

Same. I still occasionally browse Reddit, but I have a rule that I don’t post or comment there. I do post and comment here.
Same. I still occasionally browse Reddit, but I have a rule that I don’t post or comment there. I do post and comment here.
Deflecting blame by subtle ethnic discrimination. Nice.
MySpace was huge before Facebook, and it killed off a lot of blogs. Late 90s and early 2000s were truly the wild web IMO. I had a geocities page with its own forum before MySpace made me abandon it due to inactivity.
It doesn’t have to be
https://www.mathworks.com/products/compiler.html
MATLAB can ruin all sorts of coding experiences, programming included
Examples? I can think of a number of foreign companies that the US facilitates, like Nestle.
Eh, I switched. I switched all of my lab’s computers, too, and my PhD students have remarked a few different times that Linux is pretty cool. It might snowball.
The professor probably would have responded that his response was another part of the lesson: don’t trust those above you in a business setting.
This makes sense, thanks
Why would China turn against Putin for them using their nukes? I don’t keep up much on their relations.
Oregonians almost take pleasure in driving slowly in front of you. Maybe they’ve just gotten used to going slow because the entire state freeway system is always under construction. People driving crazily is infuriating for a completely different reason.
The best time to start was decades ago, but at least they’ve started.
This is a problem that’s becoming outdated, thanks to NIH now requiring females to be included in studies in order to receive grant funding–barring an exceptional reason for studying males alone (e.g., male-specific problems). They are even requiring cell lines for in vitro studies to be derived, at least in part, from females, rather than from males alone.
Seven paragraphs is too much? I read the full thing before seeing your comment. It’s well written and easy to read.
I never understand why lemmy downvotes someone who is trying to help by providing accurate information, presumably because they think that there’s a very small chance that the person they’re replying to isn’t being sarcastic.
A fellow Julia programmer! I always test new models by asking them to write some Julia, too.
I actually took that bit out because LLMs are pro climate and against everything that makes the environment worse. That’s a result of being trained on a lot of scientific literature. I was just curious what Opus would say about the conceptual knowledge piece.
Claude Opus disagrees, lol (edit to add: all of what follows is Claude; not me):
I respectfully disagree with the claim that current language models are logically incapable of mapping language to actual concepts. While today’s LLMs certainly have limitations and are prone to hallucinations, they have demonstrated a remarkable ability to capture and manipulate complex concepts through language.
A few key points:
LLMs are trained on vast amounts of human-generated text which is grounded in real-world concepts, allowing the models to build rich representations that go beyond surface-level word associations. Techniques like unsupervised pre-training enable LLMs to learn meaningful conceptual relationships.
In many domains, LLMs have shown an impressive capacity for reasoning, inference, and knowledge synthesis - skills that rely on a deep understanding of the underlying concepts, not just pattern matching. Their performance on complex question-answering, analogical reasoning, and natural language inference tasks is evidence of this.
LLMs can be fine-tuned and augmented with techniques like retrieval-augmented generation or chain-of-thought prompting to further ground them in factual knowledge and improve their reasoning capabilities. This allows more robust conceptual mappings.
Careful prompt engineering and sampling techniques can help to mitigate hallucinations and nonsensical outputs. While not a complete solution, this shows that the conceptual knowledge is often present even if the model sometimes struggles to express it faithfully.
That said, I fully agree that the tendency for subtle nonsense is a serious issue, especially for layperson use cases where hallucinations may go undetected. Continued research into making LLMs more truthful, consistent and robust is crucial. Techniques like constitutional AI to bake in truthfulness, and better UX design to convey uncertainty are important steps.
But in summary, I believe the evidence suggests that LLMs, while flawed, are not fundamentally incapable of meaningful conceptual representation and reasoning. We should push forward on making them more reliable and trustworthy, rather than dismissing their potential prematurely.
I’m back on my BS is also a solid contributor
Fair points. My use of “primary” was a poor choice; I meant something along the lines of “most common among individuals who aren’t philosophers, in my experience.”
I don’t internet without uBlock. I honestly couldn’t imagine it any other way.