Archivist David Rosenthal observes now that more material is posted online by LLMs than actual people, the bots are starting to ingest their own digital excrement, creating a negative feedback loop.
In the belief that "more is better", Large Language Models (LLMs) have insatiable appetites for training data. They started by scraping everything on the Web (robots.txt be dammed). When that ran out they downloaded the various pirate libraries (copyright be dammed). That exhausted the texts easily available in digital form, but their hunger wasn't assuaged. As for images, they partly used CAPTCHAs but mostly paid vast numbers of poor people to label the images with what they showed.
When the supply of text ran low, people observed that the LLMs were capable of generating human-like text in large quantities. The obvious idea was to pour the output of the LLMs into their training sets. This wasn't just a conscious decision, it was inevitable. The advent of LLMs rapidly polluted the Web with LLM output. Greg Druck's AI Now Writes as Many Online Articles as Humans notes that:
We observe significant growth in primarily AI-generated articles, coinciding with the launch of ChatGPT in November 2022. After only 12 months, primarily AI-generated articles accounted for 35.9% of articles published.
In Q1 2025, the quantity of primarily AI-generated articles being published on the web nearly equaled the quantity of human-written articles, 49.6% vs. 50.4%. In Q4 2025, primarily AI-generated articles surpassed human-written at 50.9%, before returning to 49.9% in Q1 2026.
Even if slop were not of undesirable quality, it is not produced by humans and thus is completely unsuitable as training data.
Previously:
(2026) A Wikipedia Clone Built on AI Hallucinations is Here to Hasten Along the Death of the Internet
(2025) When It All Comes Crashing Down: The Aftermath of the AI Boom
(2025) AI Favors Texts Written by Other AIs, Even When They're Worse Than Human Ones
(2025) What the Hell is Going on Right Now?
Related Stories
What the hell is going on right now?:
Engineers are burning out. Orgs expect their senior engineering staff to be able to review and contribute to "vibe-coded" features that don't work. My personal observation is that the best engineers are highly enthusiastic about helping newer team members contribute and learn.
Instead of their comments being taken to heart, reflected on, and used as learning opportunities, hapless young coders are instead using feedback as simply the next prompt in their "AI" masterpiece. I personally have witnessed and heard first-hand accounts where it was incredibly obvious a junior engineer was (ab)using LLM tools.
In a recent company town-hall, I watched as a team of junior engineers demoed their latest work. I couldn't tell you what exactly it did, or even what it was supposed to do - it didn't seem like they themselves understood. However, at a large enough organization, it's not about what you do, its about what people think you do. Championing their "success", a senior manager goaded them into bragging about their use of "AI" tools to which they responded "This is four thousand lines of code written by Claude". Applause all around.
I was asked to add a small improvement to an existing feature. After reviewing the code, I noticed a junior engineer was the most recent to work on that feature. As I always do, I reached out to let them know what I'd be doing and to see if they had any insight that would be useful to me. Armed with the Github commit URL, I asked for context around their recent change. I can't know for sure, but I'd be willing to put money down that my exact question and the commit were fed directly into an LLM which was then copy and pasted back to me. I'm not sure why, but I felt violated. It felt wrong.
AI favors texts written by other AIs, even when they're worse than human ones:
As many of you already know, I'm a university professor. Specifically, I teach artificial intelligence at UPC.
Each semester, students must complete several projects in which they develop different AI systems to solve specific problems. Along with the code, they must submit a report explaining what they did, the decisions they made, and a critical analysis of their results.
Obviously, most of my students use ChatGPT to write their reports.
So this semester, for the first time, I decided to use a language model myself to grade their reports.
The results were catastrophic, in two ways:
- The LLM wasn't able to follow my grading criteria. It applied whatever criteria it felt like, ignoring my prompts. So it wasn't very helpful.
- The LLM loved the reports clearly written with ChatGPT, rating them higher than the higher-quality reports written by students.
In this post, I'll share my thoughts on both points. The first one is quite practical; if you're a teacher, you'll find it useful. I'll include some strategies and tricks to encourage good use of LLMs, detect misuse, and grade more accurately.
