Blurblogs
Global
Popular
Log in
MrMarchant Stuff
519 stories
·
0 followers

The Collapse of GPT by Neil Savage
Thursday May 15th, 2025 at 3:24 PM

Communications of the ACM
1 Share

Ever since ChatGPT was released to the public in November 2022, people have been using it to generate text, from emails to blog posts to bad poetry, much of which they post online. Since that release, the companies that build the large language models (LLMs) on which such chatbots are based—such as OpenAI’s GPT 3.5, the technology underlying ChatGPT—have also continued to put out newer versions of their models, training them with new text data, some of which they scraped off the Web. That means, inevitably, that some of the training data used to create LLMs did not come from humans, but from the LLMs themselves.

That has led computer scientists to worry about a phenomenon they call model collapse. Basically, model collapse happens when the training data no longer matches real-world data, leading the new LLM to produce gibberish, in a 21st-century version of the classic computer aphorism “garbage in, garbage out.”

LLMs work by learning the statistical distribution of so-called tokens—words or parts of words—within a language by examining billions of sentences garnered from sources including book databases, Wikipedia, and the Common Crawl dataset, a collection of material gathered from the Internet. An LLM, for instance, will figure out how often the word “president” is associated with the word “Obama” versus “Trump” versus “Hair Club for Men.” Then, when prompted by a request, it will produce words that it reasons have the highest probability of meeting that request and of following from previous words. The results bear a credible resemblance to human-written text.

Model collapse is basically a statistical problem, said Sanmi Koyejo, an assistant professor of computer science at Stanford University. When machine-generated text replaces human-generated text, the distribution of tokens no longer matches the natural distribution produced by humans. As a result, the training data for a new round of modeling does not match the real world, and the new model’s output gets worse. “The thing we’re worried about is that the distribution of your data that you end up with, if you’re trying to fit your model, ends up really far from the actual distribution that generated the data,” he said.

The problem arises because whatever text the LLM generates would be, at most, a subsample of the sentences on which it was trained. “Because you generate a finite sample, you have some probability of not sampling them,” said Yarin Gal, an associate professor of machine learning at Oxford University. “Once you don’t sample, then they disappear. They will never appear again. So every time you generate data, you basically start forgetting more and more of the tail events and therefore that leads to the concentration of the higher probability events.” Gal and his colleagues published a study in Nature in July that showed indiscriminate use of what they called ‘recursively generated data’ caused the models to fail.

The problem is not limited to LLMs. Any generative model that is iteratively trained can suffer the same fate if it starts ingesting machine-produced data, Gal says. That includes stable diffusion models that create images, such as Dall-E. The issue also can affect variational autoencoders, which create new data samples by producing variations of their original data. It can apply to Gaussian mixture models, a form of unsupervised machine learning that sorts subpopulations of data into clusters; they are used to analyze customer preferences, predict stock prices, and analyze gene expression.

Collapse is not a danger for models that incorporate synthetic data but only do so once, such as neural networks used to identify cancer in medical images, where synthetic data was used to augment rare or expensive real data. “The main distinction is that model collapse happens when you have multiple steps, where each step depends on the output from the previous step,” Gal said.

The theory that replacing training data with synthetic data will quickly lead to the demise of LLMs is sound, Koyejo said. In practice, however, not all human data gets replaced immediately. Instead, when the generated text is scraped from the Internet, it gets mixed in with human text. “You create synthetic data, you add that to real data, so you now have more data, which is real data plus synthetic data,” he said. What is actually happening, he said, is not data replacement, but data accumulation. That slows the degradation of the dataset.

Simply accumulating data may stop model collapse but can cause other problems if done without thought, said Yunzhen Feng, a Ph.D. student at the Center for Data Science at New York University. As a rule, the performance of neural networks improves as their size increases. Naively mixing real and synthetic data together, however, can slow that improvement. “You can still obtain similar performance, but you need much more data. That means you’re using much more compute and much more money to achieve that,” he said.

One challenge is that there is no easy way to tell whether text found on the Internet is synthetic or human-generated. Though there have been attempts to automatically identify text from LLMs, none have been entirely successful. Research into this problem is ongoing, Gal said.

Solving with curation

There are ways, however, to make the addition of synthetic data less of a problem.

One approach is to curate the synthetic data to make sure it is of good quality. Some curation happens naturally, Gal said; people do not post everything their chatbot creates to the Internet, weeding out the material that contains false information or simply does not make sense, so that improves the training set.

Curation can also be a deliberate process to make sure high-quality data goes into a training set. Feng, for instance, has experimented with asking the LLM to assess the quality of its own output. LLMs naturally select the words they think have the highest probability of fitting into a context. In doing so, they internally generate a score rating how confident they are that they are pairing the best words together. That same mechanism can be used to assess already generated text to rate its quality, with low-scoring results removed or the highest-scoring result of several attempts selected as the best. The idea is similar to a method used to fine-tune LLMs called reinforcement learning from human feedback (RLHF), in which people provide examples of good results, thereby pushing the models toward producing similar results. In this case, though, the LLM is generating its own feedback.

How well that works varies by case, Feng said. The feedback can be improved by having other LLMs assess the same text and combining the results from different models. Including human assessments also improves the outcomes, as does applying some pre-written rules about what the output should look like. Eliminating lower-quality results from the synthetic data makes the generated data more closely resemble original data, he said. “It’s like you have a distribution of the synthetic data, you have a distribution of the real data, and you want to close the gap between them as much as possible,” he said.

Improving the quality of synthetic data could also help with another challenge LLMs are facing as they try to improve: a dearth of new data on which to train. Scientists from Epoch AI, a research institute that focuses on trends in AI, have predicted the world will run out of new text to train on sometime between 2026 and 2032. With no new data on which to train future generations of LLMs, progress could stagnate. “The interesting question is, can synthetic data lead to not just stagnation but actual improvement in the model?” asked Pablo Villalobos, a staff researcher at Epoch.

