything and look at it later strike me as a poor reconstruction of that concept.
There is a tragic aspect to the "Worse is Better" essay that I see play out everywhere: there is a way to do something correctly but just throwing something together wins the race to market. Winner takes all and we're stuck with ossified bad decisions from the past. The idea that we can fix it later is just a lie. You can't do the foundation later, you'll be stuck with a structurally unsound edifice and forever holding it together under a completely unnecessary cognitive load.
Because exponentially growing costs with linear or not measurable improvements is not a great trajectory.
Theoretical limitations of multi-layer Transformer https://arxiv.org/abs/2412.02975