The second one... is harder to categorize and would probably require a deeper study, but I think my preliminary observations are fascinating on their own.
[...] If you're a teacher and you're thinking of using LLMs to grade assignments or exams, it's worth understanding their limitations.
We should think of a language model as a "very smart intern": fresh out of college, with plenty of knowledge, but not yet sure how to apply it in the real world to solve problems. So we must be extremely detailed in our prompts and patient in correcting its mistakes—just as we would be if we asked a real person to help us grade.
In my tests, I included the full project description, a detailed grading rubric, and several elements of my personal judgment to help it understand what I look for in an evaluation.
[...] The usual hallucinations began—the kind I thought were mostly solved in newer model versions. But apparently not: it was completely making up citations from the reports.
[...] Soon after, it started inventing its own grading criteria. I couldn't get it to follow my rubric at all. I gave up and decided to treat its feedback simply as an extra pair of eyes, to make sure I wasn't missing anything.
[...] Instead of asking the LLM to identify AI-written texts, which it doesn't do very well, I decided to compare my own quality ratings of each project with the LLM's ratings. Basically, I wanted to see how aligned our criteria were.
And I found a fascinating pattern: the AI gives artificially high scores to reports written with AI.
The models perceive LLM-written reports as more professional and of higher quality. They prioritize form over substance.
The Bulletin of the Atomic Scientists published a report on the possible crash of the AI bubble:
Silicon Valley and its backers have placed a trillion-dollar bet on the idea that generative AI can transform the global economy and possibly pave the way for artificial general intelligence, systems that can exceed human capabilities. But multiple warning signs indicate that the marketing hype surrounding these investments has vastly overrated what current AI technology can achieve, creating an AI bubble with growing societal costs that everyone will pay for regardless of when and how the bubble bursts.
The history of AI development has been punctuated by boom-and-bust cycles (with the busts called AI winters) in the 1970s and 1980s. But there has never been an AI bubble like the one that began inflating around corporate and investor expectations since OpenAI released ChatGPT in November 2022. Tech companies are now spending between $72 billion and $125 billion per year each on purchasing vast arrays of AI computing chips and constructing massive data centers that can consume as much electricity as entire cities—and private investors continue to pour more money into the tech industry's AI pursuits, sometimes at the expense of other sectors of the economy.
That huge AI bet is increasingly looking like a bubble; it has buoyed both the stock market and a US economy otherwise struggling with rising unemployment, inflation, and the longest government shutdown in history. In September, Deutsche Bank warned that the United States could already be in an economic recession without the tech industry's AI spending spree and cautioned that such spending cannot continue indefinitely.
Warning signs. Silicon Valley's focus on developing ever-larger AI models has spurred a buildout of bigger data centers crammed with computing power. The staggering growth in AI compute demand would require tech companies to build $500 billion worth of data centers packed with chips each year—and companies would need to rake in $2 trillion in combined annual revenue to fund that buildout, according to a Bain & Company report. The report also estimates that the tech industry is likely to fall $800 billion short of the required revenue.
That shortfall is less surprising than it might seem. US Census Bureau data show that AI adoption by companies with more than 250 employees may have already peaked and began declining or flattening out this year. Most businesses still don't see a significant return on their investment when trying to use the latest generative AI tools: Software company Atlassian found that 96 percent of companies didn't achieve significant productivity gains, and MIT researchers showed that 95 percent of companies get zero return from their pilot programs with generative AI. [...] Claims that AI can replace human workers on a large scale also appear overblown, or at least premature. When evaluating AI's impact on employment, the Yale Budget Lab found that the "broader labor market has not experienced a discernible disruption since ChatGPT's release 33 months ago," according to the group's analysis published in October 2025.
A Wikipedia Clone Built on AI Hallucinations Is Here to Hasten Along the Death of the Internet:
There's a theory that a rising tide of LLM-generated nonsense will eventually drown both LLMs themselves and the internet as a whole. The idea goes like this: The first generation of LLMs is trained entirely on "real" material: the Gutenberg project, 4chan, that one article from Thought Catalog a decade ago, and everything in between. But as the output of those LLMs spreads across the internet, it also becomes part of the training data of future LLMs—and much of it is bullshit .