With curation of high-quality synthetic data, he said, the question becomes “whether this can be done iteratively so that each model generates better data that is used to train another model in basically the opposite of model collapse, in some virtuous circle.” He is not yet sure whether such improvement is possible, but sees some signs it could be.

Other issues arise from training new models on generated data that do not quite reach the level of model collapse. For instance, Koyejo said, synthetic data could increase the likelihood that LLMs will discriminate against people in minority groups. Because any minority is by definition a smaller part of the data distribution, losing the tails of the distribution could make minorities disappear entirely. “Data tends to anchor on majority subgroups,” he said. “It tends to be good at capturing the most popular themes and less good at capturing tails. So less represented demographics can get erased in various ways.”

While such erasure is something that could happen, he added, the issue has not been well studied. His colleague Diyi Yang, an assistant professor in the natural language processing group at Stanford, said there has been very little research into the question of how model collapse affects diversity issues. “Part of the reason is that, if you think about any existing big models, a lot of the training dynamics or checkpoints of those models actually are not really transparent or publicly available,” she said.

In the end, Gal argued, model collapse is an important consideration, but not the matter of imminent disaster that some news coverage has made it out to be. “It’s a matter for the tech companies who build these models to be aware of how the models are being used and how the models are being trained, in order to avoid training on synthetic data that they themselves generated.”

Further Reading

Read the whole story
mrmarchant
1 hour ago
reply
Share this story
Delete

Weird and Wondrous Sea Cucumbers by The Editors
Thursday May 15th, 2025 at 2:03 PM

JSTOR Daily
1 Share

Few animals seem less animal-like than sea cucumbers. The long, slender beings often appear motionless like the vegetable they resemble as they bask on the seafloor. They lack a brain, eyes or obvious features apart from a mouth and anus.

Neither are sea cucumbers especially appealing to human senses. Their bodies, often drab gray or brown, can be wart-studded and rough or slimy to the touch. “They’re not charismatic animals,” admits Annie Mercier, a researcher in marine biology at Memorial University in Canada. That doesn’t stop her from loving them.

Scientists like Mercier are learning there’s a lot more to these simple creatures than meets the eye. Some care for their young, others reproduce by dividing in two—they have an impressive ability to regrow lost tissues. To boot, many of the 1,700-plus species of sea cucumber are important for ocean health because they hoover up ecosystem-smothering algae and bacteria. “If you look at the diversity of shapes and forms and behaviors across all the sea cucumbers that exist, it’s pretty incredible,” says Mercier, who coauthored a 2025 article on the creatures in the Annual Review of Marine Science.

But many are rapidly vanishing. About eighty species are harvested commercially. Dried sea cucumbers are especially prized in East Asian cuisines; some species are being fished to the brink of extinction and are targets of illegal trade.

Scientists hope that better understanding sea cucumbers and the threats they face will help to forge paths to protect them.

To Be a Sea Cucumber

Sea cucumbers are echinoderms, invertebrates including starfish and sea urchins that have a five-rayed body arrangement. One could think of a sea cucumber as an elongated urchin that has largely lost its crunchy external skeleton. It has two openings: the food-ingesting mouth and the waste-excreting cloaca, or anus, which also hosts a breathing organ for extracting oxygen from seawater.

Beyond this basic form, sea cucumbers are astonishingly diverse. Many are brightly colored. Some are transparent, others twinkle with bioluminescent light. Some are shorter than a fingernail, others as long as a person is tall.

Though their larvae can swim, adults are slow-moving, using sticky tube feet to peruse the surfaces they meander over. Some arch their bodies upwards, extending feathery tentacles into the water to capture plankton. Others rhythmically burrow through and ingest seafloor sediment, extracting bacteria and algae and leaving moundlike trails of cleaner excreted material as they go.

With these laborious methods of locomotion, sea cucumbers were long thought to never exceed speeds of about a yard a day. But scientists recently spotted them detaching from the seafloor and moving much faster. For instance, a deep-sea species called the headless chicken sea monster was observed using winglike veils to swim, “like ethereal aquatic angels,” as noted in The World of Sea Cucumbers, a 2023 book cowritten and coedited by Mercier.

And in 2019, Mercier’s lab discovered that the North Atlantic orange-footed sea cucumber and an Indo-Pacific species known as the sandfish can float by taking up water through the mouth and anus, making themselves more buoyant. Detaching from the seafloor, they can bob along in the current for hundreds of meters and escape crowded conditions or dirty or not-salty-enough waters. “They can essentially move away from a situation that is unfavorable to them,” Mercier says.

More to Explore

Copepod

The Little Plankton That Could

Sierra Garcia
July 18, 2021
Arguably the world's most abundant animal, calanoid copepods can leap like whoa Nelly. And check out their enigmatic embryos!

Sea cucumbers have similarly diverse strategies for reproduction. In some species, the males release pheromones to attract aggregations of other sea cucumbers. The females release thousands to millions of eggs from a hole toward the front of their bodies; these are then fertilized by sperm. In other species, eggs stay attached to the female; she nourishes and protects the developing embryos.

And, about 20 species “literally just divide into two,” says marine biologist Sven Uthicke at the Australian Institute of Marine Science. In a remarkable feat of regeneration, the separated sea cucumber halves regrow the missing portion of themselves within weeks.

Sea cucumbers use their regenerative powers for survival, too. Some species spit out their guts though their mouths, “which might be a way to get rid of pollution,” Uthicke says. Others excrete parts of their innards through their anus when attacked, in order to distract predators. In both instances, they neatly regenerate the discarded body parts.

And when the silky sea cucumber senses a bite or tear, it pinches off the injured body part within 30 seconds. Other species simply dissolve their body walls when grabbed and removed from the water.

Beyond their alluring biology, sea cucumbers are critical ocean denizens—and not just as prey. Their burrowing and sediment-gobbling habits aerate the seafloor, making it a healthier habitat for crabs, mollusks and worms. They do this at massive scale: One 2021 study on an Australian reef estimated that lollyfish, or black sea cucumbers, turn around 71,000 tons of sediment per year, heavier than five Eiffel Towers.

