Logo
    Search

    “Model collapse” shows AI doesn’t have the human touch, writer says

    enSeptember 21, 2023

    Podcast Summary

    • The risk of AI-generated text becoming increasingly inhuman and nonsensical due to model collapseResearchers discovered that AI language models can start generating increasingly strange and incoherent text after a certain number of iterations, leading to a potential proliferation of misinformation and nonsensical content on the web.

      As AI-generated text becomes more prevalent on the web, there's a risk of a phenomenon called model collapse, which occurs when the AI starts generating increasingly inhuman and nonsensical text. This happens because the AI is trained on its own previous outputs, which can contain small inconsistencies or inhumanities. These inconsistencies get compounded and amplified with each new generation of AI output. Clive Thompson, a contributing writer for the New York Times Magazine and Wired, explained this concept using an analogy of photocopying a famous painting. Just as a photocopy of a painting starts to look weird after several generations due to the accumulation of small errors, AI output can start to make less and less sense as it's fed back into the models that train it. Researchers discovered this phenomenon by observing the behavior of AI language models. They noticed that after a certain number of iterations, the output started to become increasingly strange and incoherent. This can lead to a snowball effect, where the AI's output becomes more and more inhuman and nonsensical with each new iteration. The implications of model collapse are significant, as it could lead to a proliferation of misinformation and nonsensical content on the web. It's important for researchers and developers to understand this phenomenon and find ways to mitigate its effects to ensure that AI-generated content remains useful and reliable.

    • Impact of AI-generated content on language model trainingAI models trained on inaccurate or misleading information can produce increasingly incoherent output, highlighting the importance of maintaining high-quality training data

      The researchers explored the potential impact of AI-generated content on the training of future language models. They simulated this scenario by creating a model that was trained on a mix of human-written and AI-generated text. Over several generations, they observed the degradation of the model's output as errors compounded. The final output, after nine generations, was essentially gibberish and no longer related to the original prompt. This experiment illustrates the risks of AI models being trained on inaccurate or misleading information, leading to the production of increasingly incoherent output. The researchers' findings underscore the importance of maintaining high-quality training data to ensure the reliability and accuracy of AI models.

    • Model collapse in large language modelsModel collapse in large language models could lead to unpredictable and incorrect responses, potentially affecting their performance and accuracy over time. Companies may intervene by providing new content or saving older training data to prevent further deterioration.

      The concept of model collapse in large language models raises concerns about their potential deterioration in performance and accuracy over time. As models become more complex and are trained on larger datasets, they may start to generate increasingly erratic and unpredictable responses. This could lead to a decline in their ability to understand and generate human-like language. The 7th and 9th generation language models were mentioned as examples of this phenomenon, with the 9th generation model's output described as "babbling about jack rabbits" and a poor representation of the original Mona Lisa. The implications of model collapse are significant, as it could result in models that provide even more unpredictable and incorrect responses than they do now. However, it's unlikely that companies developing these models will let them deteriorate without intervention. They may choose to pay humans to write new content for the models, or use techniques like saving older training data to refresh the models and prevent further deterioration. Ultimately, model collapse underscores the importance of human communication and the value it brings. While language models can mimic human language to a certain extent, they cannot replicate the depth, nuance, and creativity of real human expression.

    • The importance of human language in creating effective AIHuman language, written by humans, is crucial for developing good AI. The unique qualities humans possess can't be replicated by AIs yet.

      Creating effective AI requires authentic human language. This was emphasized by tech writer Clive Thompson during a discussion on Marketplace Tech. He highlighted the importance of real prose written by humans in developing good AI. Thompson also pointed out that there seems to be a unique quality or "lightning in a bottle" that humans possess, which AIs don't have yet – we're not sure if they ever will. Meanwhile, for kids trying to understand the complexities of the world, there's Million Bazillion, a Webby-winning podcast from Marketplace. Each episode tackles the awkward and intriguing questions kids have about money. Some examples include "What is a college account, and how does it work?" or "What are unions, and what are they for?" By answering these questions, Million Bazillion helps kids grasp the role money plays in our world. Listen to Million Bazillion wherever you get your podcasts.

    Recent Episodes from Marketplace Tech

    Lessons to learn from the massive CrowdStrike outage

    Lessons to learn from the massive CrowdStrike outage

    Last Friday felt like something out of a Y2K nightmare after the cybersecurity company CrowdStrike, pushed a software update to all its clients — including health care systems, banks and the federal government — that ended up crashing computer systems worldwide. The fallout is still being felt, particularly in the travel sector, as airliners try to reschedule canceled flights while trying to get everything back to normal. It’s also become something of a reminder that the internet and a lot of the online services we rely on are delicate. Marketplace’s Lily Jamali spoke with Kate Conger, a reporter at The New York Times who recently wrote about this with her colleague David Streitfeld.

