Future

The Smartest People in AI Believe This Is the World’s Most Important ChartBen Cohen | The Wall Street Journal ($)

“Anthropic refers to [recursive self-improvement] as ‘closing the loop,’ a cycle in which each improvement makes the machines better and faster at improving themselves. AI companies say they are now seeing the early signs of recursive self-improvement, which is speeding up their research—and freaking everyone out.”

The post This Week’s Awesome Tech Stories From Around the Web (Through November 7) appeared first on SingularityHub.

Five scientific experts discuss whether artificial intelligence is—or might one day become—truly intelligent.

This article was originally published on Undark. Read the original article.

Is it possible that AI is—or might one day become—truly intelligent? We put this question to five experts working in a variety of fields, including linguistics, computer science, cognitive science, and psychology. Not surprisingly, they offered very different answers, though if there was one point of agreement, it was that the term “intelligence” is vague and poorly suited for the conversation.

“Intelligence is a heavily contested concept,” Benjamin K. Bergen, a professor of cognitive science at the University of California, San Diego, wrote in an email to Undark. “Psychologists disagree on what it is.” Nevertheless, these five experts helped to parse just what our machines are doing—and what, if anything, separates AI from humans and from the natural world.

Some of these entries are drawn from email exchanges while others came from Zoom interviews. In some cases, they have been edited for length and clarity.

Emily M. Bender
Professor of linguistics, University of Washington

Setting aside the undefined terms “AI” (a marketing term, not a coherent set of technologies) and “intelligent” (a eugenicist way to rank people), can we ever create artificial systems that engage in cognitive behaviors or experiences similar to people? Probably not, and definitely not through language modeling (like LLMs). The only evidence we have that LLMs can “think” comes from our ability to make sense of the synthetic text they extrude.

Linguistics—the study of how languages work and how we work with language—provides two crucial insights here: The first is that languages are systems of signs, which pair form (the spelling or pronunciation of a word) with meaning (what the word means in general and what it is being used to mean). Language models manipulate form, but without reference to meaning.

The second lesson concerns how we interpret language. Far from simply unpacking meaning that an author packed into the words, we instead keep in mind everything we believe about the author’s state of mind, our common ground with them, and their beliefs about their intended audience’s state of mind. Against that background, we ask ourselves: What must the author be trying to convey by using those words and in that way?

In order to interpret text, we have to imagine a mind behind the text. We do this instinctively and reflexively—we can’t help ourselves, making rigorous evaluation of these systems difficult. As we wrote in the Stochastic Parrots paper in 2021: “Our perception of natural language text, regardless of how it was generated, is mediated by our own linguistic competence and our predisposition to interpret communicative acts as conveying coherent meaning and intent, whether or not they do.”

If we don’t account for this, we will not be able to make wise decisions about this technology.

Melanie Mitchell
Computer scientist and professor, the Santa Fe Institute

There’s this notion in AI circles that there’s this easily defined split between what people will call “cognitive intelligence” versus “physical intelligence.” When these big companies, for instance, like DeepMind or OpenAI, define what they’re calling “artificial general intelligence,” they talk only about cognitive tasks, as opposed to physical tasks. So, ChatGPT is not going to come fix your roof.

But it’s not clear to me how you can so easily separate the so-called cognitive kinds of tasks from the more physical kinds of intelligence that our brains were actually evolved to enable us to have. Or the more social kinds of intelligence that we humans rely on so much.

Undark: Can you say more about why it’s hard to draw a distinction between cognitive and physical intelligence? If I ask AI, “how do I fix my dishwasher?” it will give me instructions. Then I go and do the physical act of fixing the dishwasher. That seems like a clean distinction.

But in interpreting those instructions and applying them to your dishwasher, you’re adding quite a bit of the intelligence in that. If it says, “Press the red button on the left side of the door,” you—human—are using a lot of your intelligence to figure out what that means. Maybe there’s two red buttons, but one of them is obviously the one you’re supposed to press.

Undark: My experience with AI is that it’s quite bad at social intelligence.

One notion of social intelligence is that we can coordinate with other people to get things done: So, I don’t know how to fix my car, but I know how to find someone to fix my car. I might ask my friends, “Who do you know who knows how to fix a car?” And then also figuring out, what’s the right way to explain the problem to another person? I think of that as social intelligence.

I don’t know why AI systems are so bad at that. It may be that they don’t have the right training. It may be that they don’t have the right kind of participation in the social world.

Anima Anandkumar
Professor of computing and mathematical sciences at Caltech

Let me give my definition of intelligence: the ability to learn from data, and the ability to adapt to surroundings, or environment, based on that. You have now many AI systems that can do both.

You have AI that interacts with humans as chatbots and then adapts through that feedback—what we call reinforcement learning. These models can be made to align to certain goals. Could be doing more accurate mathematical proofs. Where it’s been most popular has been language models because we interact with those daily.

