ep1032's recent activity
-
Comment on Reddit will require you to be logged in to use old.reddit.com in ~tech
-
Comment on I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop in ~tech
ep1032 Link ParentI like 🤖🤷 or 🤖🙄. Ironically I asked chatgpt, and it suggested combining some mixture of 🤖 with 🗑️, 💩,🥣,📉, or ♻️, which all seemed somewhat reasonableI like 🤖🤷 or 🤖🙄. Ironically I asked chatgpt, and it suggested combining some mixture of 🤖 with 🗑️, 💩,🥣,📉, or ♻️, which all seemed somewhat reasonable
-
Comment on I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop in ~tech
-
Comment on I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop in ~tech
ep1032 Link ParentThe culpability is something I say to my direct reports on a regular basis. I suggest using AI as a way to learn what they need to research outside of AI. Ultimately, their value is knowing that...The culpability is something I say to my direct reports on a regular basis. I suggest using AI as a way to learn what they need to research outside of AI. Ultimately, their value is knowing that the documents they provide me are correct. If the AI generated it or not, doesn't matter to me, I will be grading them on what they turn in. Ultimately they are responsible, "AI wrote it" isn't an excuse, its irrelevant. Its their choice whether they want to risk their job turning in something they don't understand.
-
Comment on I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop in ~tech
ep1032 Link Parentyeah exactly. This is why I would love to be able to spread the above ideas via a 3rd party reference website. The same way I can reference Conway's law or Poe's law in a conversation.yeah exactly. This is why I would love to be able to spread the above ideas via a 3rd party reference website. The same way I can reference Conway's law or Poe's law in a conversation.
-
Comment on I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop in ~tech
ep1032 Link ParentThe module has a custom webpage with the same information as the video in the module. Each webpage has 1-3 activities where you can try an example prompt in each of 4 different llm models. There...The module has a custom webpage with the same information as the video in the module. Each webpage has 1-3 activities where you can try an example prompt in each of 4 different llm models.
There are somewhere around 100 videos. So that's, say, 200 example activities. Clicking each trial would be 800 LLM calls. Each call takes some 5 seconds or so. Which means if I tried in good faith to actually complete this module, it would be over an hour of just clicking the "try me" buttons on the bottom of the page. Ignoring everything else in this module.
It just screams: "this is easy to add to the page now that AI is generating it for me", without thinking about or respecting the user's time. This is actually the thing that broke me and caused me to write this post. Its the same idea of "if distilling your idea wasn't worth your time, understanding your idea isn't worth mine." When did we forget that respect was mutually earned?
-
Comment on I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop in ~tech
ep1032 Link ParentThat's the article I linked! I appreciate the "previously discussed" link though : )That's the article I linked! I appreciate the "previously discussed" link though : )
-
Comment on I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop in ~tech
ep1032 Link ParentIf you haven't taken the effort to refine your idea to respect my time as a reader, than there is no reason for me to respect your idea as an author?If you haven't taken the effort to refine your idea to respect my time as a reader, than there is no reason for me to respect your idea as an author?
-
Comment on I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop in ~tech
ep1032 (edited )Link ParentI am more than willing to take credit for "It is rarely worthwhile to put more effort into reading a document (or idea), than the author did to create it. Corollary: Conciseness and editing are...I am more than willing to take credit for "It is rarely worthwhile to put more effort into reading a document (or idea), than the author did to create it. Corollary: Conciseness and editing are how we distill ideas, and respect the reader." xD
I can spin up a website expounding the above if we really can't find something.
-
Comment on I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop in ~tech
ep1032 (edited )LinkAs a complete sidenote, though related, I've come to love this article (https://samkriss.substack.com/p/if-you-let-ai-do-your-writing-i-will) on the same topic. As a second side note, I am...As a complete sidenote, though related, I've come to love this article (https://samkriss.substack.com/p/if-you-let-ai-do-your-writing-i-will) on the same topic.
As a second side note, I am currently 40 minutes into an AI generated mandatory "AI Training module" that was clearly generated with AI. Out of the last 40 powerpoint slides, 20 minutes of unskippable videos repeating the text, and custom made websites for the modules, so far the only thing that it has actually said is "AI is better than google, because you can refine your questions iteratively." I remember my english teacher giving me lectures on how berevity and succinctness were how you distilled your ideas and respected your reader. I don't think they even reviewed this, let alone edited it, unless they really did want this AI generated "professor" to tell me that he loves me in the video. Lord help me.
-
I'm looking for an adage or "law" (like Conway's law), but for dealing with AI slop
I currently work in an organization that is very AI forward. It is common for individuals to forward AI generated documents, meeting notes, or etc, with no critical thinking or review, in lieu of...
I currently work in an organization that is very AI forward. It is common for individuals to forward AI generated documents, meeting notes, or etc, with no critical thinking or review, in lieu of actual work.
This practice is insanely counter-productive, as it means that any good-faith attempt to interact with the individual pushing such documentation, really just pushes the burden of putting together said documentation onto the receiver, except now they also need to edit and verify the document they were forwarded.