    Marketplace Tech
    enJuly 23, 2024

    EU regulators request information on Amazon’s algorithms

    EU regulators request information on Amazon’s algorithms

    Online sales in the U.S. surpassed $14 billion during Amazon Prime Day last week, according to Adobe Analytics. Amazon’s heft and promotional power continue to drive sales, even for rivals, during the shopping jamboree. But in Europe, an important market for the e-commerce giant, lawmakers have become increasingly sensitive to Amazon’s relations with its rivals, as well as its partners and customers. They’ve requested that Amazon hand over information about its product recommendation algorithms, along with data on ads, by Friday. It’s part of compliance with the European Union’s Digital Services Act, a sweeping set of tech regulations that took effect in recent years. Marketplace’s Meghan McCarty Carino discussed it with Theo Wayt, who covers Amazon for The Information.

    Marketplace Tech
    enJuly 22, 2024

    Tech Bytes – Week in Review: Vance’s Silicon Valley ties, Prime Day injuries and Starbucks bets on EVs

    Tech Bytes – Week in Review: Vance’s Silicon Valley ties, Prime Day injuries and Starbucks bets on EVs

    A new Senate report finds Amazon Prime Day is prime time for warehouse injuries. Plus, Starbucks is teaming up with Mercedes-Benz to supercharge electric vehicle infrastructure. But first, several Silicon Valley billionaires have thrown their support behind former President Donald Trump in his quest to reclaim the White House, thanks in part to his pick for vice president, Sen. J.D. Vance. Marketplace’s Meghan McCarty Carino spoke with Jewel Burks Solomon, a managing partner at Collab Capital, for her take on these stories in Marketplace Tech Bytes: Week in Review.

    Marketplace Tech
    enJuly 19, 2024

    How to deal with misinformation about the assassination attempt on Donald Trump

    How to deal with misinformation about the assassination attempt on Donald Trump

    FBI officials are still looking into what motivated the 20-year-old gunman who attempted to assassinate former President Donald Trump on Saturday. As of this episode, investigators have yet to publicly share any conclusions about his reasons for the attack, which killed a rally attendee and injured the former president and two others. But the lack of information didn’t stop misinformation from flooding online channels. Marketplace’s senior Washington correspondent, Kimberly Adams, speaks with Molly Dwyer, director of insights at PeakMetrics, and Lisa Fazio, associate professor of psychology at Vanderbilt University, about the false narratives surrounding the shooting and how to separate fact from fiction in the aftermath of a violent event.

    This conversation is part of “Marketplace Tech’s” limited series “Decoding Democracy.” Watch the full episode here or on our YouTube channel.

    Marketplace Tech
    enJuly 18, 2024

    The complicated reality of school cellphone bans

    The complicated reality of school cellphone bans

    Pop quiz: What’s a policy supported by political rivals in California and Florida? The answer is banning cellphones in school. Florida is among a handful of states that have restricted mobile devices in the classroom. California has not, though Gov. Gavin Newsom has pushed the Legislature to act. The policies are intended to reduce distraction and mitigate addiction and other mental health concerns attributed to phone use. But Liz Kolb, a clinical professor of education at the University of Michigan, tells Marketplace’s Meghan McCarty Carino it’s not that simple.

    Pop quiz: What’s a policy supported by political rivals in California and Florida? The answer is banning cellphones in school. Florida is among a handful of states that have restricted mobile devices in the classroom. California has not, though Gov. Gavin Newsom has pushed the Legislature to act. The policies are intended to reduce distraction and mitigate addiction and other mental health concerns attributed to phone use. But Liz Kolb, a clinical professor of education at the University of Michigan, tells Marketplace’s Meghan McCarty Carino it’s not that simple.

    Marketplace Tech
    enJuly 17, 2024

    The digital surveillance of transgender people

    The digital surveillance of transgender people

    In recent years, we’ve seen a surge of state laws and policies affecting trans people. Half of states have banned or restricted gender affirming care for minors, with some adding restrictions for adults. The ACLU is tracking more than 500 bills that have been introduced across the country. The enforcement of such laws, as with recent bans on abortion and related reproductive care, have raised concerns about tracking people’s digital footprints. So much of daily life is conducted online, and there are currently no federal data privacy protections. KB Brookins, a writer based in Austin, Texas, wrote about a personal experience that drove home concerns about their trail of digital data.

    Marketplace Tech
    enJuly 16, 2024

    Online speech cases sent back to the lower courts

    Online speech cases sent back to the lower courts

    Sandwiched between some blockbuster Supreme Court rulings last month came a decision — or more so, a non-decision — that is reverberating through the tech world. NetChoice, big tech’s lobbying arm, challenged a pair of laws in Florida and Texas that sought to restrict how social media platforms moderate content. The high court kicked both cases back to lower courts with some added commentary. Marketplace’s Meghan McCarty Carino spoke with Lauren Feiner, senior policy reporter with the Verge, who wrote about what this means for future attempts to regulate tech.

