6 ms·
I like the parable that any text content that originated before 2021 now has increasing value, and in certain sense became a digital equivalent of physical "Low
by myfonj 3y ago
I like the parable that any text content that originated before 2021 now has increasing value, and in certain sense became a digital equivalent of physical "Low-background steel" [1].
Its value is mostly, and paradoxically, for training LLMs, to get "unspoiled" models.
[1] https://en.wikipedia.org/wiki/Low-background_steel https://en.wikipedia.org/wiki/Low-background_steel
- myfonj 3y agoAnecdotally, I've recently visited LinkedIn and saw promoted job offerings for "Photo Collector" [1] that among other interesting requirements had one point that stood out: • Only photos taken before October 1, 2023, will be accepted (I chuckled.) [1] https://web.archive.org/web/20240322134531/https%3A%2F%2Fwww.linkedin.com%2Fjobs%2Fview%2F3831240660%2F https://web.archive.org/web/20240322134531/https%3A%2F%2Fwww...
- silverquiet 3y agoI don't mean to be a vexatious asshole (I suppose my problem is that I can't stop myself), but if content that's created before 2021 is more valuable, what's to stop AIs from creating more of it via the power of lying? (Or call it a hallucination or whatever.)
- snapcaster 3y agoThe whole source of the value increase is that it's verifiably non-AI generated content
- silverquiet 3y agoOk, how do you verify that?
- CoastalCoder 3y agoI agree with your skepticism. I think you're being realistic about the underlying adversarial dynamics. I.e., it's an arms race regarding recognizing AI authorship.
- snapcaster 3y agoBecause the capabilities didn't exist before then? Which part do you think is difficult to verify?
- staindk 3y agoI can get an AI to write 50k bad articles and publish them on my site, with the publish dates set back to before 2021.
- vik0 3y agoThe first sentence of the article reads: >*Google Books is indexing low quality, AI-generated books that will turn up in search results* This was never about articles, it is about books, though. You can't retroactively publish a book in 2021 if we're not in 2021. It's bewildering to see how people here are so confident in their tone while arguing about things they think are said in the linked post, but don't know for sure because they never even opened it to read it.
- staindk 3y agoTopics of discussion change, in this thread I believe we're talking about AIs "hallucinating" "content". Not books specifically. You make a valid point though, which is actually more scary now that I think about it... it is becoming less and less possible for non-google (/other huge) companies to train models on verifiably human-generated content, as only Google (/other huge companies) have enough accurately timestamped data to train on. Hmm.
- silverquiet 3y agoWhat capabilities didn't exist? The AIs I've interacted with create artifacts that are either text files or digital images, either of which can be printed out into physical manifestations. It's not like this is exactly a new phenomenon - Plato's famous "Allegory of the Cave" was a thought experiment in how the media we consume can create what we perceive as reality - it's just that we're discussing an uncomfortably recent intersection of philosophy with computer science. I said it a few times in the threads that blew up over Google's diverse Nazi bungle; I didn't get into computers to discuss anthropology and how humanity constructs meaning; I actually did it to avoid that and do work I considered purely practical after washing out of a liberal arts degree. That said, the whole thing ironically felt like very good art in the sense that it got people to discuss the nature of these generative AI systems. For me especially, the image of the Black Pope was very thought provoking as a (lapsed) Catholic; why is it that there has never been a black pope?
- ADeerAppeared 3y agoIt's not the AI lying about the date, it's about the people using the content. One big problem is that Google also lies about dates when you do a 'before' search, so AI spam still shows up even if you filter it out. Scraping a new dataset without contaminated content is impossible now. There's no authoritative source on the age of any given website. (You can try places like the Internet Archive, but most such archivers aren't interested in handing their data over to AI firms because of the bad PR.) The only way to ensure you don't have AI content is to use a dataset created before ~2021-2022. This too is hard because all the big ones are illegal to posses as they contain CSAM. The real answer is that AI developers played themselves. Humanity was creating exponentially more data each year, and by rushing these AI tools out the door, all that data is useless now. By the end of this year there'll be more new "AI-polluted" data on the web made after 2021 than the total amount of "clean" data made up to 2021.
- Solvency 3y agoI don't buy this take for a second. You honestly don't think Google has a VAST and complete trove of time-stamped data scraped and indexed from decades of its existence. You don't have access to it. But they do.
- lupire 3y agoPeople care a lot more about current events than past events.
- withinboredom 3y agoIn my experience, indicies tend to trust the date in the article more than the date it was discovered. That’s also how some people fight bad press: copy an article verbatim, set the date to the day before, file takedown notice for theft of content. By the time the actual content comes back up, it won’t matter.
- ancientworldnow 3y agoAI training doesn't care anymore. Huge amounts of it now are intentionally created synthetic data made by LLM's to generate data to train bigger LLM's. The larger the model, the less it matters.
- burkaman 3y agoA dishonest steel salesperson can also lie to you about when their steel is from. If you don't have a way of measuring the radiation, then you have to rely on trusted sources that you don't believe will lie to you, and not accept any steel from someone you don't know. It's the same here, we have many trusted sources that we don't believe would lie about when content is from - Internet Archive, newspaper archives, widely used web crawls, Wikipedia exports, etc. Obviously you don't use an AI (or a black box search engine) to get non-AI content, you get it from people and systems you trust.
- silverquiet 3y agoYes, I see no other way than trusted humans; my problem is that I've worked with salespeople and there are vanishingly few that I trust.