Language models mirror a lot of human intelligence, but I want to talk about nature’s intelligence. The ability to learn from microscopic scales—how atoms interact, how reactions happen, that enables us to create better drugs, better materials—to planetary scales and even cosmological scales: understanding how weather systems form. Is there a hurricane forming? How quickly will it make landfall? Can we understand the fine details of it to really get the predictions correct and do that early on? Because that has huge impact on saving human lives, creating evacuation policies early on.

We created the first AI-based weather model about five years ago that’s tens of thousands of times faster than what traditional forecasting could do, and now exceeds even the accuracy of traditional forecasts. So it’s faster, it’s more accurate, and we can create better predictions of extreme weather events and also assess risk. And now with changing climate, you could also make the systems predict longer-term climate, be able to really accurately incorporate the laws of physics, so that they can be accurate, even in newer conditions.

People focus so much on human intelligence. I want to talk about the intelligence that’s all around us—that’s invisible to us—but really, all of life is built upon it.

Alison Gopnik
Professor of psychology at UC Berkeley, author of popular books about babies and cognitive development

This question, and the phrase “artificial general intelligence” assume that there is a single thing called “intelligence” and that creatures, artificial or natural, have more or less of it.  This is a common kind of explanation in folk theories and in early scientific ones. But as science progresses, explanations become more complex and cognitive science describes many varied, complex, cognitive capacities that serve different, and often conflicting, functions—many different “intelligences.” The intelligences that allow human children to learn so much are still far beyond current models.

The remarkable success of the current large models depends on the ability to detect statistical patterns in the text and pictures that humans have already produced at scale and produce new text and pictures based on those patterns. This allows human users to take advantage of the information previous humans have produced—these models are a technology for cultural transmission. But this is fundamentally different from the capacities that allow people to produce that information in the first place.

Another kind of “intelligence” is the “exploit” ability to take actions to achieve particular goals. This kind of intelligence is captured by reinforcement learning. AI systems use reinforcement learning to master skills like Atari games or coding, and it plays a crucial role in “post-training” the large models.

Cultural transmission and reinforcement learning play a role in human intelligence. But very young children actively explore a changing world and build new models of it—“explore” capacities that are at odds with “exploit” intelligence. They also rely on a very different kind of intelligence—the intelligence of care. Children can explore because other people take on the cognitively challenging task of caring for them. These capacities go beyond current large models, although different types of artificial intelligences might begin to approach them in the future.

Benjamin K. Bergen
Professor of cognitive science at UC San Diego

For most definitions of intelligence, modern AIs already display behaviors consistent with at least some degree of intelligence. They react to their environment, learn from data, and generalize beyond their training materials. By some measures, they already appear to reason logically and display sensitivity to the context and feelings of the humans they’re interacting with.

But whether these behaviors make them “truly intelligent” … depends, in part, on what you think behavioral evidence can actually tell you. Let’s say you’re trying to decide whether a horse is truly intelligent. When you pose arithmetic problems, the horse stomps its hoof the correct number of times to indicate the answer. Is the horse intelligent? Well, if the horse is able to understand problems articulated in human language, perform mathematical operations, and respond appropriately, you might reasonably conclude that it has some degree of intelligence.

But if you learn instead that the horse is actually just taking subtle cues from its trainer, who knowingly or not makes subtle movements when the horse has reached the correct number of stomps, this horse is displaying a different kind of intelligence. It has learned to stomp in a particular kind of situation, to attend to the trainer for subtle cues, and to stop appropriately. This is an example of learning from experience, generalizing to new situations, and perhaps sensitivity to the emotions of others in a social context. 

But the lesson from this case is that intelligent-seeming behavior can be produced by many different kinds of underlying mechanisms, which may be different in the kinds or degree of intelligence that they engage. Similarly, AIs can answer questions that seem to require profound reasoning but which upon further inspection are actually spelled out in their entirety in the voluminous training data the model has consumed.

The post What Does the ‘I’ in AI Really Mean? appeared first on SingularityHub.

Rather than relying on text alone, the algorithm looks for clues in a paper’s reasoning.

Science is no longer a purely human endeavor. Machine minds are increasingly merging with human creativity to generate hypotheses, experiments, and ideas.

But with promise comes misuse. Like TikTok videos and AI-generated images splattered across the internet, AI slop has infiltrated science. Academic journals and conferences are being flooded with AI-generated content, and editors are drawing the line. As long as they disclose it, scientists can use chatbots to help write papers, as long as the underlying ideas are their own. Submitting a whole paper dreamed up by AI, with little human input, is another matter.

Publishers are turning to AI detectors to weed out slop. But detectors aren’t perfect. Most scrutinize turns of phrase, word choice, and other intricacies associated with machine-generated text. They can mislabel legitimate research because of how the paper is written, rather than on the merit or origin of its ideas. That’s especially problematic for scientists whose first language isn’t English and who rely on AI for heavy editing.

