For real, it almost felt like an LLM written article the way it basically said nothing. Also, the way it puts everything in bullet points is just jarring to read.
Hi! I am Creesch, also creesch on other platforms :)
For real, it almost felt like an LLM written article the way it basically said nothing. Also, the way it puts everything in bullet points is just jarring to read.
True, though that isn’t all that different from people doing knee jerk responses on the internet…
I am not claiming they are perfect, but for the steps I described a human aware of the limitations is perfectly able to validate the outcome. While still having saved a bunch of time and effort on doing an initial search pass.
All I am saying is that it is fine to be critical of LLM and AI claims in general as there is a lot of hype going on. But some people seem to lean towards the “they just suck, period” extreme end of the spectrum. Which is no longer being critical but just being a reverse fanboy/girl/person.
I don’t know how to say this in a less direct way. If this is your take then you probably should look to get slightly more informed about what LLMs can do. Specifically, what they can do if you combine them with with some code to fill the gaps.
Things LLMs can do quite well:
These are all the building blocks for searching on the internet. If you are talking about local documents and such retrieval augmented generation (RAG) can be pretty damn useful.
You are glossing over a lot of infrastructure and development, when boiled down to the basics you are right. So it is basically a question of getting enough users to have that app installed. Which is not impossible given that we do have initiatives like OpenStreetMap.
It depends on the platform you are using. But, for platforms like github and gitlab there are extensions for popular IDEs and editors available that allow you to review all changes in the editor itself.
This at the very least allows you to simply do the diffing in your own editor without having to squash or anything like that.
At least for the instance this was posted on: the February 2024 Beehaw Financial Update
Well, the problem is you don’t know what you don’t know.
This is true, even recognized in the paper. People that spend more time on writing prompts (probably knowing that this is important) actually did manage to do reasonably well. Which is exactly what I in the previous reply was hinting at.
Because, let’s be honest, this statement is true for everything where someone starts out new. In the past (and probably still) you had people blindly copying code blocks from stackoverflow not understanding what the code or realizing how outdated the answer might be.
So the key is still education of people and making them aware of their limitations. You can try to block the usage of tools like this, some companies actively do so. But people will be people and as long as the tools are available they will try to use them. So the more realistic approach, in my opinion, is to educate them in the usage of these tools.
For LLM training I do wonder if they assigned a weight, but I doubt it.
Given my experience with models I think they might actually do assign a weight. Otherwise, I would get a lot more bogus results. It also isn’t as if it is that difficult to implement some basic, naive, weighing based on the amount of stars/forks/etc.
Of course it might differ per model and how they are trained.
Having said that, I wouldn’t trust the output from an LLM to write secure code either. For me it is a very valuable tool on the end of helping me debug issues on the scale of being a slightly more intelligent rubber ducky. But when you ask most models to create anything more than basic functions/methods you damn well make sure it actually does what it needs it to do.
I suppose there is some role there for seniors to train juniors in how to properly use this new set of tooling. In the end it is very similar to having to deal with people who copy paste answers directly from stack overflow expecting it to magically fix their problem as well.
The fact that you not only need your code/tool to work but also understand why and how it works is also something I am constantly trying to teach to juniors at my place. What I often end up asking them is something along the lines of “Do you want to have learned a trick that might be obsolete in a few years? Or do you want to have mastered a set of skills and understanding which allows you to tackle new challenges when they arrive?”.
If everything you have read is saying that it is fine, then why does it not feel right for you? Looking around I do get the same impression, it is non-combustible so there is not really a concern there. Basically from what I gather as long as you use the proper wire for use in walls/isolation, leave enough space and generally take good practices in account like using conduit where needed you should be good to go.
I am not an electrician though and certainly not aware of your local code and regulations.
Talking about electricians, if you are worried about doing it not right, why not hire one to do it for you?
Most code on GitHub either is unsecure, or it was written without needing to be secure.
That is a bit of a stretch imho. There are myriads of open source projects hosted on github that do need to be secure in the context where they are used. I am curious how you came to that conclusion.
I’m already getting pull requests from juniors trying to sneak in AI generated code without actually reading it.
That is worrysome though. I assume these people have had some background/education in the field before they were hired?
Long term wearing of vr headsets might indeed be not all that good. Though, the article is light on actual information and is mostly speculation. Which for the Apple Vision Pro can only be the case as it hasn’t been out long enough to conduct anything more than a short term experiment. So that leaves very little data in the way of long term data points.
As far as the experiment they did, there was some information provided (although not much). From what was provided this bit did stand out to me.
The team wore Vision Pros and Quests around college campuses for a couple of weeks, trying to do all the things they would have done without them (with a minder nearby in case they tripped or walked into a wall).
I wonder why the Meta Oculus Quests were not included in the title. If it is the meta Quest 3, it is fairly capable as far as pass through goes. But, not nearly as good as I understand the Apple Vision Pro’s passthrough is. I am not saying the Apple Vision Pro is perfect, in fact it isn’t perfect if the reviews I have seen are any indicator. It is still very good, but there is still distortion around edges of vision, etc.
But given the price difference between the two I am wondering if the majority of the particpants actually used Quests as then I’d say that the next bit is basically a given:
They experienced “simulator sickness” — nausea, headaches, dizziness. That was weird, given how experienced they all were with headsets of all kinds.
VR Nausea is a known thing even experienced people will get. Truly walking around with these devices with the distorted views you get is bound to trigger that. Certainly with the distortion in pass through I have seen of Quests 3 videos. I’d assume there are no Quests 2 in play as the passthrough there is just grainy black and white video. :D
Even Apple with all their fancy promo videos mostly shows people using the Vision pro sitting down or in doors walking short distances.