As a result, the quality of newer LLMs' training data is inferior to that of their predecessors—and by extension, so is their output. And as that output accumulates on the internet, it becomes part of future training data, and the cycle continues. With each passing day, the proportion of the internet that's low-quality LLM-generated bullshit increases, until eventually all that's left to train LLMs is the gibberish created by their predecessors.
The end result is a sort of RAM-hoovering, water-guzzling, bullshit-munching ouroboros, an unholy circular undulant with Jensen Huang's face at one end and Sam Altman's at the other, slowly human-centipeding both itself and the internet into oblivion. If humanity hasn't set fire to the planet by that point, then we start a new internet, hopefully with lessons learned along the way.
And even if the doomsday scenario of the internet drowning in a sea of em dashes and it's-not-just-x-it's-y constructions never comes to pass, people are starting to take the idea of using LLMs to poison LLM training data and run with it.
Take, for example, Halupedia , an absurdist Wikipedia-esque site whose pages are entirely populated by content that an LLM has made up—sorry, hallucinated— on demand. If you search for a topic that someone has previously entered, you'll get the existing nonsense. If your search is the first of its kind, the LLM will carefully assemble your very own small mound of nonsense from a list of possible topics.
According to the site's tips-for-tokens page , Halupedia appears to be the work of one Bartłomiej Strama. The page also provides a little more insight into the purpose of the project, which isn't 100% clear at face value—Strama tells one contributor, "Your contribution towards polluting LLM training data will surely benefit society!"
Of course, quibblers might argue that there's more than enough LLM-generated rubbish on the internet already without sites deliberately adding to the pile. Google pretty much anything these days and you'll find umpteen long-winded articles that purport to explain the topic in question, but really just waffle for paragraph after paragraph without saying anything at all. This is certainly true, but there's some virtue in the fact that Halupedia's output is openly and exuberantly absurd as opposed to content that is superficially credible and doesn't reveal its true nature without closer inspection.
Although... you may also find yourself wondering which topics other users have been entering into Halupedia. After all, you can basically enter any subject into the site's "search" bar and have it write an article for you. The answer lies in the site's list of trending topics, and... sigh.
Yep, it's the usual mix of shitposts, nonsense, and unabashed racism—or, in other words, it's basically the internet's id in microcosm. In fairness, some of these pages have been deleted—click on "niggabutt" and you get this:
But since the page title still shows up in the sidebar, it's not like it's been entirely banished. On the tip page, Strama also comments on the challenges of moderation: "The moderation sometimes is too restrict, but at least it's not griefed now." That's as it may be, but it's hard to see this ending well once 4chan gets a hold of it. This is why we can't have nice things, etc.
(Score: 3, Interesting) by JoeMerchant on Monday July 06, @11:51PM (18 children)
How accurate is accurate enough?
If a 1st round trained LLM reading 'virgin' human digital excrement can synthesize "truth" 80% of the time, is that better or worse than the baseline training set?
I submit: if the output is even 1% more accurate / reliable / true than the base training set, then each round of re-training should incrementally improve.
There was a time, perhaps 2 years ago now, when accuracy rates were below 50%, and that is a losing game, but lately in some domains 80% seems very realistic for common / widely published information.
I will say: 2 hours with Gemini and I had a working path through the city building department variance system which multiple 20+ year experienced contractors were unable to find for their friends / potential clients after years of trying. This was very much an LLM assisted search / synthesis process, one which simple keyword matches failed to produce helpful results for.
🌻🌻🌻🌻✌️ [google.com]
(Score: 2, Informative) by khallow on Tuesday July 07, @02:54AM (10 children)
Obvious rebuttal: why is the second time going to be any better than the first time? The clean up work was done the first time. I see this as something like photocopying. The first copy can be better, filtering out bad parts of the original copy. Making recursive copies past that point won't improve noise cleanup, but they will add noise due to introduction of error.