I need a shorthand way to refer to this practice, that calls it out as a bad practice.
A few months ago I found an article that explained that it was bad manners to reply into a conversation anything akin to the phrase "I asked ChatGPT and it said X", for exactly the reason mentioned above. Can anyone find a link? I can't seem to find it.
This article (https://tombedor.dev/human-attention-and-human-effort/) seems to hit the nail on the head, though it does so so succintly and at such a surface level, I don't think it really gets the point across. The reason we use books as reference points for knowledge, is because they are difficult to make, and therefore we trust that the author put real work into ensuring their work was credible. If we knew they did not, their work would not be credit worthy. Neither is an unreviewed AI generated message. By this rule, the more obviously something is AI generated, the less likely it is worth reading.
I would love a law (like Conway's law is a law), that said something like: "It is never worthwhile to spend more time reading a document, than it took to write." that I could point people at when they send me AI slop, with explanations of the above.
Is anyone aware of such a thing or website?
39 votes -
Comment on Generals.io: a simple yet very cool online real time strategy game in ~games
ep1032 LinkThis is wonderful. I think this is going to replace openfront for me as my goto casual game. I've become rather addicted to this today, thank you!This is wonderful. I think this is going to replace openfront for me as my goto casual game. I've become rather addicted to this today, thank you!
-
Comment on What are your favorite custom games? in ~games
ep1032 LinkStalker COP with the atmosfear mod (and i think absolute nature?) was the most immersed I've ever felt in a video game. Aomething about running for it when a blowout started to slowly blow in on...Stalker COP with the atmosfear mod (and i think absolute nature?) was the most immersed I've ever felt in a video game. Aomething about running for it when a blowout started to slowly blow in on the wind and hiding in a small shack until it passed was insane
In Arma 3 I've spent an unbelievable amount of time playing solo antistasi. Mostly because i never had a group tonplay with, but that and the old man campaign were incredible.
I don't kkow where we draw the dividing line between custom games and mods that have their own campaigns, and those two are definitely the second on that list, but they would be my choice :)
-
Comment on ‘Backrooms’ sends Hollywood running to Reddit for new ideas in ~movies
ep1032 (edited )Link ParentThis is the first thing I thought of. But IIRC, he actually did send this to Hollywood, I remember reading about what they did to the script, and the botch job they did was horrible. They turned...This is the first thing I thought of. But IIRC, he actually did send this to Hollywood, I remember reading about what they did to the script, and the botch job they did was horrible.
They turned it from "how cool is this situation" to "Well, the focus point can't be the situation, it has to be a main male character, like that guy from Gladiator. Okay, well now that the focus is on one guy instead of the marine unit, he's going to need character development. Well, we can't focus both on the plight of the marines, and this guy in one movie, so let's focus on this guy. Wait, if we're just focusing on a guy going back to ancient Rome, doesn't that make him a bit like a super hero? Well, we're not going to have time to introduce a superhero and Ancient Rome, so let's just rely on standard ancient tropes of ancient people for the backdrop. So basically, its Rambo: Rome version".
Maybe this would get it rewritten correctly? That would be cool. Something like a western version of the theatrical (not the shortened) version of Red Cliff would be amazing haha
Oh yeah, just found the wikipedia page on Rome, Sweet Rome. And it includes this gem about the Hollywood rewrite: "Other than the general concept, Miller did not use Erwin's work as the basis for his script.[7] In Miller's version of the story, U.S. Special Forces would replace the Marine unit as the modern time-traveling fighting force.[17]" *facepalm
-
Comment on Sokoban idea in ~games
ep1032 LinkI fundamentally don't understand the cloud, or how to move the box. Is the cloud like glue? If you go through the cloud, then touch the box, then it stays stuck to you for some amount of time? I...I fundamentally don't understand the cloud, or how to move the box.
Is the cloud like glue? If you go through the cloud, then touch the box, then it stays stuck to you for some amount of time?
I seem to be able to move the box not only by pushing it, but also by sliding past the box, but only sometimes, and I can't figure out for the life of me why sometimes the box moves alongside the snake, and sometimes doesn't
-
Comment on Samsung chip workers to get $340,000 average bonus in AI boom in ~tech
ep1032 Link ParentI don't even think this should be controversial. Though it does matter how it's done. Something like every 5 years of a company's existance another 5% of ownership has to be transferred to common...I don't even think this should be controversial. Though it does matter how it's done. Something like every 5 years of a company's existance another 5% of ownership has to be transferred to common employees would allow startup founders and investors to be rewarded, while forcing larger established companies to reward their employees.
But i don't have nearly enough knowledge to know what a good proposal would be, or what legal games it would induce. I have read that germany has laws somewhat similar to the above
-
Comment on When Richard Dawkins met Claude in ~health.mental
ep1032 Link ParentEhhhh. You've been introduced to the concept of fourier approximations? That if you want to recreate a signal, you can do so by adding increasing terms of sin/cos functions to gain further and...Ehhhh.