A team at the University of Maryland in College Park is now taking a different approach: Strip away the prose until only a skeleton of concepts remains. Their detector, called IdeaLens, learned to find patterns in how human and AI models organize ideas, build narratives, and construct arguments, all of which are integral to every scientific paper.

IdeaLens alone won’t stem the deluge of AI slop polluting scientific literature. But it joins a growing toolset that recognizes human contributions, even as the boundary between scientist- and AI-generated ideas grows blurry.

“Discerning the extent to which AI performs critical thinking—a task that, in our view, rightly belongs to human authors—is increasingly difficult,” wrote editors of The Journal of Clinical Psychiatry, who were not involved in the new work. But, “using AI to deliberately fabricate data or automate the interpretation of study findings crosses ethical lines and undermines scientific integrity.”

Fighting Fire With Fire

There’s no doubt that AI has already made waves in science.

Algorithms are tackling particle and space physics and deciphering biology. Drug companies are betting AI can accelerate drug discovery and fight cancer, Alzheimer’s, and other medical scourges that continue to haunt millions worldwide.

Amid controversy, OpenAI recently announced an AI-powered breakthrough in fluid dynamics—one of the longest-standing, toughest puzzles in mathematics. While welcomed by some, the advance, alongside a barrage of other results, has triggered an “existential crisis” within the field and sparked debate over who deserves credit.

Generative AI has boosted research productivity over 89 percent, according to a study in late 2025. But there are growing pains. One estimate suggests roughly 12.5 percent of biomedical articles show signs of substantial AI involvement. And even AI conferences have struggled with AI submissions: Some papers accepted to the prestigious 2025 NeurIPs conference were flagged as likely AI-generated and 100 or more hallucinated citations slipped past human reviewers.

Publishers are pushing back.

The popular preprint server arXiv has tightened submission rules for computer science reviews, requiring prior peer review to combat the flood of submissions partially fueled by AI.

Others are using detectors. AI writing has particular quirks in word choice and sentence structure. In a well-known example, models tend to overuse constructions like, “It’s not X, it’s Y.” One popular tool, Pangram, claims 99 percent detection accuracy and is now integrated into alphaXiv, an arXiv mirror site, to automatically flag potential machine-written manuscripts. Meanwhile, scientists are turning papers into AI agents, making it easier to reproduce findings and spot questionable research.

Instead of focusing on wording, IdeaLens asks if the underlying concepts are human or AI.

A New Detective

The tool’s creators previously found AI-written fiction differs structurally from human work. Stories generated by machines tend to be tidier, more predictable, and more on the nose about their message. Humans write sloppier but more colorful tales. Beyond surface-level word choice or tone, the team found differences embedded into the narrative itself. For example, human writers often wander into subplots, let readers infer themes, explore morally debatable choices, and leave ambiguous endings. These detours create rich, lived-in worlds that reflect the messiness of real life. AI prefers straightforward narratives and neatly tied loose ends.

These findings inspired the team to develop IdeaLens, which looks past prose and focuses on structural and conceptual patterns.

The researchers first used Pangram to label roughly a million texts, such as news articles, blog posts, and academic papers, as human- or AI-generated. A large language model then stripped away the original wording, leaving a detailed outline of the main points. Scientific papers, for instance, were distilled to background information, claims, evidence, and arguments. Another model, called a classifier, then learned to find patterns that distinguished human- from AI-generated texts in these skeletal rewrites.

Like many modern AI systems, IdeaLens is a “black box.” The team can’t fully explain how it lands on verdicts, although they did find human-derived outlines tend to contain more study-specific details, rather than generic points often used by AI that could apply to other studies.

Still, it worked.

When challenged with documents written by ChatGPT and other large language models, IdeaLens mistakenly flagged them as human work just one percent of the time. For texts lightly edited by AI, it correctly recognized human origins in about 95 percent of cases. Across evaluation datasets, it missed roughly 10 percent of AI-generated work.

That might seem like a big problem. But “we would much rather have that AI document slip through than humans wrongly being flagged as AI,” study author Rishanth Rajendhran told Nature. “Because of the reputational damage that these things carry, it is important that we are careful with what we do with the prediction.” Once flagged, human reviewers should jump in for a closer look at the questionable submission.

Like other detectors, IdeaLens has blind spots. The team didn’t individually verify the origins of every document in its training data. The tool was also never tested on genuinely hybrid articles, where humans and AI models co-develop the underlying concepts, which is an increasingly common scenario in science. As scientists turn to chatbots to refine hypotheses or spark new ones, intellectual ownership becomes even murkier.

For now, IdeaLens sees the world in black and white, assigning just two labels: human or AI. Future versions may introduce more nuance, such as assigning the estimated percentages of their contributions. You can try it here.

The team stresses the detector shouldn’t discourage scientists from embracing AI. Rather, the differences it found highlight the value of human ingenuity. Without a dose of human messiness and out-the-box thinking, AI slop risks reshaping not just scientific literature, but our understanding of the world too.