(Score: 2) by JoeMerchant on Tuesday July 07, @03:05AM (8 children)
> why is the second time going to be any better than the first time?
Ever hear of successive approximation? Multi-stage filters? Progressively finer grit polishing processes?
> The first copy can be better, filtering out bad parts of the original copy. Making recursive copies past that point won't improve noise cleanup, but they will add noise due to introduction of error.
In some copy systems, yes. Other systems can benefit from multiple passes.
One thing current LLM training methods aren't: fully developed.
🌻🌻🌻🌻✌️ [google.com]
(Score: 1) by khallow on Tuesday July 07, @03:34AM (7 children)
Indeed, but I also get that the approximation is to a system that's being altered in real time by the approximator.
(Score: 2) by JoeMerchant on Tuesday July 07, @03:53AM (6 children)
> I also get that the approximation is to a system that's being altered in real time by the approximator.
Sounds vaguely reminiscent of Runge-Kutta. RK1 and RK2 are intuitive to the point of being something you'd just come up with in an iterative process looking at the cumulative errors you get and what obvious tricks might make them smaller. RK3 and RK4 take a little more formal methodical approach to develop - I never bothered to go that far before some Uni math class taught me the name of the thing(s).
Conceptual tokenization, statistical determination of truth based on not only frequency but context, diversity of sources, citation reputation of sources, replicatability and God knows what all else they've been cooking up in the labs... there's probably already a filter looking at likelihood that content is AI generated and treating that content appropriately to maximize its value in determination of "best truth."
I, under my cap of tin foil, am virtually certain that Project 2025 and similar have their "Rand Corporation think tank" equivalents who have predicted for their employers (for the past 10+ years) that "now is the lowest cost best opportunity to significantly affect the subjective truths that society will accept into the coming 100 years through the influencing of online content and the LLM systems being trained on that content." And, in response, they have created real-life equivalents of Lex Luthor's monkey army based online troll farms [fandom.com] from the recent James Gunn Superman movie - not to mention direct ownership of major LLM development efforts at X/Grok, Meta/Llama and friends, Bezos' Nova-Olympus, etc.
🌻🌻🌻🌻✌️ [google.com]
(Score: 1) by khallow on Tuesday July 07, @11:06AM (5 children)
Think of it more like the approximator is altering the original differential equation in addition so that after a few iterations you are deviating from the true solution, not due to the error of the approximation, but because you are solving a perturbed equation rather than the original one. With AI, it is also possible that the subtle hallucinations are more influential than the real stuff so that deviation happens faster than expected.
I strongly doubt those think tanks have thought that far ahead. They're paid to turn patron demands into acceptable propaganda. My point instead is that even with sincere development of AI, it will deviate from reality based on the insertion of AI output into the input. And this deviation may be faster than expected due to an AI propensity for accepting AI input (not necessarily from itself directly). While I can't locate the story presently, I recall AI chatbots found arguments from other AI chatbots - even from different AI sources - to be more persuasive than human sources.
(Score: 3, Interesting) by JoeMerchant on Tuesday July 07, @12:10PM (4 children)
>With AI, it is also possible that the subtle hallucinations are more influential than the real stuff so that deviation happens faster than expected.
I see no reason why the AI tag makes susceptibility to influence by hallucinations any more, or less, likely. You might say that meatbags' lived experiences give them a "sixth sense" of truth, but I think that's equivocal with their susceptibility to cargo cult and other false pattern following.
>I strongly doubt those think tanks have thought that far ahead.
I strongly doubt their patrons think that far ahead, but I'm relatively confident that their patrons _do_ pay "smart fellers" to make these kinds of analyses for their consideration of how to direct their wealth/power in order to make their propaganda more effective.
>My point instead is that even with sincere development of AI, it will deviate from reality based on the insertion of AI output into the input. And this deviation may be faster than expected due to an AI propensity for accepting AI input
My point is that a sufficiently powerful "bullshit detector" should reach a tipping point at which additional passes across a dataset can refine and remove remaining residual bullshit. I would say that is impossible at a hallucination rate greater than 50%, unlikely at hallucination rates just below 50% for the kinds of reasons you are stating, but... somewhere... feels like not much less than 20% to me... there should be a tipping point at which self-reflection yields a more consistent worldview, not less.