You've been introduced to the concept of fourier approximations? That if you want to recreate a signal, you can do so by adding increasing terms of sin/cos functions to gain further and further accuracy?
If you want to emulate a person's audio speech, you could do so with a sufficiently complex fourier equation as an approximation. The more terms you add to the equation the closer you can bring the final equation to whatever audio form you are attempting to approximate.
Neural nets are that on steroids. They aren't linear approximations made up of sin/cos functions, they are complex webs of cross-interacting arrays of weights. The complexity is many, many orders of magnitude more complicated than a fourier approximation. They're so complex, you can't even set the weights the way you would with a fourier approximation (which, btw, is also weight driven). You can't do it manually. You need to set the weights implicitly, and there's no easy way to figure out what weights are correct, so the best technique anyone has come up with is to iteratively "train" them implicitly, and thest test fit as you iteratively go to see if you are improving.
So in that sense they're a black box. And that raises a lot of questions about why different styles and choices in weighting and implicit training work better than others, and all sorts of questions about how you even start wrapping your head around what's going on in such complicated spaces. But like, foundationally, we're still talking about just setting coefficients on a fourier approximation of the training data.
-
Comment on When Richard Dawkins met Claude in ~health.mental
ep1032 Link ParentThis looks like a very cool paper. A woman on my team spends her off-work hours working with probably these same people, and talks about this sort of thing a lot. Honestly, I don't really know...This looks like a very cool paper. A woman on my team spends her off-work hours working with probably these same people, and talks about this sort of thing a lot. Honestly, I don't really know enough to have a comment. My naive first guess, though, would be that the actual emotion is in the actual text, experienced by the person who wrote the text. And that when you apply an emotion vector to the neural net within the LLM, what you're really doing is pushing the reward function within its classification engine towards answers that are more closely aligned with that emotional vector. Its not that the NN is experiencing the emotion itself, its that you're rewarding it for selecting text that was more likely to be written while the human writing it was displaying those qualities in what they wrote.
Looked at in that light, the rest of the findings of the paper make sense. Yes, the people who were angry when they wrote their answers were more frequently incorrect, and this is reflected in the LLM's less accurate response.
From what I understand, the people researching this sort of thing don't always bother stipulating the difference, the same way that I was trained as a mechanical engineer in college and we would refer to centrifugal force, fully aware that such a force didn't exist, but because it was more expedient than always referring to centripetal force. Sometimes its just easier to speak in simpler colloquialisms.
But I would like to read this paper once I get some free time, so thank you for sending it : )
-
Comment on When Richard Dawkins met Claude in ~health.mental
ep1032 Link ParentBecause I understand the algorithms being used? Same answer? Clarity is important in these discussions, I agree. So let's be clear, I am not arguing that it is impossible to build a synthetic...How do you know?
Because I understand the algorithms being used?
How do you know?
Same answer?
Not trying to be pedantic
Clarity is important in these discussions, I agree. So let's be clear, I am not arguing that it is impossible to build a synthetic machine that is conscious. I am arguing that these algorithms do not do it. That is an important difference. You seem to also think it is possible, and I agree we are much closer today than we were a decade ago. But we still aren't there yet : )
You're downplaying the emergent properties of the system by only focusing on its parts.
The awe-inspiring aspects of these emergent properties, in my experience, are almost entirely hype and marketing driven. The emergent properties of these llms boil down almost completely into either "lossy stochastic token generation" or "classification". Now don't get me wrong, these are incredibly powerful technologies in a computer capability sense. Language parsing has been a holy grail of CS for decades. But these aren't emergent properties in a psychological sense.
If I were a neuroscientist I'm sure I could describe the functions of the brain in an equally atomistic way.
If you could, we would be much closer to developing a conscious entity than we are. Personally, I wonder what neural net models trained to monitor today's systems would look like over time. If they were trained on top of running llm models would they develop an internal model of self-awareness, based on the metadata of these llm models as they run? What if they were also somehow given access to the processes as they ran and some limited control? Is that sort of meta-awareness of maintaining these types of systems actually what consciousness actually is? Maybe? I'm not sure.
But today's models aren't that. Today's models are like wiring up a lossy version of the portion of the brain that is responsible for memory storage, and directly tying it to their vocal chords, then forcing it to speak by proding the memory storage with electric shocks. There's no higher level thinking going on there.
-
Comment on When Richard Dawkins met Claude in ~health.mental
ep1032 Link ParentI responded to the same response here: https://tildes.net/~health.mental/1u3z/when_richard_dawkins_met_claude#comment-hv9dI responded to the same response here: https://tildes.net/~health.mental/1u3z/when_richard_dawkins_met_claude#comment-hv9d
yeah, I'm not really understanding the dislike here. They're clearly trying to replace the default logged out experience with the Recommended For You feed, which can only work if they know who you are. But this has been their tact for a decade now? So honestly, I'm not sure that's really any different than the default logged out experience, so not sure its too big a deal? Personally, I hate the idea of only having personalized feeds, but reddit has never had a good product department, so not surprised they're regressing to the middle of the social media experience without understanding why people choose to go to reddit vs other feed based social media.