The post Science Is Littered With AI Slop. A New Detector Could Filter It Out. appeared first on SingularityHub.

The system, MAI-DxO, was not only more accurate, but it also spent much less on tests to make its diagnoses.

Doctors are like detectives. They scrutinize symptoms, piece together medical histories, and narrow down possible diagnoses. Simple cases may be solved on the spot. But trickier ones require tests to gradually zero in on the culprit.

The stakes are high. A misdiagnosis can delay treatment, allowing diseases to fester and cause lasting harm or prolonged suffering. Costs also stack up. Even with health insurance, multiple tests, some unnecessary in hindsight, can leave patients with staggering bills.

Medical debt is already a major burden in the US. According to one estimate, more than one-third of households carried some type of it in 2024. Some patients even opted out of subsequent care because of costs, which are often opaque.

A faster, more efficient path to the right diagnosis could vastly improve medical care. To that end, AI could be a useful tool. This month, in a paper published in Science, Microsoft researchers described MAI-DxO, a sophisticated system that acts like a crack medical team.

When challenged with extraordinarily complex cases, MAI-DxO was more than twice as accurate as physicians armed with textbooks and the internet—and nearly four times as accurate as those relying solely on their own knowledge. The system also surpassed doctors on typical emergency room cases, achieving 80 percent accuracy versus 47 percent for physicians.

“I’ve been working on principles and applications of AI in medicine since the 1980s,” study author Eric Horvitz told Science. “My hope is that these advances will help bring timely and effective care to more people, and that’s what makes this progress so exciting to me.”

The ER Pitt

AI already outperforms doctors at spotting breast cancer on mammograms and detecting skin cancer in curated datasets. Systems can also detect the telltale signs of cancer spread and heart disease. But while modern models routinely ace standard medical licensing exams and tackle hypothetical clinical scenarios, these tests gauge a physician’s knowledge, not the messy realities of diagnosis. All clues are given up front, and the AI simply picks an answer.

In practice, diagnostics is a winding journey. Doctors continuously gather symptoms, order tests, and assess whether new results support their initial hunch or torpedo it, requiring them to start anew.

Decades ago, physicians proposed using benchmarks to judge how AI doctors navigate real-world cases. One of these is the New England Journal of Medicine (NEJM) clinicopathological series, which is often used to challenge novice doctors with complex medical mysteries. And earlier this year, a team at Harvard Medical School and colleagues developed a benchmark to compare the performance of AI to physicians. Pitting OpenAI’s o1-preview reasoning model against two expert physicians on emergency room cases, the AI performed better across clinical tasks, from triage and chart reviews to deciding whether to admit or discharge.

High Standards

The Microsoft team took benchmarking a step further. In the new study, they selected 304 complex cases from NEJM and a similar number of typical emergency room encounters from a database. Together, “they span a broad spectrum of clinical presentations and documentation styles, with final diagnoses ranging from common conditions to rare disorders,” wrote the authors.

The team transformed each case into a standardized, step-by-step interactive challenge for both AI and physicians. As you might expect in a clinical setting, the cases started with limited information—say, a 30-year-old man presenting with chest pain. Participants then had free rein to ask about medical history, order tests, and update hunches before committing to a diagnosis.

The benchmark measured two main outcomes, said study author Harsha Nori. “Did you get the right diagnosis? And how expensive was the path you took to get there?”

The team then built MAI-DxO. Powered by GPT-5, the AI acts like a physician panel. One specialist proposes possible diagnoses. Another suggests questions and tests and iteratively challenges the emerging diagnosis. The last considers costs. Meanwhile, an “orchestrator”—think head physician—coordinates the debate, weighs new clues from test results, and decides whether another test, and which one, is worth the cost.

Pitted against 40 doctors, each with roughly a dozen years of clinical experience, MAI-DxO excelled. On complex NEJM cases, it reached an accurate diagnosis 83 percent of the time. In comparison, physicians were 22 percent accurate when relying on their own knowledge. Giving them access to the internet and other resources only boosted their performance by a smidge. For typical emergency room cases, the AI was nearly twice as accurate as physicians.

MAI-DxO was proficient even under tight budgets. When diagnosing NEJM cases, a cost-conscious configuration maintained nearly 80 percent accuracy while spending an estimated $1,300 dollars per case, based on 2023 hospital pricing data from a national healthcare system. Doctors, in contrast, racked up a bill of over $4,700.

Although GPT-5 supplies the brainpower, MAI-DxO performed better than when the team used the model alone. In a tricky case of scurvy, caused by low vitamin C levels, the system connected the patient’s dietary history to bleeding gums and tooth loss. After ordering a targeted vitamin C blood test, it reached the correct diagnosis after spending just over $1,100. The standalone version of GPT-5, meanwhile, asked for an extensive neurological workup, resulting in a bill that was nearly 7.6 times more expensive.