>I recall AI chatbots found arguments from other AI chatbots - even from different AI sources - to be more persuasive than human sources.
I don't doubt that has been the case for some training algorithms, but it sounds like the kind of phenomenon which can be tuned, even tuned to a negative gain - with hillariously disastrous consequences when done arbitrarily without evaluation/tuning control loops - but somewhere in there it should be able to damp out the AI influence to "reveal underlying truths" better than synthesized hallucinations.
The simple method I have stumbled upon to reveal hallucinations is: verify everything. When a chatbot tells me it has found a person, or a case citation, or whatever: great, now show me evidence of that - multiple writings of the person from independent sources, multiple publications citing the case, etc. It should be relatively easy (if expensive) to add a self-evaluation layer during training - the "I'm from Missouri, you have to show me" filter. AI itself tells me that this kind of RAG verification costs 3-10x as much "context window" as simply accepting text at face value, diminishing the "brain capacity" of your LLM from the equivalent of six bumblebees to 2, or less, but those diminished capacity bumblebees are functioning like they're "from Missouri" and they aren't as easily led astray.
🌻🌻🌻🌻✌️ [google.com]
(Score: 0) by Anonymous Coward on Tuesday July 07, @03:06PM (1 child)
> My point is that a sufficiently powerful "bullshit detector" should reach a tipping point at which additional passes across a dataset can refine and remove remaining residual bullshit.
Maybe, as you say, this can work if the amount of BS is below some value, but I wonder if that is a reasonable assumption--for example if the LLM training only included one original source for the "fact" in question and everything else derived from that?
Let me take you back a couple of decades, here's Jaron Lanier's critique of Wikipedia (which persistently maintained errors in his biography), https://s6cmedia.wordpress.com/2016/03/23/jaron-lanier-criticism-of-wikipedia-web-2-0/ [wordpress.com] Here's the first part of that page (which also includes links to more reading and papers),
Seems to me that everything above also applies to LLM's that feed on their own slop, but possibly much worse than the results of group-editing in Wikipedia.
(Score: 4, Interesting) by JoeMerchant on Tuesday July 07, @03:33PM
>collectively created works may be manipulated behind the scenes
Not just collectively created, whether by anonymous groups of authors, or algorithms tuned to specific purposes contrary and hidden from the user base (https://www.npr.org/2025/07/09/nx-s1-5462609/grok-elon-musk-antisemitic-racist-content), hell - Benjamin Franklin writing as Silence Dogood was behind the scenes manipulation of popular opinion...
It's a perpetual concern, an endless arms race, and I'll say that the single best measure of relative progress is: transparency. How functionally transparent is the entire system? The less transparent, the less it should be trusted.
🌻🌻🌻🌻✌️ [google.com]
(Score: 2) by hendrikboom on Thursday July 09, @10:14PM (1 child)
Reality may imply consistency, but consistency does not imply reality.
Do we merely want the LLM output to be consistent? Or do we want it to be correct?
(Score: 2) by JoeMerchant on Thursday July 09, @10:59PM
>>there should be a tipping point at which self-reflection yields a more consistent worldview, not less.
>Reality may imply consistency, but consistency does not imply reality.
Say you have different methods of determining "truth." All are restricted to accessing information via the internet, but... some methods yield consistent results, while other methods yield inconsistent / conflicting / non-repeatable results. The more you compare your "internal picture of reality" against these methods, assuming your internal picture isn't too flawed, you can home in on the most reliable methods.
Not just that A, B and C agree with me, so I'll just read A, B and C in the future... more like:
Pass 1: A, B and C have developed reliability ratings of 85, 94, and 72% respectively (continue for hundreds of thousands of sources...)