By vacuuming microbes out of the sand, sea cucumbers reduce the growth of ecosystem-smothering algae as well as nasty bacteria that cause coral disease: A 2024 experiment found that removing sea cucumbers on two Pacific reefs led to a surge in sickness and death among many staghorn corals. Sediment-gobbling also releases nutrients like nitrogen, phosphorus and calcium carbonate from the sand, to be used by other animals.

On top of this, many sea cucumbers host dozens of life-forms in and on their bodies—including the cloaca-inhabiting pearlfish. Sea cucumbers “are not just an animal,” says marine biologist Chantal Conand, an honorary associate at the National Museum of Natural History in Paris. “They have a whole ecosystem living with them.”

Conserving these animals is no easy task. Demand for dried sea cucumbers has skyrocketed, especially in China, where they’re considered to have health-boosting properties and are traditionally served as a symbol of status and generosity. As the Chinese population has become wealthier, “the market is growing for these products,” says marine biologist Alessandro Lovatelli, an aquaculture officer at the Food and Agriculture Organization of the United Nations.

While demand initially focused on temperate species along the Chinese coast, burgeoning demand has been driving traders to look further afield. Sea cucumber harvesting for East Asian buyers is now an important livelihood in South America, Africa, the Pacific and the Caribbean. The animals are easy to catch, dry and store, and can fetch hundreds of dollars per pound, driving locals to harvest many tropical species to the brink. As they become rarer, fishermen are using more dangerous diving tactics to get at deeper-living sea cucumbers. In Mexico, fishermen have engaged in violent disputes over sea-cucumber fishing territories.

Reversing these trends is difficult, especially since there are sparse data on species declines, which are key to getting them legal protections. And even where species have been assessed as endangered and placed on lists that prohibit their trade, illegal smuggling persists.

Lovatelli and his colleagues have helped many countries to develop management plans for sustainable sea cucumber harvesting—but many populations have collapsed regardless because governments lack the resources to enforce fishing rules. Lovatelli’s team has also crafted guides to help custom officers catch endangered species crossing borders, but many exports from small islands are collected by boats and bypass customs.

Some countries have started farming sea cucumbers in contained areas, but that hasn’t entirely stopped the harvesting of wild animals. “It’s a difficult industry to control,” Lovatelli says.

It will take concerted effort by international organizations, governments and fisheries to ensure that threatened sea cucumbers continue to exist. But at least there’s now an uptick of interest in studying and protecting these ugly yet prized creatures of the sea, and that gives Mercier hope. “I think that’s definitely a step in the right direction,” she says.

This article is republished from Knowable Magazine under a Creative Commons license. Read the original article.

Support JSTOR Daily! Join our membership program on Patreon today.

The post Weird and Wondrous Sea Cucumbers appeared first on JSTOR Daily.

Read the whole story
mrmarchant
2 hours ago
reply
Share this story
Delete

American Schools Were Deeply Unprepared for ChatGPT, Public Records Show by Jason Koebler
Thursday May 15th, 2025 at 12:40 PM

404 Media
2 Shares
📄
This article was primarily reported using public records requests. We are making it available to all readers as a public service. FOIA reporting can be expensive, please consider subscribing to 404 Media to support this work. Or send us a one time donation via our tip jar here.
American Schools Were Deeply Unprepared for ChatGPT, Public Records Show

In February 2023, a brief national scandal erupted: Several students at a high school in Florida were accused of using a tool called “ChatGPT” to write their essays. The tool was four months old at the time, and it already seemed like a technology that, at the very least, students would try to cheat with. That scandal now feels incredibly quaint.

Immediately after that story broke, I filed 60 public records requests with state departments of education and a few major local school districts to learn more about how—and if—they were training teachers to think about ChatGPT and generative AI. Over the last few years, I have gotten back thousands of pages of documents from all over the country that show, at least in the early days, a total crapshoot: Some states claimed that they had not thought about ChatGPT at all, while other state departments of education brought in consulting firms to give trainings to teachers and principals about how to use ChatGPT in the classroom. Some of the trainings were given by explicitly pro-AI organizations and authors, and organizations backed by tech companies. The documents, taken in their totality, show that American public schools were wildly unprepared for students’ widespread adoption of ChatGPT, which has since become one of the biggest struggles in American education.

Last week, New York magazine ran an article called “Everyone Is Cheating Their Way Through College,” which is full of anecdotes about how generative AI and ChatGPT in particular has become ubiquitous in the education system, and how some students are using it to do essentially all of their work for them. This is creating a class of students who are “functionally illiterate,” one expert told New York. In the years since generative AI was introduced, we’ve written endlessly about how companies, spammers, and some workers have become completely reliant on AI to do basic tasks for them. Society as a whole has not done a very good job of resisting generative AI because big tech companies have become insistent on shoving it down our throats, and so it is asking a lot for an underfunded and overtaxed public school system to police its use.

The documents I obtained are a snapshot in time: They are from the first few months after ChatGPT was released in November 2022. AI and ChatGPT in particular have obviously escaped containment and it’s not clear that anything schools did would have prevented AI from radically changing education. At the time I filed these public records requests, it was possible to capture everything being said about ChatGPT by school districts; now, its use is so commonplace that doing this would be impossible because my request would encompass so many documents it would be considered “overbroad” by any public records officer. All documents and emails referenced in this article are from January, February, or March 2023, though in some cases it took years for the public records officers to actually send me the documents.

💡
Are you a teacher? I want to hear how AI has affected your classroom and how your students use it. Using a non-work device, you can message me securely on Signal at jason.404. Otherwise, send me an email at jason@404media.co.

And yet, the documents we obtained showed that, in the early days of ChatGPT, some state and local school districts brought in pro-AI consultants to give presentations that largely encouraged teachers to use generative AI in their classrooms. Each of these presentations noted potential “challenges” with the technology but none of them anticipated anything as extreme as what is described in the New York magazine article or as troublesome as what I have heard anecdotally from my friends who are teachers, who say that some students rely almost entirely on ChatGPT to make it through school.