MAI-DxO also outmatched a range of general language models, including Anthropic’s Claude, Google’s Gemini, and xAI’s Grok. “If you run it in this configuration, the AI systems get to the right diagnosis far more often and at far less cost than out-of-the-box runs,” said Nori. And the system isn’t tied to GPT-5. Swapping in the newer GPT-5.5 further boosted accuracy.

These findings, while promising, don’t mean we’ll be seeing unsupervised AI diagnosticians in hospitals anytime soon. The experiments took place in controlled, text-based simulations. Real-world clinicians navigate inaccurate and incomplete patient histories and analyze medical images and other clinical data—often under time pressure. Chronic or evolving diseases, such as cancer, could also stump the system.

Then there’s the human factor. Patients have the final say over whether to undergo a test. A diagnostic plan may look perfect on paper, but if someone refuses a procedure, the AI must be able to pivot. “Critically, we have to consider each patient’s understandings and their own goals and preferences,” said Horvitz.

The team has partnered with Beth Israel Deaconess Medical Center in Boston to evaluate MAI-DxO in different hospital settings. One of their goals is to figure out how best to fit the system into everyday clinical practice. Another is to see how patients directly interact with it and how its performance stacks up against assessments from doctors.

“Working alongside clinical teams will show where the technology succeeds, where it falls short and how it can fit into busy workflows,” the team said in a blog post.

If all goes well, unexpected emergency room visits might come with a little less sticker shock.

The post Microsoft’s Medical AI Is Over Twice as Good Diagnosing Diseases as Doctors appeared first on SingularityHub.

New AI tools could help more biologists experiment with problems once restricted to a handful of specialists.

Designing new proteins from scratch usually requires in-depth knowledge of their structures and chemistry, as well as familiarity with highly specialized software. But researchers have now developed an AI tool where users simply describe what they want a protein to do in plain language.

Since AlphaFold cracked protein folding in 2020, AI has revolutionized the field of structural biology. Since then, a wave of generative models have begun to work in the opposite direction, learning from vast collections of amino acid sequences, protein structures, and labels detailing their function to design proteins that do not exist in nature.

These tools remain difficult to use, however, as scientists have to guide them with formal technical specifications like structural constraints or instructions to mutate particular sites. That means trying to design something like a new plastic-degrading enzyme or a fluorescent imaging marker takes significant specialist expertise.

But in new research published inNature, scientists at Westlake University in China detail an AI system called Pinal that can generates protein sequences from natural-language prompts describing a desired protein function. And while the model outputs a lot of duds, lab tests showed that with enough tries, the system could reliably achieve a user’s goal.

“For decades, the field has largely required specialized expertise to translate desired functions into low-level structural or sequence constraints,” the authors write. “The importance of Pinal lies not only in its performance but also in reframing protein design as a problem that can be specified through functional language.”

Pinal works in two stages, neither of which asks the user for details about the molecular structure or sequence of amino acids. The first module takes the plain-language description of what the user wants and turns it into a structural blueprint. A second module then reads both that blueprint and the original text to decide the sequence of amino acids.

Crucially, the software doesn’t put all its eggs in one basket. The first module generates several candidate structures, all of which are fleshed out with their own amino acid sequences. Pinal then scores each of these combinations before settling on a final design. The researchers say this accounts for the fact that the best protein designs don’t always have the most obvious shapes.

Training the model took a vast amount of data. The researchers assembled roughly 135 million protein sequences and used large language models to write plain-language descriptions of their functions, generating about 1.7 billion protein-text pairs. They then used AlphaFold to predict the structures of the sequences.

The team used this corpus to train the two modules separately. The first learned to predict a structural blueprint from a text prompt, and the second learned to pick the right amino acid for each position in the sequence, given both the blueprint and the prompt.

To test the model, the team asked it to design four proteins: one that glows green, an enzyme that breaks down PET plastic, an enzyme that processes alcohols, and a shuttle protein that ferries chemicals between enzymes in a metabolic pathway.

The model was able to solve all four challenges, but its hit rate was a mixed bag.

When asked to produce a glowing protein, it created 50,000 sequences. These were filtered down to 29 for testing, and only one actually fluoresced. The system was better at producing the shuttle protein. Five of the 20 candidates worked, and the best one boosted a multi-enzyme carbon dioxide fixation pathway to 1.7 times the output of the natural version.

A different collaboration between Hugging Face and academics in China and Italy recently applied similar ideas to DNA sequences. The group released a 14-billion-parameter open model called Carbon-X that can answer questions about a gene sequence, predict what a mutation will do, or generate new DNA.

That model is based on Qwen3-14B, but was retrained on 300 billion tokens that mixed genomic sequences, biology text, and ordinary English. Getting it to learn DNA without forgetting English required a delicate balance though, the researchers write in a post, and forced them to do a long slow training run with a large model.