Pass 2: Drop the lower quartile of sources by reliability ratings - now re-rate the remaining sources considering only those which were not dropped in Pass 1
Pass 3: Again drop the lower quartile of sources by ratings after pass 2, but bring back the top 50% of sources that were dropped in pass 2, if they were truly unreliable, they'll be dropped again in this pass...
Pass 4: Again drop the lower quartile of sources by ratings after pass 3, but bring back the top 50% of sources that were dropped in pass 3....
And, variations thereof, tweaking lower quartile and bring back rates for optimal apparent self consistency.
>Do we merely want the LLM output to be consistent?
Of course not, but, lacking the ability to perform actual experiments on hypotheses beyond meta-studies of published data...
>Or do we want it to be correct?
It can only be as correct as the best data in its training set. If the entire training set is writings by people who believe that the earth is flat, an LLM would be unlikely to derive a better truth from that bulk of publication... although, flat earthers often do publish self-contradictory conclusions, so if there is any non-flat earth support in there at all, even basic geometric measurements of sun and shadow angles, the LLM might very well throw all that out. This is the power of diversity: including many different kinds of source material instead of exclusively focused specialist training. It works for people as well as statistical models.
🌻🌻🌻🌻✌️ [google.com]
(Score: 2) by JoeMerchant on Tuesday July 07, @02:42PM
>>I submit: if the output is even 1% more accurate / reliable / true than the base training set, then each round of re-training should incrementally improve.
>Obvious rebuttal: why is the second time going to be any better than the first time?
Less obvious rebuttal to the rebuttal: diversity.
The current frontier model development landscape has a great deal of diversity. Some of the more successful + efficient models are called MoE: Mixture of Experts. The continuing development and application of different types of BS detectors would give that opportunity for model A trained on virgin data -> model B trained on data + A output -> model C trained on data + B output, etc. the later generation models have the opportunity to score progressively better on hallucination tests than model A did. In a sort of genetic programming contest, the ones that improve continue, while the ones that digress are abandoned. MoE presumably specializes in various domains then feeds questions to the most appropriate expert(s), possibly attempting to build consensus the way early Space Shuttle flight computers (which resulted in a LOT of launch delays) would seek agreement from their independent units to guide operational decisions.
Of course, applications are myriad and legion, so some models are focusing on specific applications and testing for advancement there, while others work on different areas of focus - ignoring or neglecting the AGI brass ring to deliver more actual value to specific users.
One early thing to focus on is: improvement of the specifications and their corresponding tests - as long as they are sloppy, the systems developed to optimize for them will output slop.
The question then becomes: how well are AI systems improving their own specifications? How much improvement are they achieving in isolation, without direct human guidance?
🌻🌻🌻🌻✌️ [google.com]
(Score: 4, Interesting) by coolgopher on Tuesday July 07, @04:09AM (4 children)
The math says no: https://smsk.dev/2026/04/26/ai-cannot-self-improve-and-math-behind-proves-it/ [smsk.dev]
(Score: 2) by JoeMerchant on Tuesday July 07, @04:21AM
I think the application of the label "math" is... a stretch.
🌻🌻🌻🌻✌️ [google.com]
(Score: 1) by khallow on Tuesday July 07, @11:08AM
(Score: 0) by Anonymous Coward on Tuesday July 07, @05:07PM
These things don't exist in isolation slurping up their own poop.
Humans respond too. So there is slop and response to slop in the new slop. Hooman don't like slop gets fed back into the slop generator. Figure that shit out.
(Score: 2) by DadaDoofy on Tuesday July 07, @05:29PM
Absolutely. Nice to know, but just observe unmitigated feedback in any system, an audio amplifier for instance. The end result is always useless noise.
(Score: 1) by shrewdsheep on Tuesday July 07, @10:19AM (1 child)
In principle, statistical regression models, LLMs are an example of, cannot be improved by generating new data from the fitted regression model and adding that data to re-fit the model. The best outcome is that your new model gives the same predictions, only you falsely believe that you are very certain about the predictions. More likely, the model deteriorates with respect to the underlying true model. I believe we had a story about scientific papers around LLM degradation on a re-feed loop.