    Marketplace Tech
    enJuly 15, 2024

    Bytes: Week in Review — FTC’s latest tech crackdown, Trump’s pro-crypto campaign and Threads turns a year old

    Bytes: Week in Review — FTC’s latest tech crackdown, Trump’s pro-crypto campaign and Threads turns a year old

    The Republican Party officially adopted former President Donald Trump’s 2024 platform this week. The GOP is now taking a friendly approach to cryptocurrency and artificial intelligence, with plans to roll back regulation of both. Plus, Meta’s Threads platform celebrates its one-year anniversary this month. But first, the Federal Trade Commission is cracking down on the popular messaging app NGL. Marketplace’s Meghan McCarty Carino is joined by Natasha Mascarenhas, reporter at The Information, for her take on this week’s tech news.

    Marketplace Tech
    enJuly 12, 2024

    More website links are expiring. Is it a bug or a feature of the internet?

    More website links are expiring. Is it a bug or a feature of the internet?

    The internet is full of all manner of unsavoriness that is surely corroding our minds and societies. But the kind of rot we’re talking about here is link rot — the disappearance of online content when links turn into “404 Page Not Found.” A recent study from Pew Research suggests almost 40% of all webpages that existed in 2013 are no longer accessible. That includes important government links, citations on Wikipedia and hyperlinks in news articles. Marketplace’s Lily Jamali recently talked about this with Clare Stanton, product and research manager at Harvard Law School’s Library Innovation Lab, who also works on a webpage preservation project, perma.cc.

    Marketplace Tech
    enJuly 11, 2024

    Progressive TikTok creators turn against Biden

    Progressive TikTok creators turn against Biden

    About a third of adults under 30 regularly get their news on TikTok, according to the Pew Research Center. And in this election season, the messages from young, left-leaning creators on the short-form video app are pretty different from last time around. In 2020, a coalition of influencers united to back presidential candidate Joe Biden’s campaign, and historically high youth turnout helped propel him to a win. But after almost four years of the Biden presidency, the TikTok tide has turned, according to Taylor Lorenz, online culture columnist at The Washington Post, who recently wrote about this shift.

    Marketplace Tech
    enJuly 10, 2024

    Related Episodes

    How an AI pope pic fooled us

    How an AI pope pic fooled us
    An AI-generated image of Cool Pope in immaculate drip went viral over the weekend and most everyone thought it was real. The Verge’s James Vincent explains how we should navigate our new internet reality. This episode was produced by Amanda Lewellyn, edited by Matt Collette, fact-checked by Avishay Artsy and Siona Peterous, engineered by Paul Robert Mounsey, and hosted by Sean Rameswaram.  Transcript at vox.com/todayexplained   Support Today, Explained by making a financial contribution to Vox! bit.ly/givepodcasts Learn more about your ad choices. Visit podcastchoices.com/adchoices

    A Clockwork Metaverse - (Hour 2)

    A Clockwork Metaverse - (Hour 2)
    Alternate Current Radio Presents: BOILER ROOM - Learn to protect yourself from predatory mass media cartels and social media feedback loops.

    Hour 2
    In the second hour of the podcast the BOILER ROOM is looking at the Neil Young vs Joe Rogan story is unfolding in the wake of the outrage at Joe Rogan interviewing Dr. Robert Malone. Who does the mass media and fortune 100 cartels look to as voices of analysis for this...? Well, a Unicef spokesperson and a one-hit-wonder viral video 'anti-golf' tiktok personality who's now tried to rebrand herself as a "anti-conspiracy researcher." Emperor Fauci is on the jab path again, pushing for jabbing children under the age of five years old with experimental biotech gene therapy mRNA drugs, just following the W.H.O. coming out and saying there was little evidence that this type of experimenting on children is helpful at all. Whats up with mass media platforming someone lightly roasting Schwab and the Davos crowd? Some discerning readers might think they backhandedly want to paint swaths of people as "antidemocratic populists."

    Featured guests: Ruckus (The Daily Ruckus), Oddman Out The Oddcast), Infidel Pharaoh (Magic Carpet Ride), Mindspace Art (Soul Purpose)

    Hour 2 Reference Links:

    An Open Letter to Spotify: A call from the global scientific and medical communities to implement a misinformation policy
    https://spotifyopenletter.wordpress.com/2022/01/10/an-open-letter-to-spotify/

    ‘A Menace to Public Health’: Doctors Demand Spotify Puts an End to Covid Lies on ‘Joe Rogan Experience’
    https://www.rollingstone.com/culture/culture-news/covid-misinformation-joe-rogan-spotify-petition-1282240/

    Ministry of Truth on Abbie Richards
    https://en.wikipedia.org/wiki/Abbie_Richards

    Unicef wants your money
    https://www.unicefusa.org/support-campaign-end-pandemic

    The High Cost of Disparaging Natural Immunity to Covid
    https://www.wsj.com/articles/the-high-cost-of-disparaging-natural-immunity-to-covid-vaccine-mandates-protests-fire-rehire-employment-11643214336

    HOW KLAUS SCHWAB BUILT A BILLIONAIRE CIRCUS AT DAVOS
    https://www.vanityfair.com/news/2022/01/how-klaus-schwab-built-a-billionaire-circus-at-davos