Across five benchmark tests, Carbon-X beat Opus 5.5 and Gemini 4 Argon and trailed GPT-6.1 Sol by less than a percent. And the model can run on a MacBook, which could be a major boon for researchers eager to avoid sharing sensitive or unpublished sequences with a commercial model provider.

Neither tool is perfect though. Pinal’s own developers concede that their screening heuristics “may restrict exploration to a limited region of the design space,” while Carbon-X struggles with long sequences and remains prone to hallucinations. Validating the sequences they produce also still requires a well-equipped lab and experienced scientists.

But these tools could nonetheless empower a far wider range of biologists to experiment with problems previously restricted to a handful of computational specialists.

The post Scientists Can Now Ask AI to Design Proteins Using Plain Language appeared first on SingularityHub.

Artificial Intelligence

Mistral Says Its AI Model ‘Le Chonk’ Has Leapfrogged China in the Open-Weight RaceJoel Khalili | Wired ($)

“Mistral now says Le Chonk is the ‘world’s best’ open-weight model for automated coding and cybersecurity. The advances, Mistral says, were made possible by feeding the model increasingly difficult cybersecurity and coding problems, which allowed the trillion-parameter Le Chonk to match or surpass models from leading Chinese labs like DeepSeek and Z.ai.”

The post This Week’s Awesome Tech Stories From Around the Web (Through October 31) appeared first on SingularityHub.

How is AI affecting collaboration and creativity? Cognitive scientists ran an experiment to find out.

The productivity gains of generative AI are often touted as the main reason for its rapid adoption. Much like how a calculator can instantly solve mathematical equations too long for the human brain to compute, generative AI can help someone quickly summarize long articles and draft emails.

But beyond productivity, a key question remains: How is AI affecting quintessentially human activities, such as collaboration and creativity?

As cognitive scientists who study creativity, we recently conducted an experiment that explored how working with AI—rather than humans—affects originality and other aspects of creativity.

Speedy and Wordy Doesn’t Mean Creative

Creativity is often tied to what cognitive scientists call “divergent thinking.” One way of measuring divergent thinking is to ask someone to come up with a range of uses for a common item, such as a brick or a book. This is known as the alternate uses task.

Researchers then compare the number of ideas and the originality of these ideas to assess creative capacity.

Large language models appear to have the “number of ideas” part down pat: They can generate many more ideas than humans can. This ability to produce more—and produce that high volume more quickly—partially explains why more than half of new articles online are now generated by AI.

These texts, however, are often verbose, with redundant information. Researchers who tested 12 AI chatbots across different companies on thousands of factual questions found that their answers were often many times longer than human-written answers, with the actual answer making up only about 42% of the text. The rest was largely extra padding: repetitive information, unnecessary background details, and explanations that went into too much detail or veered off topic.

Large language models also seem to struggle with the originality of their outputs.

An ‘Original’ Idea—Again and Again

In our research, we compared the results of people working on the alternate uses task with another person to those who collaborated with ChatGPT on the same task. We randomly assigned 208 participants to work with either AI or another person.

We found that compared with two people taking turns together, people working with ChatGPT wrote longer responses on the alternate uses task. But human pairs came up with more creative ideas than the humans working with ChatGPT.

We also found that the ideas people came up with when working with other people tended to be more distinct and diverse than those who worked with large language models. Individual AI suggestions could be quite creative—using a book as a sound dampener, for example—but ChatGPT presented that same idea across many separate sessions, making its overall pool of ideas more homogeneous.

Other studies have highlighted the homogeneity of generative AI’s outputs. One group of researchers found that using AI repeatedly over time to generate images often results in bland, generic pictures—what they call “visual elevator music.” Other studies have found that although AI may help an individual write a more creative story when compared with their preexisting work, AI-assisted writing as a whole, across all users, tends to be more similar.

All of this makes sense when you think about how large language models work. After being trained on massive sets of texts, large language models predict the next most likely words. Because their training processes were similar, different large language models, whether it’s ChatGPT, Claude or Gemini, tend to have similar cadences, diction, syntax, and suggestions—particularly when compared with the work of two random human writers.

In other words, generative AI’s ideas may seem unique in a vacuum. But if everyone’s using AI, the ideas and writing style start to converge.

Creative Collaboration Is a Long Game

Finally, we also found that the creativity benefits of working with another person became even stronger as the collaboration progressed.

Ideas can be measured through semantic distance, which is a way of quantifying how distinct individual concepts are. Over time, interacting with other humans seems to result in larger semantic distance than interacting with a chatbot. In other words, people may traverse more creative terrain when collaborating with other people. Meanwhile, the tendency for ideas to converge in chatbot-assisted brainstorming could accumulate, potentially narrowing the range of ideas people explore over the course of a long period of time.

This is important because many creative projects involve long-term engagement.

For example, to launch a new product, researchers have found that it takes an average of 27 months to go from the initial idea to the finished product. More innovative products can take as long as 54 months.