You can argue, of course, that each LLM is trained on the output of many other models apart from "true" data. The above argument is not fully watertight for ensemble methods that also get to see real new data but it seems applicable enough.
(Score: 3, Interesting) by JoeMerchant on Tuesday July 07, @11:48AM
>ensemble methods that also get to see real new data
Of course this is one way for "new truth" to be introduced into the next generation results.
I believe a logical flaw of the "cannot be improved by generating new data from the fitted regression model and adding that data to re-fit the model" argument rests in 1) assuming the same fitting algorithm is used repeatedly, and 2) looking at the problem as one of "finding truth" rather than one of "rejecting falsehood".
For a well worn analogy, consider the polishing process. A mirror surface is perfectly flat, coherent incident light all reflects at the same angle. A dull, matte surface scatters light, instead of all coherent inbound light following the "one true" outbound trajectory "imperfections" in the surface send it at scattered outbound angles. The polishing process doesn't start with the finest grit, that would take forever to achieve a flat surface, instead the largest imperfections are "taken out" with a coarse grit which itself leaves - and even creates - imperfections in the surface, but smaller ones than the surface started with. For crystal clear plastic, or mirror smooth metal, this process repeats with finer and finer grit, each step smoothing the surface closer to a "true" flat, until the final polish with very fine grit achieves a surface close enough to flat as to be visibly indistinguishable - of course imperfections remain, but at a certain level of polish we can no longer perceive them with our eyes and ordinary light sources.
We don't know perfect truth, the best we have is only what our best methods can discern. It's going to be rare (but still occasionally possible) that a statistical analysis of our existing writings uncovers "new truths" somehow hidden in "plain sight":
Of course, most fields aren't rigorously demonstratable - are subject to endless debate and prejudices/proclivities of the judges.
In the 1990s I developed a (very effective) ensemble method to tease cardiac volume traces out of a whole chest plethysmograph which was dominated by respiratory motions 5 to 20x larger than the cardiac motions. Fortunately, the cardiac motions were highly repetitive and synchronized to their electrically detectible "R-wave", so you simply synch up your time series data to the R-waves and ensemble average it - about 60 to 120 times. This produced a picture of a single cycle of cardiac motion from the average of the previous - say 100 - motions, reducing the (mostly) uncorrelated respiratory motions to ~1% of their former relative amplitude. Some of the art lay in creating a continuous cardiac waveforrm via interpolated stitching together of the averages - and of course the cardiac rate (period) is irregular, constantly varying, so the length of the average picture needs appropriate stretching - no stretching near the R-wave, greater stretching applied further away into the times between R-waves. This method would produce an ensemble curve which nearly perfectly overlapped the raw signal when the subject stopped all breathing motions, yet continued to accurately track the cardiac waveform (including changes induced by exercise and interventions like administration of nitroglycerine) during normal breathing. In 30+ years in and around signal processing fields, I have never found another application for ensemble methods remotely as valuable/illuminating as that one.
One might posit that LLM training could attempt to emulate such a system by identifying a reliable signal analogous to the EKG R-wave, say: peer reviewed publication cited by 100+ studies running 99% or better confirmation, said studies themselves cited and confirmed by at least 10 more peer reviewed papers also confirming their findings - the "truth signal" of the scientific echo chamber if you like - having identified such a signal, the LLM might extrapolate away from that signal to another "truth signal" "averaging" the "signals" found between them to determine the most statistically likely agreement of the content in between the "truth spikes".
Who knows what voodoo they are actually trying in the frontier development labs, probably something like what I just outlined and a thousand other approaches, all being evaluated by arbitrary measures of "success" in a kind of genetic algorithm development powered by meatbags, funded by the billions-trillions in investment the field is currently receiving
🌻🌻🌻🌻✌️ [google.com]
(Score: 4, Insightful) by driverless on Tuesday July 07, @12:58AM (2 children)
This has been written about extensively. Search for the term in the subject heading.
(Score: 2, Informative) by Anonymous Coward on Tuesday July 07, @03:33AM (1 child)
> the bots are starting to ingest their own digital excrement [dshr.org], creating a negative feedback loop.