American Schools Were Deeply Unprepared for ChatGPT, Public Records Show
A slide from ChatGPT and AI in Education
American Schools Were Deeply Unprepared for ChatGPT, Public Records Show
A slide from ChatGPT and AI in Education
American Schools Were Deeply Unprepared for ChatGPT, Public Records Show
A slide from ChatGPT and AI in Education
American Schools Were Deeply Unprepared for ChatGPT, Public Records Show
An excerpt from a slide from ChatGPT and AI in Education
American Schools Were Deeply Unprepared for ChatGPT, Public Records Show
A slide from ChatGPT and AI in Education

For example, the Louisiana Department of Education sent me a presentation it said it consulted called “ChatGPT and AI in Education,” made by Holly Clark, the author of The AI Infused Classroom, Ken Shelton, the author of The Promises and Perils of AI in Education, and Matt Miller, the author of AI for Educators. The presentation includes slides that say AI “is like giving a computer a brain so it can learn and make decisions on its own,” note that “it’s time to rethink ‘plagiarism’ and ‘cheating,’” alongside a graph of how students can use AI to help them write essays, “20 ways to use ChatGPT in the classroom,” and “Warning: Going back to writing essays—only in class—can hurt struggling learners and doesn’t get our kids ready for their future.” One graphic in the slide imagines a chatbot conversation where a teacher says “AI please grade my papers,” and the bot says “On it! Give me 30 seconds.” It also includes this slide:

American Schools Were Deeply Unprepared for ChatGPT, Public Records Show
A slide from ChatGPT and AI in Education via BonnieChelette

New York City’s Department of Education initially banned ChatGPT before quickly unbanning it. Records obtained by 404 Media show that Project Recess, a “Google for Education Partner,” gave the education system a presentation called “ChatGPT: The Convenience and Challenge of A.I. in the Classroom,” which is about the “capabilities and potential benefits and challenges of using ChatGPT for education. By the end of the workshop participants will be able to start exploring ways to incorporate ChatGPT into their teaching.” 

American Schools Were Deeply Unprepared for ChatGPT, Public Records Show
A slide from ChatGPT: The Convenience and Challenge of A.I. in the Classroom
American Schools Were Deeply Unprepared for ChatGPT, Public Records Show
A slide from ChatGPT: The Convenience and Challenge of A.I. in the Classroom

It notes ChatGPT is “similar to the personal computer, high speed internet, [and] the iPhone” and says “ChatGPT can assist students in improving their writing skills by providing suggestions on grammar, spelling, and style. Additionally, it can help students generate ideas and outlines for their writing assignments,” and suggests that teachers can use ChatGPT to “generate customized quizzes and exams for teachers to use in the classroom,” quotes a New York Times Kevin Roose article about ChatGPT’s “potential for mental health services,” and said it can be used “to evaluate student writing and provide feedback on grammar, sentence structure, and content.” It also does note the potential for cheating and plagiarism, but the presentation ends with a guide for teachers to request it to be unblocked in their classrooms. 

The New York State Education Department, meanwhile, put together a relatively straightforward informational presentation called ChatGPT & Generative AI, which listed several generative AI concerns such as “How can we ensure students are submitting their own work? Will AI give students wrong or biased information? Are we still preparing students with the right knowledge and skills?,” and “What privacy concerns should my school be aware of and how can they keep students safe?” Missouri took a similar approach. The state department of education sent me a presentation it gave teachers called “Right-sizing ChatGPT/AI in the Secondary Classroom,” which was created by a University of Missouri assistant teaching professor. The presentation is level-headed about AI and suggests “simple tweaks that you can make to your existing assignments to engage students in the process of writing, assessing and giving feedback along the way” to teach them how to actually write. 

American Schools Were Deeply Unprepared for ChatGPT, Public Records Show
A slide from Right-sizing ChatGPT/AI in the Secondary Classroom

The documents and emails I obtained also had a lot of administrators and teachers discussing how they felt about ChatGPT, and in some cases showed that state departments of education had not thought about ChatGPT even after it became popular.

In California, when the principal of a school asked the state if it had curricula or frameworks for how it should think about ChatGPT, they were told “unfortunately, the topic of ChatGPT has not come up in our circles” and “I am not aware of any plans to develop a framework in this area.” When a California public information officer sought details from one of his colleagues about what the state was doing about ChatGPT in schools, or how the state was thinking about it, his colleague said “I have never heard of ChatGPT prior to your email. I am not aware of discussions among educators and teachers about this program.” These specific emails were from January and February 2023, which was soon after ChatGPT launched, but even then, it was a phenomenon, and experts and teachers were warning about how it would change teaching. 

In Idaho, one of the state’s heads of curriculum told a colleague in an email that ChatGPT “is AMAZING and can be used in so many ways for teachers. Writing letters to parents, writing letters of recommendations for students, and giving unit and lesson ideas.” The colleague wrote back saying they were concerned because “it can be prone to factual errors or biases in its generated responses, and that “The people who'd need the most help with planning are the ones who probably wouldn't catch any problematic stuff the bot generates.” The head of curriculum wrote back “it's kind of like Wikipedia‐ some people think it is the worst thing to ever happen but really it is one source and actually a fairly decent one. The more it is used, the better it gets,” then explained that they tested having it make a quiz. “I bet when spell check came out teachers went crazy as well. But I know I for one would not want a world without it!” 

Another email shows that a Michigan department of education worker sent a TikTok about using ChatGPT in schools to a curriculum coordinator in Idaho, who responded “That is some serious application…I love it!” An administrator in Missouri told colleagues that ChatGPT is “quite incredible. When we google a question we typically get multiple sites that we have to sort through to find the answer. This website uses artificial intelligence software to sort through websites and articles and literally just gives you the answer to your question.”

In Minnesota, an education department worker told colleagues that “ChatGTP [sic] is creating some buzz in education circles,” then shared a screenshot of a ChatGPT conversation in which he asked it “Explain how AI may impact education.” “It’s here and free for anyone to use,” he said. In Mississippi, a teacher undergoing coaching was taught how to use ChatGPT to make quizzes and activity plans. 