Ben Affleck and Matt Damon co-wrote the screenplay for ‘Good Will Hunting,’ which won the 1998 Academy Award for best original screenplay.

Chat-Human Before You ChatGPT

If you want to work with a chatbot, how can you preserve your creativity in the face of AI’s tendency toward homogeneity?

Some research suggests using what’s called “persona prompting.” Like talking to other people with different perspectives, users could provide an extensive prompt to give chatbots a pretend personality that can simulate different types of people. For example, you might prompt it to respond as if it is “a male college professor from India who is an expert in teaching” or “a 58-year-old female founder of a construction company.”
Researchers found that while this persona prompting increased ChatGPT’s idea diversity, the overall range of ideas still fell short of that produced by human writers. These findings echo other recent research suggesting that chatbots are poor simulators of human personality.

If you need help with a creative task, we suggest going back to the basics: Chat with a person before chatting with AI. They’ll have idiosyncratic personal experiences that can’t be captured by a large language model that’s trained to generate average responses.

Composing a sentence, however modest, is a creative act, whether it’s in a text to a friend or a wedding speech. As AI is integrated into workflows, it’s important to remember that other people can still be a valuable resource. Work completed in collaboration with other humans feels more meaningful.

They might not respond as quickly or as comprehensively. But chances are they’ll have something more interesting to say.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

The post AI May Quickly Generate Ideas—but Teams of People Are Still More Creative appeared first on SingularityHub.

The first patient received a dose of the high-precision gene editor in a new clinical trial for Wilson disease. Researchers hope it’s a one-and-done fix.

Prime editing is finally ready for prime time. Introduced in 2019, the CRISPR-based gene editor wowed scientists with its versatility and precision. A search-and-replace engine for the human genome, it can swap individual DNA letters, insert new genetic instructions, or delete chunks, all with minimal damage to the genome. In theory, it could correct up to 89 percent of known genetic variants associated with disease.

Now, scientists at Prime Medicine in Cambridge, Massachusetts, are testing the technology’s ability to make edits inside the body for the first time. In a global clinical trial for Wilson disease, the team is giving participants a prime editor that directly rewrites DNA in liver cells. The first patient received an infusion earlier this month, but results showing whether it’s safe and how well it works aren’t expected until next year.

The therapy has “the potential to offer patients a one-time curative treatment, ideally before irreversible liver and neurological damage occur,” Mohammed Asmal, chief medical officer of Prime Medicine, said in a press release.

Version Update

Gene editing made history in 2024 when regulators approved Casgevy for the inherited blood disorders sickle cell disease and beta thalassemia. It was the first CRISPR-based treatment to receive the green light. Casgevy uses CRISPR-Cas9 to edit blood stem cells outside the body, switching on a fetal form of a blood protein that compensates for the faulty adult version.

The classic version of CRISPR cuts through both DNA strands. Over the years, scientists have expanded on it significantly. Rather than simply disabling a gene, some variants can insert new genes, flip DNA sequences, or delete millions of DNA letters across a chromosome. Other spin-offs target RNA—the messenger molecule delivering genetic instructions—rather than DNA itself.

But for all these upgrades, CRISPR-Cas9 has a glaring weakness: It’s rough on the genome. Much of what we call “editing” is actually cells scrambling to repair broken DNA strands. Sometimes, they botch the job. In one glaring example, a single cut triggered the loss of large portions of a chromosome, which one expert called “genome vandalism.”

To rein in collateral damage, scientists redesigned the system to nick only one DNA strand. Base editing emerged as a powerful alternative. Here, the tool swaps one DNA letter for another without deliberately breaking the double helix.

Base editing made medical history last year, when doctors gave a baby, KJ Muldoon, a custom-made base editor to treat a life-threatening genetic disorder. Delivered to his cells inside lipid nanoparticles, the editor corrected a DNA mutation and eased his symptoms in weeks. Multiple clinical trials are now exploring base editing for other diseases. However, the technology can make only certain types of DNA-letter swaps.

Prime editing takes the concept a step further, turning the gene editor into a genetic word processor.

Like base editors, prime editing cuts a single DNA strand. A second component, pegRNA, carries a blueprint of the desired edit and steers the machinery to its target. Once there, a third component copies the RNA instructions into DNA, which the cell can incorporate into its genome. In a variety of tests, the inventors showed prime editing could swap any DNA letter for another, add new genetic code, or delete short stretches with high precision.

But there was a problem. Prime editors are hefty molecular machines. Getting them into the right cells is tricky.

Special Delivery

In prime editing’s first medical foray, researchers targeted chronic granulomatous disease, a rare immune disorder. Here, genetic defects prevent certain microbe-fighting immune cells from attacking invaders, leaving patients vulnerable to dangerous infections.

Scientists have tried multiple fixes. One, a viral gene therapy, restored neutrophil function in six of seven surviving patients but risked disrupting normal gene activity. In another effort, researchers turned to base editing and a clinical trial is now recruiting patients.