So yeah, certainly with the current state of technology I am not surprised there are all sorts of weird side effects and distorted views of reality.
What I’d be more interested in, but what is not really possible to test yet, is what the effects will be when these devices become even better. To the point where there is barely a perceivable difference in having them on or off. That would be, I feel, the point where some speculated downsides from the article might actually come into play.
They’re for different needs.
Yes… but also extremely no. Superficially you are right, but a lot of the arguments of why many new distros are created is just because of human nature. This covers everything from infighting over inane issues to more pragmatic reasons. A lot of them, probably even a majority, don’t provide enough actual differentiators to be able to honestly claim that it is because of different needs. In the end it all boils down to the fact that people can just create a new distro when they feel like it.
Which is a strength in one way, but not with regard to fragmentation.
I am not quite sure why there are all these bullet points that have very little todo with the actually issue.
Researchers at the University of Wisconsin–Madison found that Chrome browser extensions can still steal passwords, despite compliance with Chrome’s latest security standard, Manifest V3.
I am not sure how Manifest V3 is relevant here? Nothing in Manifest V3 suggests that content_scripts can’t access the DOM.
The core issue lies in the extensions’ full access to the Document Object Model (DOM) of web pages, allowing them to interact with text input fields like passwords.
I’d also say this isn’t directly the issue. Yes, content_scripts needing an extra permissions to be able to access password input fields would help of course.
Analysis of existing extensions showed that 12.5% had the permissions to exploit this vulnerability, identifying 190 extensions that directly access password fields.
Yes… because accessing the DOM and interacting with it is what browser extensions do. If anything, that 12.5% feels low, so I am going to guess it is the combination of accessing the DOM and being able to phone home with that information.
A proof of concept extension successfully passed the Chrome Web Store review process, demonstrating the vulnerability.
This, to me, feels like the core of the issue right now. The behavior as described always has been part of browser extensions and Manifest V3 didn’t change that or made a claim in that direction as far as I know. So that isn’t directly relevant right now. I’d also say that firefox is just as much at risk here. Their review process over the years has changed a lot and isn’t always as thorough as people tend to think it is.
Researchers propose two fixes: a JavaScript library for websites to block unwanted access to password fields, and a browser-level alert system for password field interactions.
“A javascript library” is not going to do much against content_scripts of extensions accessing the DOM.
The alert system seems better indeed, but that might as well become browser extension permission.
To be clear, I am not saying that all is fine and there are no risks. I just think that the bullet point summary doesn’t really focus on the right things.
It still does? That is an entirely different page and still shows the newest videos of channels you are subscribed to. At least, for me it does.
Have you tried placing the <br>
tag directly after the </h1>
closing tag?
I am dissapointed in that I have not been able to get a single mathematic equation produced (like famous ones), but I know they can?
Well, my understanding is that they actually can’t. LLM’s do “language” mostly based on what is called “next word prediction” so they basically look at the word and predict what the next most logical word would be. (Somewhat simplified). So numbers to them are not numbers but words, which is why they are fairly bad at them.
Opera has Aria, which is like the cleanest version of ChatGPT
Pass, not sure what stake the chinese owners have these days but Opera is a bit too… feature rich in everything.
I do like working with just chat.openai.com for simple stuff. It is great at helping my debug things in areas I don’t quite have all the knowledge I’d like. For example, I had to work on a shell script earlier in bash. Something I don’t do often and as an added bonus it needed to work on both macOS machines and the bash version shipped with “git bash” on windows. MacOS GNU utils already function slightly differently at times, but git bash on windows is entirely broken in some areas. Where yesterday I spend an hour trying to find something relevant based on my input and the error I got through google chatGPT just managed to point out the pain point right away.
And that is where I feel chatGPT (in this case anyway) does a great job, troubleshooting issues about things that are not necessarily bleeding edge. I just presented it with a clear problem and a bit of context and asked why that could be the case. It also got it wrong a few times, but that is fine, it did safe me a bunch of time in the end.
Bing and Google Bard keep disappointing me. Bing for some reason only picks up on half of what I ask. Which is extremely odd as it is supposedly is ChatGPT based and ChatGPT gives pretty good answers on the same queries. The only problem with the latter is that a lot of it is of course outdated.
Bard might just be broken for me. I keep getting I'm a text-based AI, and that is outside of my capabilities.
or similar responses.
Yes, of course :) This is the desktop web front-end, so the “default” .
What do you mean by “it”? The chatGPT interface? Could be, but then you are also missing the point I am making.
After all, chatGPT is just one of the possible implementations of LLMs and indeed not perfect in how they implemented some things like search. In fact, I do think that they shot themselves in the foot by implementing search through bing and implementing it poorly. It basically is nothing more than a proof of concept tech demo.
That doesn’t mean that LLM’s are useless for tasks like searching, it just means that you need to properly implement the functionality to make it possible. It certainly is possible to implement search functionality around LLMs that is both capable and can be reviewed by a human user to make sure it is not fucking up.
Let me demonstrate. I am doing some steps that you would normally automate with conventional code:
I started about by asking chatGPT a simple question.
It then responded with.
The following step I did manually, but is something you would normally have automated. I put the suggested query in google, I quickly grabbed the first 5 links and then put the following in chatGPT.
It then proceeded to give me the following answer
Going over the search results myself seems to confirm this list. Most importantly, except for the initial input, all of this can be automated. And of course, a lot of it can be done better, as I didn’t want to spend too much time.