Vermont appeared to be a bit more forward thinking, with a teacher proactively reaching out to the state department of education to say they wanted to put together tools for teachers about the risks of ChatGPT: “I was initially only thinking about doing this for my local district, but I believe there is a larger need. The vast majority of educators are ignorant about this technology … Readily accessible AI can already generate whole essays on nearly any topic, and so it is understandable that educators are concerned about the implications for VT students,” they wrote. “The best teaching strategies and practices already mitigate many of the concerns this tech raises. I’d like to help teachers understand what those are, and what it means to teach in a way that asks students to do more than simply generate formulaic content that a machine can spit out in seconds. At the same time, I’d like to assist teachers in thinking about ways to help students interact with content generators in healthy and responsible ways—as is our obligation as educators. The tech is here to stay; we can’t hide from it or ignore it. What we can do, however, is use it as an opportunity to remind ourselves what high-quality teaching looks like.”

In New Hampshire, a district superintendent told administrators at schools that their current ways of teaching were likely to get “overwhelmed” by AI: “ChatGPT is merely an example of how AI is now (currently, not soon) impacting content creation — including research and writing,” he wrote. “As we work on curriculum alignment among our schools, then move from there into instruction and assessment, we CAN NOT continue to embrace traditional notions of instruction and assessment. Not ‘shouldn’t’ — ‘can’t.’ There are some obvious strategies that our educators can employ; others will not be so obvious. This is important, and even if you feel like we are drinking from a firehouse already, we don’t have any choice but to engage on these challenging, disruptive ideas.”

Read the whole story
mrmarchant
3 hours ago
reply
Share this story
Delete

Nuclear reactors for dummies by Alex Chalmers
Thursday May 15th, 2025 at 10:01 AM

The Works in Progress Newsletter
1 Share

We’re experimenting with turning on comments on our Substack. Please leave your thoughts below! We’re also hosting a Stripe Press pop-up coffee shop and bookstore on Saturday, June 28, in Washington, DC. RSVP here if you can make it.

At the bottom of the periodic table, you’ll find the ‘heavy’ atoms. Their tiny nuclei are crammed with over a hundred protons and neutrons. Heavy atoms, like uranium-235 are naturally unstable. The electrical repulsion inside them is so great that they are perpetually on the edge of fragmenting – a single, slow-moving neutron is enough to make a nuclear rupture.

When the break comes, the ‘heavy atom’ will split into two mid-sized atoms, commonly pairs such as krypton and barium or strontium and xenon, along with a few neutrons. Each split also releases about 200 million electron-volts of energy; a gram of U-235 yields roughly the heat of a few tonnes of coal.

Image source: IAEA.

Meanwhile, the neutrons generated by each split can strike other heavy nuclei, triggering a carefully managed chain reaction. The steady torrent of heat this process generates is used to boil water, produce the steam that spins turbines, and generate electricity.

If you are interested in nuclear energy, you will hear throwaway references to different reactor types. To an outsider, it may not be obvious how the many types of water or gas reactor projects differ from each other or which experimental designs are likely to be close to fruition. This post outlines the main categories of reactors, their advantages, drawbacks, and the prospects for different concepts.

The different reactor types have some features in common. They need nuclear fuel containing fissile material, usually uranium that has been enriched to contain a higher proportion of fissile isotope uranium-235. Most reactors also have a moderator that slows the flying neutrons so they linger near the uranium nucleus longer, allowing it to absorb them more easily and keep the chain reaction going. Cooling systems then transfer heat from the reactor core to turn a turbine and generate electricity while preventing overheating (which would cause a meltdown of the core).

They key differences across reactors tend to concern:

  • Moderator: Different moderators slow neutrons with varying efficiency. All moderators absorb a proportion of the neutrons, but every absorbed neutron is a neutron that cannot fission a nucleus, making it harder to keep a chain reaction going. Moderators with lower absorption rates can use less enriched or entirely natural uranium.

  • Coolant: The coolant establishes the maximum operating temperatures and pressures. When the steam is at a higher temperature, a higher proportion of its energy is converted into mechanical work turning turbine blades, meaning it has a higher thermal efficiency. Some coolants have more failsafes, preventing meltdowns, or minimising their impact if they do happen.

  • Fuel composition: The form and enrichment level of fuel determines energy density, frequency of refueling, and the amount and type of waste that’s produced.

    Share

I’ve not focused on size in this piece, as many of the reactor types in this piece could theoretically be built either as large reactors or adapted to smaller and modular designs, often known as SMRs. Currently, the most common basis for SMRs is the pressurized water reactor, although essentially every experimental design is currently being approached in a small format for economic reasons.

Operating reactors

Pressurized water reactor

Image source: Energy Encyclopedia.

Moderator: water

Coolant: water

Operating temperature: 290-330 celsius

Thermal efficiency: 33 percent

Enrichment: 3-5 percent

The Pressurized Water Reactor (PWR) is the most widely used nuclear reactor design in the world. Roughly 70 percent of operational civilian nuclear reactors are PWRs. The vast majority of reactors in France, South Korea and China are PWRs, along with two thirds of the US fleet.

In a PWR, nuclear fission heats water in a primary coolant loop kept under high pressure to prevent boiling. This pressurized water transfers heat through steam generators to a separate secondary loop where water is allowed to boil and produce steam, which then drives turbines to generate electricity.

First developed in the US to power submarines, they were adapted and scaled up for civilian use as part of the 1950s Atoms for Peace program, with the first commercial PWR beginning operation at Shippingport, Pennsylvania in 1957.

Shippingport Atomic Power Station. Image source: Department of Energy.

The PWR’s dominance was driven by two main factors.

First, its military origins ensured that there was a readymade industrial base for civilian deployment. Companies like Westinghouse that designed and built reactor components for military purposes already had equipment, supply chains, and in-house expertise.

Secondly, the use of water came with inherent advantages. As the reactor gets hotter or if any coolant is lost, the nuclear reaction slows, acting as a passive safety measure. This is because water becomes less dense as it heats up, making it less effective at slowing down neutrons to the optimal speed needed for uranium fission. This is known as a negative temperature coefficient and is an example of ‘passive safety’, when natural physical processes build in a degree of safety without the need for human intervention.

