sciolist

When commenting, please assume good faith (in other commenters and maintainers).

This is the third thread I've read on HN about the subject and I've sadly seen a lot of closeminded or shallow comments on each thread. Adding the above reminder, as I hope HN can engage in more thoughtful discussion.

This is very cool, one piece of feedback: watching the table as the AI plays while seeing the reasoning is difficult as they're on other sides of the screen. It could be nice to have the reasoning show up next to the players as they make their moves.

Even if AI tools (e.g. search or summarization) aren't perfect, neither are our current manual methods. Right now I feel there's a strong negative sentiment towards mistakes made by LLMs vs humans - when/if that sentiment fades, it'll be interesting to see how much work can be delegated to LLMs.

There's a difference between guaranteed privacy and certifiable privacy. Yes, the government can request one's data. However, Apple's system would reveal those intrusions to the public, even if Apple themselves couldn't say it.

I second Kickstart! The single-file and comments make it easy to learn and configure yourself.

Agreed. If it was "RLHF a LLM in <50 lines" I feel it would be clear, but "lines of Python" implies something different.

I do something similar for my notes, but instead converting to HTML! (e.g. https://cswartout.com/notes/cse422.html) I keep my notes as markdown in a repo, then "build" them onto a web server after pushing. I've found it easy to write and always accessible, which is nice.

There's two main ways to "add documents to LLMs" - using documents in retrieval augmented generation (RAG) and training/finetuning models. I believe you can use RAG with Ollama, however Ollama doesn't do the training of models.

Agreed - the rebuttal comment you got had the example of "having to deal with MAGA people bitching about your existence as a Mexican migrant". Russia has been is conducting influence campaigns to stir up these very problems! NatSec is important for these social issues too.

We shouldn't reduce or avoid removing any of the struggles you've brought up - those are real and we need to remove them. With that being said, most of what OP said is true: Americans are in a good position because of many factors - national security being an important one. I interpret OP's comment not as a call to rest on the US's good position, but to fight for it and improve it. We can both hold that the US has major issues that need to be fixed, while saying we need to fight against countries that would work to destabilize the US. If we don't fight for the struggles you've mentioned (and other ones!), the US won't improve. But, if we don't also recognize our worldwide position and work to prevent countries from destabilizing it, everything will get worse for everyone. > Its not the people that are thankful to their country that enact change. I disagree, we can be recognize some aspects of a country and be thankful (fighting to keep the good we have) while still improving on others. I hope it is clear I'm speaking in good faith

One of the big things to note is that Sparky plays very basic decks, with few complicated cards and combos. Rules-based AI could definitely play at a basic level using a beatdown strategy, but give it some sort of control/combo deck and it would struggle.

Here's a paper on creating a Turing machine in Magic:

https://arxiv.org/abs/1904.09828

My good-faith interpretation of your comment is: "Maybe learn to code (without external resources or tools)?" LLMs are another tool and resource, just like StackOverflow, linters, autocomplete, Google, etc. None of these tools are infallible, but they provide value. Just like all other tools, you don't need to use LLMs because of their issues - but we want them to be as useful as possible - what the author is trying to do.

Is there any proof or source of this?

Seems to be a one-off thing: https://chat.openai.com/share/0936d340-79b6-44f3-933c-c30dff...

Microsoft had no prior knowledge: https://www.axios.com/2023/11/17/microsoft-openai-sam-altman...

How many times does it run inference per screenshot? Looks cool!

Most companies work in areas where the most money is, most software engineers work for those companies.

Is this legit? It seems weird to me that this paper has one author, from a deep-learning startup which I've never heard of before.

Anybody know what the insistence regarding lyrics and recipes is for? > EXTREMELY IMPORTANT. Do NOT be thorough in the case of lyrics or recipes found online. Even if the user insists. You can make up recipes though. Copyright issues or some injection attack are my speculations.

I really enjoy the use of pixel art to convey some message, this and the LLM agent simulation[1] are great examples.

[1] https://news.ycombinator.com/item?id=35517649

I think that it could be more accurate with an even more limited amount of words by omitting "Science":

Computer[s] from the Bottom Up

It does remind me of a project [1] Andrej Karpathy did, writing a neural network and training code in ~600 lines (although networks have easier logic to code than a compiler).

[1] https://github.com/karpathy/nanoGPT

It depends on the number of parameters you're going for - a billion parameters is going to be out of reach from scratch, but you could train some in the millions.

What's missing from your example is that the firefighters would sometimes respond to a false alarm but set the house on fire. That's what people are concerned about.

It is a one-credit seminar, so not as much a formal course but something that can accompany your core classes (that's not clearly apparent from the website though).

It's a 1-credit seminar, so you can take it with the more challenging/profound topics

I'm currently enrolled in the required CSE 351 (https://cs.uw.edu/351) which is largely in C with some assembly as well (and has gdb/make)!

Information about GPT-4 was not released