This is backwards, negative feedback is self-regulating and is used to reduce distortion.
Imo, "Archivist David Rosenthal" needs to take a basic circuit theory class. What he should be saying is positive feedback which leads to a hard-over error, possibly oscillating on the way to the final state.
(Score: 5, Informative) by canopic jug on Tuesday July 07, @09:04AM
What he should be saying is positive feedback which leads to a hard-over error [...]
The wording in the summary is mine, not his. Thus that mistake is mine, not his. Perhaps it would be better to have written, a feedback loop with ever worsening results, or something similar there in the summary.
Money is not free speech. Elections should not be auctions.
(Score: 2) by looorg on Tuesday July 07, @10:55AM (4 children)
So LLM has finally, ok it's been going on for a while, turned into the human centipede. Feeding on it's own output. Things are going downhill from here on. The only interesting aspect is how this will change human interaction and language. Humans will adopt to LLM output as if it was the new norm or correct.
(Score: 0) by Anonymous Coward on Tuesday July 07, @11:23AM (1 child)
This might be the word you are looking for? https://en.wikipedia.org/wiki/Ouroboros [wikipedia.org]
> The ouroboros (/ˌʊərəˈbɒrəs/[2]) or uroboros (/ˌjʊərəˈbɒrəs/[3]) is an ancient symbol depicting a snake or dragon[4] eating its own tail. The ouroboros entered Western tradition via ancient Egyptian iconography and the Greek magical tradition. It was adopted as a symbol in Gnosticism, Hermeticism, and alchemy.
I'm ready to lump LLM's in with Gnosticism, Hermeticism, and alchemy if you are!!
(Score: 2) by looorg on Tuesday July 07, @12:16PM
There is that. Mine was more a reference to the horror movie -- the human centipede -- https://en.wikipedia.org/wiki/The_Human_Centipede_(First_Sequence) [wikipedia.org] . I do not recommend it unless you are into that sort of thing. I wish I had not seen it. It now lives rent free in my mind forever.
(Score: 2) by VLM on Tuesday July 07, @03:04PM (1 child)
It'll be friction based as with the spread of all memes.
Internal corporate emails will be essentially impossible to translate even with AI assistance because the friction is extremely high. It'll be full on slop warfare at big companies. A strange side effect is this will make big companies even less productive than they already are, making AI a net benefit to small companies and contractors.
With lower friction areas like media, slop will be abandoned by most of the population. Look at the low popularity of legacy media. It'll never really die because of the addicts and superfans (generally one and the same) but the general public avoid it like the plague. For example capeshit movies will become even less comprehendible to the general public and even less popular, but they'll never go away because they will not need to pay scriptwriters or actors.
There will be strange side effects that I think are hard to predict. From the capeshit movies example above, at least in the early days, licensed IP fees will go UP not down because "they don't need to pay for everyone who got fired by AI" so they have money to harvest. Eventually as revenue dries up the licensed IP fees will have to drop in parallel, but not in the short term.
Another weird side effect will be "side issue slop". Hobby Lobby is not directly in the business of slop. But have you set foot in one lately? Every piece of "decor" they sell looks like AI generated slop now. Its not just a genre thing like homemade vs highly detailed machine made. Its really sloppy in there now. Likewise, you can expect architects to be wiped out at the low end so expect megacorporate franchise chains to look very sloppy.
(Score: 1, Insightful) by Anonymous Coward on Tuesday July 07, @05:39PM
It's the fructose corn syrup of white collar work.
(Score: 0, Redundant) by QuickButterfly on Wednesday July 08, @04:05AM (1 child)
I'm so sorry to be that guy. This is a classic example of positive, not negative, feedback.
(Sorry).
(Sorry).
(Score: 1, Informative) by Anonymous Coward on Wednesday July 08, @12:29PM
> This is a classic example of positive, not negative, feedback.
It has already been addressed in previous comments:
https://soylentnews.org/article.pl?sid=26/07/05/1921216#1447511 [soylentnews.org]