The two-loop design also creates a barrier between the radioactive primary coolant and the turbine system, making maintenance easier and reducing the number of components that need radiation shielding.

Despite their dominance, PWRs do have notable disadvantages, which have inspired much of the subsequent innovation in reactor design. The pressure level of 150-160 bar (one bar is roughly equal to atmospheric pressure; 150 is roughly the crushing pressure that would be experienced under the weight of a 1,500 metre column of water) requires highly robust pressure vessels and piping, along with a rigorous inspection regime. The two-loop design also builds in a degree of complexity.

Water also has its drawbacks, most notably its lower steam temperature. This means that PWRs typically achieve lower thermal efficiency than most of the competition, with only about a third of the heat generated being converted into electricity. (This is less of a problem if there is some other use for the spare heat, for example a district heat network as at Beznau in Switzerland.)

Boiling water reactors

Image source: Energy Encyclopedia.

Moderator: water

Coolant: water

Operating temperature: 288 celsius

Thermal efficiency: 35 percent

Enrichment: 3-5 percent

The Boiling Water Reactor (BWR) is the second most common nuclear reactor worldwide, accounting for roughly a fifth of nuclear power generation. In a BWR, water circulates through the reactor core where it is heated directly by nuclear fission until it boils, creating steam that drives turbines to generate electricity. Unlike PWRs, BWRs employ a single-loop design where the same water serves as both coolant, moderator, and then boils directly inside the reactor core to generate steam.

Subscribe now

Developed in the 1950s by General Electric, BWRs emerged as an alternative to PWRs with the first commercial unit entering operation at Dresden, Illinois in 1960. Their development was driven by the pursuit of simplified systems and lower construction costs. Alongside the US, major BWR adopters include Japan, Sweden, and Germany.

Dresden Generating Station. Image source: Nuclear Engineering International/

BWRs offer several advantages compared to other reactor designs. By employing half the operating pressure of a PWR, the BWR does not need such thick pressure vessels and demands fewer pressure-resistant components. The single-loop design strips out some complexity, while retaining some of the passive safety benefits of water. The steam in a BWR is generated directly by boiling conditions in the reactor core, which removes some of the temperature and efficiency loss. This means that despite operating at a lower temperature than PWRs, thermal efficiency is marginally higher thanks to higher steam quality.

Despite these benefits, BWRs face certain disadvantages. The single-loop design means radioactive water comes into contact directly with the turbine system. This means that all the components require radiation shielding and makes maintenance more complicated. BWRs also need complicated control systems, because steam bubbles can displace the liquid water moderator, potentially causing the reactor’s power to oscillate and interfering with cooling. Additionally, when steam forms in the upper part of the core, it creates uneven heating patterns that operators must carefully manage to prevent any areas from becoming too hot and maintain stable operation.

Pressurized Heavy Water Reactor

Image source: Energy Encyclopedia.

Moderator: heavy water

Coolant: heavy water

Operating temperature: 290 celsius

Thermal efficiency: 30 percent

Enrichment: 0.7 percent

A Pressurized Heavy Water Reactor (PHWR) uses deuterium oxide (heavy water) as both moderator and coolant instead of ordinary water. This allows the use of natural uranium (uranium-238), because the deuterium atoms in heavy water absorb significantly fewer neutrons than the hydrogen atoms in ordinary water.

The concept of heavy water moderation has been extensively trialled since the 1940s. Canada, India, and Argentina have all opted for PHWRs, as they allow countries to build reactors without relying on a uranium enrichment programme. Deuterium only makes up 0.0156 percent of hydrogen atoms in standard water and is extracted by circulating hydrogen sulfide gas between hot and cold towers to concentrate the naturally occurring deuterium.

PHWRs use separate systems for coolant (hot heavy water under pressure) and moderator (cold heavy water at low pressure) to maximize neutron efficiency. This separation requires hundreds of horizontal pressure tubes running through a large tank (calandria), with each fuel channel needing complex sealing connections and individual cooling circuits. While technically ingenious, this is builds in significant cost and complexity.

This cost of manufacturing and maintaining these hundreds of individual pressure tubes means that PHWRs remain relatively niche, providing only five percent of the world’s nuclear energy.

Advanced gas-cooled reactors

Windscale AGR prototype. Image source: Nuclear Power magazine.

Moderator: graphite

Coolant: carbon dioxide

Operating temperature: 650 celsius

Thermal efficiency: 41 percent

Enrichment: 2.5-3.5 percent

Advanced gas-cooled reactors (AGRs) were the great hope of the British nuclear industry and served as the backbone of the UK’s nuclear fleet, building on the UK’s original gas-cooled Magnox reactors. While eight of the second-wave ‘advanced’ reactors remain in operation, the UK never succeeded in exporting the advanced design, and it exported only two of the original Magnox (though its plans were shared as part of the Atoms for Peace program, and the North Korean nuclear program was based on it).

They used carbon dioxide as a coolant and graphite as a moderator instead of water, which allowed the AGRs to operate at a higher temperature, providing a meaningful efficiency improvement over PWRs.

There was another advantage. In AGRs, the graphite moderator is a large, solid block in which the fuel channels are embedded. This made it theoretically possible to open individual fuel channels for refueling, while keeping the reactor running, as the graphite would stay in place and continue to slow neutrons effectively. By contrast, a water moderator involves the coolant cycling continuously, so its removal or disturbance during fuel handling would interfere with the moderation process, necessitating a shutdown for safe refueling.

Construction proved significantly more complex and expensive than expected and engineers later discovered that the graphite moderator was prone to cracking. It was also necessary to shut the reactor down for safe refueling.

While much maligned, the AGR was relatively cheap by the standards of British infrastructure mistakes. By the time they are all decommissioned in the 2030s, the UK will have gained fifty years of reliable electricity from these reactors, which collectively cost around £30 billion, significantly less than Sizewell C, and about as much as was spent on capping energy prices during recent energy price spikes.