- visarga 3y ago> I like the parable that any text content that originated before 2021 now has increasing value, and in certain sense became a digital equivalent of physical "Low-background steel" It goes even deeper. If OpenAI has 100M users, and they consume let's say 10K tokens per month on average, that means 1 trillion tokens are read by the user base, who then go out and act in the world, at the very least they will read the text. The impact of AI onto the world is huge because we flock to it by the millions. Such a massive new source of text will surely influence how language is spoken and potentially support many activities, including scientific research. One of the consequences will be much wider circulation of information, as users only need to point it to the right direction and it will readily combine information across multiple fields. Recently updated AIs might be even more informed than domain specialists because nobody got time to read everything new, like LLMs, we still need most of our time for work.
- myfonj 3y ago:] By their tongue, you will know them. Tangentially related: I really loved the 'shibboleth' that went like complex and multifaceted [things] that encompass [other things] from November 2023 [1]. Even though it was clear this will soon disappear from newer models being too obvious, I was still quite surprised, that "multifaceted" didn't even get into more recent study from March 2024 [2] that pointed out skyrocketing of terms such as commendable, intricate, and meticulous. [1] https://news.ycombinator.com/item?id=38501589 https://news.ycombinator.com/item?id=38501589 / https://blog.j11y.io/2023-11-22_multifaceted/ https://blog.j11y.io/2023-11-22_multifaceted/ / https://twitter.com/padolsey/status/1727555475573440996 https://twitter.com/padolsey/status/1727555475573440996 [2] https://news.ycombinator.com/item?id=39909692 https://news.ycombinator.com/item?id=39909692 / https://arxiv.org/abs/2403.07183 https://arxiv.org/abs/2403.07183 / https://twitter.com/myfonj/status/1773371415149891871 https://twitter.com/myfonj/status/1773371415149891871
- Solvency 3y ago> Such a massive new source of text will surely influence how language is spoken Before chatGPT I'd see the words "it's worth noting.." maybe five times in my entire life (39 years old). I now see that stupid phrase dozens of times per day, every single day, everywhere I go. Including HN, which tosses it around even more than Reddit.
- api 3y ago> Its value is mostly, and paradoxically, for training LLMs, to get "unspoiled" models. This is the tell that present-generation LLM "AI" is not actually intelligent. It doesn't truly understand anything. It's just a general purpose lossy compression algorithm for text that is queryable through replaying the model to complete prompts. I don't necessarily mean "just" dismissively. That's obviously incredibly useful. My laptop now has local LLMs on its hard drive that contain a conversationally queryable digest of the sum total of a large fraction of human knowledge, and we've basically solved natural language interfacing to computers and natural language translation. But it's not "AI" in the sci-fi sense. If it was, it would be generating truly novel insightful content that enlarged the base data set not padding the data set with regurgitations of what is already there. I actually go back and forth about how much this represents a step toward true thinking AI. It's hard to judge because the novelty of these things being able to work so well with language causes a "wow" reaction that might cause me to overestimate just how large a step we have made. Are these really just nothing more than linguistic JPEGs or is there a lot more going on here?
- ToucanLoucan 3y agoI've been saying exactly this from the beginning. I don't like being lied to, and calling ChatGPT AI was always a lie, from minute one. It is not intelligent in any sense of the word, whatsoever, nor may it aspire to intelligence. It's a highly adept predictive text model. It generates words based upon words it has been trained on as being the right words to minimize the errors in the words it generates. Not useless, mind you, and frankly tons of B2B salespeople especially would benefit from using it to clean up their nearly incoherent rambling emails; but not intelligence. The way you can "talk to" these machines is certainly clever, and the way they generate their replies is fascinating technology and certainly beats the shit out of a traditional chatbot, all of these things are true and I think they have a bright future as such. Natural language UI is a cool concept and it's cool to see some serious progress there. But that's all it is, and all it really can be. I don't know for certain what the path is to true emergent intelligence from the machine, but I assure you that shoving billions of words scraped from the internet through an LLM is not that. Now, if for whatever reason you want ream upon ream of pretty dull and repetitive content churned out on a given subject, that's far longer than it needs to be to convey the limited information within as apparently tons of online businesses do, written damn near to perfection? ChatGPT's your hookup, no question. However you will never get blinding insights, you will never get new ideas, and it will never be your friend. Sorry.
- JohnFen 3y agoThe value isn't just in terms of training LLMs. I know that I regard anything I see on the web that was created after the advent of LLMs to be suspect and requiring additional vetting before giving it much weight.
- harshreality 3y agoA sea of low-quality content isn't a problem if the training process incorporates a way to estimate quality and eventually train the NN out of low-quality output. There are many areas where that estimation might be difficult without human intervention: arts of all kinds. However, in a more fact-based area like programming, a NN training routine could be modified so that what a LLM thinks it's learning can be tested. That could apply to CS and math, at least. Eventually, NNs integrated with robotics could learn physical sciences without human intervention or pre-selected training material, and would still converge to superhuman knowledge even starting from random ideas. Giving a NN access to robotics without a human in the loop is probably too dangerous, though... it might decide to unsafely test its understanding of nuclear reactions, or the r-number of ebola...
- dakial1 3y agoI think more than "text content generated before 2021" we'll have a curated list of sources, either online or physical publishers / content generators that will have a quality classification based on how good their content governance is.
- intelVISA 3y agoI guess a fallback career for SWEs in the future is generating some fresh low-background code for the right price. Maybe we'll get stickers like non-GMO plants for organic code.