“The chatbot gave wildly different answers to the same math problem, with one version of ChatGPT even refusing to show how it came to its conclusion.”
It’s getting worse. And because it’s a black box model they don’t know why. The computer science professor here likens it to how human students make mistakes… but human students make mistakes because they don’t have perfect recall, mishear things being told to them, are tired and/or not paying attention… A bunch of reason that basically relate to having a human body that needs food, rest and water. A thing a computer does not have.
The only reason ChatGPT should be getting math wrong is that it’s getting inputs that are wrong, but without view into it they can’t figure out where it’s getting it wrong and who told it the wrong info.
It’s almost certainly because OpenAI is throwing less computing power at it in order to decrease the cost.
I mean, they’ve gotta to be blowing absurd amounts of money at it. It’s not remotely cheap to build a massively complicated web service at that scale, and eventually the numbers need to start adding up. I’m sure they have several good monetization plans, but not every instance of a business attempting to stop hemorraghing money is a conspiracy. You’d be doing the exact same thing in their shoes.
Enshittification is not a conspiracy because a conspiracy requires communication and planning. Enshittification is just how idiots act when trying to make money.
And there are more and more offline GPT AIs available for free. Now everyone with an above average computer can have their own chatGPT.
I mean an “average” computer would require a pretty beefy set of hardware. I think most of the average local llama’s would run fairly decently on a MacBook without issue nowadays (that m3 is going to be a pretty awesome beast). But the quality is pretty reduced even compared to something like 3.5 which most people thought wasn’t all that great.
But really, I’m excited about researchers have access to more computer for smaller amounts (see this https://www.chatgptguide.ai/2023/07/20/worlds-largest-supercomputer-for-ai-training-is-out/) currently we have 1T models that are good, but we could pretty soon have 100T models from the open source community. Let’s see whether we can scale the hardware needs with the parameter growth so we don’t need A100s to run a decent model.
It’s still pretty rough to selfhost an LLM. You can get one that’s kind of okay on an average computer, but to get a really competitive one running locally at a good speed, you need a huge amount of RAM that is still beyond most average users (VRAM for GPU based projects).
I’ve been trying to get Vicuna going and the RAM usage is rough, 60gb is suggested, and I’ve got 64 and I think I need a lot more honestly.
Huh… so after months of being exposed to people that aren’t quite as smart as world class computer scientists and engineers, it gets dumber. Maybe it’s more human that I previously thought.
I wonder if it is in fact learning from people’s prompts; I didn’t think that was part of the operation. That’s a huge design flaw if so.
You think that’s bad? My calculator can’t even finish a simple sentence.
A single word can be a full sentence, unless answers to either/or questions are not sentences.
Or is this one of those logic things where a train is only a train when the railway engine is connected to something?
For me it’s like using a coffee machine to measure a time lapse, and then complaining that it doesn’t always yield the same time lapse.
For any question the number of incorrect answers is larger than the number of correct answers.
This is a fundamental problem, constrained by energy costs, and one that will only be exacerbated as training datasets becomes more and more tainted by generated content.