Sodium-cooled fast reactor

Image source: Wikimedia.

Moderator: none

Coolant: liquid sodium

Operating temperature: 550 celsius

Thermal efficiency: 41 percent

Enrichment: 15-30 percent

A Sodium-Cooled Fast Reactor (SFR) uses liquid sodium metal as its primary coolant instead of water. This allows the reactor to operate at much higher temperatures compared to traditional water-cooled reactors. The sodium coolant transfers heat efficiently and remains liquid without requiring high pressure. These higher temperatures significantly improve thermal efficiency, while the low pressure operation making maintenance simpler.

SFRs are among the only ‘fast breeder’ reactors to have ever been built. These reactors have the potential to produce more fissile material than they consume, theoretically making them self-sustaining after their initial fueling.

Subscribe now

However, sodium's violent chemical reactivity with water and air presents serious safety challenges. The liquid metal ignites spontaneously in air and reacts explosively with water, requiring sophisticated containment systems, inert gas blankets, and multiple heat exchange barriers to prevent potential sodium fires.

The concept of sodium cooling has been extensively tested since the 1950s. Several countries have operated experimental and demonstration SFRs, including long-running plants like France's Phénix, Russia's BN-series, and Japan's Monju. A number of commercial projects are under construction, including TerraPower’s Natrium reactor in Wyoming.

Proposed or experimental reactor types

While the range in operational reactor design is quite limited, start-ups and government-funded R&D projects have been exploring a range of potential innovations. These are largely focused on finding alternative coolants to water so the reactor can operate with greater thermal efficiency.

There are few ‘original’ ideas in nuclear, so these concepts often rely on ideas that were explored during the 1960s or 1970s, but abandoned on the grounds of cost or complexity. Innovators are hoping that advances in engineering, technology, and material sciences can help to overcome some of these hurdles.

It’s important to note that the thermal efficiency numbers for these advanced concepts are R&D targets or design estimates – no reactor has ever demonstrated efficiency significantly above 40% at utility-scale.

Molten salt reactors

Image source: US Department of Energy.

Moderator: molten salt

Coolant: molten salt

Operating temperature: 600-800 celsius

Thermal efficiency: 40-45 percent

Enrichment: 5-19 percent

A molten salt reactor uses a mixture of molten salts as its primary coolant and, in many designs, as the medium in which the nuclear fuel is dissolved. The molten salt stays liquid at high temperatures without needing high pressure, allowing efficient heat transfer without complex high-pressure systems. This allows the reactor to operate at much higher temperatures, typically around 700 degrees centigrade compared to traditional water-cooled reactors.

These higher temperatures improve efficiency, while the lack of pressurization reduces the likelihood of an explosion if the reactor malfunctions. Molten salts’ chemical properties can act as a passive safety feature. The molten salt keeps most fission products dissolved in the liquid and, if a leak occurred, would drain into a catch tank and quickly freeze, trapping the radioactivity. Because the salt boils only at around 1,400 celsius and carries heat well, the core can ride out a loss of active cooling for long periods without overheating.

The idea of using molten salt isn’t new. Oak Ridge National Laboratory deployed an experimental reactor between 1965 and 1969, but there was little incentive at the time to invest billions of dollars in commercializing them, given the success of light water reactors.

Control room of Oak Ridge National Laboratory’s molten salt reactor Image source: Oak Ridge National Laboratory..

Governments are giving the technology a second look now and a number of start-ups are tackling the challenge. Google has struck an agreement to buy energy from Kairos Power, which is working on small molten salt reactors.

High-Temperature Gas-Cooled Reactor

Moderator: Graphite

Coolant: helium

Operating temperature: 750-950 celsius

Thermal efficiency: 48-51 percent

Enrichment: 17 percent

The High-Temperature Gas-Cooled Reactor (HTGR) uses helium gas as coolant and graphite as moderator. Helium is chemically inert and remains gaseous at high temperatures, which allows for high heat transfer without the risk of dangerous chemical reactions, while the graphite can absorb significant amounts of heat.

HTGRs use specialized fuel consisting of uranium particles coated with multiple layers of ceramic materials, which can withstand extreme temperatures while containing radioactive materials (even in case of a meltdown). There are two main design variants: prismatic block designs where hexagonal graphite blocks contain channels for fuel and coolant, and pebble-bed designs where tennis ball-sized fuel spheres move through the reactor core.

These high temperatures, however, require specialized materials and the specialised fuel particles require intricate manufacturing processes.

The US experimented with HTGRs, operating the Fort St. Vrain plant from 1979 to 1989. While the reactor itself worked, consistent faults in other parts of the plant’s setup led to its early decommissioning on economic grounds. There have also been experimental reactors in Germany and the UK. There is currently only one HTGR in operation globally, the pebble-bed Shidao Bay Nuclear Power Plant in Shandong province. Meanwhile, US start-up Valar Atomics is planning a test reactor in the Philippines.

Cooling tower of the THTR-300 in Germany, demolished in 1991 Image source: Wikimedia.

Lead-cooled fast reactors

Moderator: none

Coolant: liquid lead

Operating temperature: 480-580 celsius

Thermal efficiency: 42 percent

Enrichment: 16 percent

A Lead-Cooled Fast Reactor (LFR) uses liquid lead or lead-bismuth blend as its primary coolant instead of water or gas. This allows the reactor to operate at much higher temperatures, typically around 480-580 degrees celsius compared to traditional water-cooled reactors. The lead coolant remains liquid at these high temperatures without requiring high pressure, enabling efficient heat transfer while eliminating the need for complex high-pressure containment systems.

Lead also has some intrinsic advantages. It doesn't react violently with water or air, doesn't burn, and it effectively shields radiation. The high boiling point of lead (1,749 celsius) creates a substantial safety margin, as it's nearly impossible to boil the coolant during operation.

The concept of using liquid metal as a reactor coolant has been explored since the early days of nuclear power, with the Soviet Union pioneering lead-bismuth cooled reactors for their submarine fleet. However, this design comes with several challenges. The high density of lead creates substantial structural loading on reactor components. The coolant's opacity makes visual inspection of submerged components impossible, complicating maintenance. Additionally, lead's high melting point requires sophisticated heating systems to prevent freezing during shutdowns, which could damage reactor structures.

There are no LFRs currently in operation, but a few demonstrator reactors have been planned. For example, Belgium, Italy and Romania have formed a consortium with Westinghouse, although the project is in its early stages.

Supercritical Water Reactor

Moderator: Supercritical water

Coolant: Supercritical water

Operating temperature: 500-625 celsius

Thermal efficiency: 44-48 percent

Enrichment: 5-7 percent

Image source: US Department of Energy.

A Supercritical Water Reactor (SCWR) uses water heated and pressurized beyond its critical point (374 celsius and 22.1 megapascals) as both coolant and moderator. In this supercritical state, water has properties of both liquid and gas, with exceptional heat transfer capabilities. This allows the reactor to operate at significantly higher temperatures than many of its competitors.

These higher temperatures substantially improve thermal efficiency, but the extreme operating conditions create formidable materials challenges. Supercritical water's extreme temperature, pressure, and radiation exposure creates a powerfully corrosive environment that rapidly degrades conventional materials, necessitating specialized alloys for reactor components.

Several countries have research programs exploring SCWR designs, but the significant costs mean that technology remains at the conceptual and small-scale experimental stage.

Alex Chalmers is an editor at Works in Progress, focused on AI and energy. He’s also the author of Chalmermagne, a Substack covering technology, policy, and finance.

Thanks for reading The Works in Progress Newsletter! Subscribe for free to receive new posts.

Read the whole story
mrmarchant
6 hours ago
reply
Share this story
Delete

ChatGPT Diminishes Idea Diversity in Brainstorming, Study Finds by msmash
Thursday May 15th, 2025 at 9:59 AM

Slashdot
1 Share
A new study published in Nature Human Behaviour reveals that ChatGPT diminishes the diversity of ideas generated during brainstorming sessions. Researchers from the University of Pennsylvania's Wharton School found [PDF] that while generative AI tools may enhance individual creativity, they simultaneously reduce the collective diversity of novel content. The investigation responds to previous research that examined ChatGPT's impact on creativity. Their findings align with separate research published in Science Advances suggesting AI-generated content tends toward homogeneity. This phenomenon mirrors what researchers call the "fourth grade slump in creativity," referencing earlier studies on how structured approaches can limit innovative thinking.

Read more of this story at Slashdot.

Read the whole story
mrmarchant
6 hours ago
reply
Share this story
Delete

To the Dad Coaching Third Base, Wearing a “Women Want Me, Fish Fear Me” Hat by Tyler Gooch
Thursday May 15th, 2025 at 9:59 AM

McSweeney’s
1 Share

One hat, two independent claims, three bases, two outs, and zero energy from you, the third-base coach wearing a hat that alleges, WOMEN WANT ME, FISH FEAR ME.

I have no standing to doubt that women want you. I’m willing to take you at your word that there are women out there who want a lethargic man who just stands there as a runner that he should’ve told to slide is casually tagged out at third. My issue is with your hat’s second claim.

I do not believe fish fear you.

I believe fish are unaware of your existence. And I think it would be impossible for any animal, land- or sea-faring, to fear an entity it has never known. For instance, you or I could reasonably fear aliens as a concept, but we do not fear one specific alien we’ve never met, like Zorpthop the Destroyer (though we should).

I also believe fish would never fear a man with the cowardice to stop a runner when he had a clear path to home. No creature would watch the outfielder bobble the ball like that, see you fail to send the runner home, and quiver at your existence.

Am I to assume your hat means you’re a skilled fisherman, and fish fear you because your fishing prowess means you represent certain death for the fish? If that is what the hat is getting at, you should know that is impossible. In order for you to be seen as an aquatic angel of death, you must kill all the fish you catch. And if you kill all the fish you catch, then word of you would never spread throughout the fish community. Your legend would never grow underwater.

You can take the hat off, but it’s been three innings. I already saw the hat, and I’m not going to let it go.

The only way for you to develop a reputation underwater as the fish-slaughtering third-base coach would be if you threw the fish back after catching them. However, if you throw the fish back, the fish would not fear you. They would be annoyed by you (as I am). They would see you as some sort of mysterious deity that inconveniences them and potentially causes them to be late to their fish child’s baseball game. But they would not fear you.

Maybe you catch five fish, kill the first four, then show the fifth one all of its dead compatriots before returning it to the water so that it can spread word of the horrors it saw at your hand. But in that scenario, I still don’t think the fish would fear you. The fish would recognize that they have, in their midst, a fish who was spared by you. A fish that somehow defeated the mystifying harbinger of evil on the fishing boat. Whatever fear the fish may have for you would be outweighed by the respect they have for their maritime messiah.

No, I won’t shut up and let you focus on coaching.

If you knew anything about baseball, you would signal a bunt right now, you coward. The five-year-old T-ballers on the other team don’t even fear you.

There is only one plausible scenario I can dream up in which both women want you and fish fear you. Perhaps you work at a pet shop or a dentist’s office, some place with a fish tank. And maybe you also enjoy attending either horror conventions or Slipknot concerts. Occasionally, you have to get changed into your Slipknot/horror costume at work so you can drive straight to the event after work. There are women at these concerts/conventions who love the lengths you go to to dress up, and thus, they want you. And there are fish in the office fish tanks you pass by on your way out who look out and are frightened at the sight of your Slipknot / Freddy Krueger mask, as they are unfamiliar with the discography of Slipknot or the Freddy Krueger franchise.

Oh, okay, Ump, his absurd hat is fine, but you’re going to kick me and my kid out because of my yelling and my hat that says I CAN’T GO INTO DETAILS BUT I INSPIRED AN EPISODE OF LAW AND ORDER SVU. That’s fine, Bozo. My kid’s game doesn’t even start for another forty-five minutes.

Read the whole story
mrmarchant
6 hours ago
reply
Share this story
Delete
Next Page of Stories