Prime Medicine turned to prime editing to tackle a common form of the disorder caused by two missing DNA letters.

In an earlier study, their prime editor restored those letters in blood stem cells collected from two patients—a repair that classic CRISPR-Cas9 and base editors can’t reliably complete. Within a month of receiving the edited cells, both patients had neutrophils capable of producing pathogen-killing chemicals. For one, the benefits lasted at least half a year. Side effects were mostly caused by the chemotherapy used to clear existing stem cells before the infusion.

“We created prime editing five and a half years ago as a versatile, precise genome editing technology that in principle could correct almost all mutations known to cause genetic diseases,” David Liu, a cofounder of Prime Medicine and whose lab developed prime editing, said in a press release last year. “Today’s data represent a milestone in medicine.”

Despite the early success, Prime Medicine announced it would no longer develop the therapy on its own, shifting their focus instead toward editing genes directly inside the body. Part of the decision was driven by cost: Each patient needed a custom batch of their own edited cells.

The new Wilson disease trial reflects that pivot. Patients with the condition harbor mutations in a gene encoding a protein that shuttles copper out of liver cells. Symptoms often emerge during childhood, and without treatment, copper buildup progressively damages the liver, nerves, and brain.

Though a lifelong medical cocktail can manage the disease, a one-time cure could be a game changer. Prime Medicine’s drug, PM577a, packages its editor in liver-targeted lipid nanoparticles and delivers it into the bloodstream. Not every liver cell needs fixing. According to the company, correcting between 20 and 30 percent may be enough to see a benefit.

Early animal results have been promising. In monkeys, a single infusion edited roughly half of liver cells. There were no notable changes in liver function, blood chemistry, or immune response. In mice modeling Wilson disease, the editor precisely corrected 80 percent of cells with minimal off-target edits. Prime Medicine presented these findings at the 2024 American Association for the Study of Liver Diseases, but they have not yet been published in a peer-review journal.

Dosing the first patients is a milestone. But the real test is whether the treatment is safe and gets rid of excess copper over the long haul. If so, prime editing could be adapted to tackle the other mutations behind Wilson disease, and potentially entirely different genetic disorders.

Although PM577a targets “the most prevalent disease-causing mutation in the United States and Europe,” the same system could help expand the treatment to patients elsewhere in the world, where different mutations are more common, said Prime Medicine CEO Allan Reine.

The post CRISPR Prime Editing Rewrites DNA in the Human Body for the First Time appeared first on SingularityHub.

By chilling roundworms, scientists traced a molecular pathway hinting at one way cold might extend lifespan.

From ice baths and cold showers to whole-body cryotherapy, cold exposure has long been hyped as a wellness shortcut. The supposed benefits include boosted dopamine, weight loss, stronger immunity, or even slower aging.

Scientific evidence for these claims is scant, but a telltale link between temperature and longevity has intrigued scientists for over a century. Lowering ambient temperature by just a smidge can extend lifespan in the roundworm C. elegans, fruit flies, and even mice. Plenty of mechanisms could explain the curious effect, but scientists are still largely in the dark.

A new study may shed some light on the matter. When scientists grew roundworms at moderately cold temperatures, the worms steadily ramped up production of BH2, a helper agent for many enzymes, including those that build brain chemicals. Though initially produced in our muscles, BH2 roams to other parts of the body too, including the intestines and brain. Here, in the study, it boosted tyrosine production and contributed to a longer lifespan.

Mapping a longevity pathway in worms isn’t just about satisfying scientific curiosity. When the scientists supplemented human cells with BH2 or tyrosine, the treatments fended off toxic protein clumps associated with neurodegenerative diseases.

But don’t go ordering tyrosine just yet. A biological response in cells doesn’t automatically translate into a health benefit, especially one as complex as slowing aging. While worms are a useful model in longevity studies, they’re a far cry from humans. That said, an existing drug used to treat a metabolic condition can tweak the pathway, and examining clinical data from patients taking the drug could offer a useful test to see if the mechanism holds up in people.

More broadly, the findings point to muscles as active players in the aging process. Beyond movement, their chemical output could change how other tissues age. The findings “highlight the muscle’s ‘metabolome’—its full set of metabolites—as a potential source of signals that promote longevity,” wrote Hanseul Lee and Seung-Jae Lee at the Korea Advanced Institute of Science and Technology, who were not involved in the study.

Cold Case

As time ticks by, basic cell functions gradually falter. DNA’s protective end caps, called telomers, fray. Mutations accumulate in our genetic blueprint. Chronic inflammation flares across the body. Senescent “zombie” cells bathe healthy tissues in a toxic chemical brew. Cellular energy factories and waste disposal plants begin to sputter.

Dubbed the hallmarks of aging, these processes only partly depend on genetics. Environment, social factors, and lifestyle—diet, exercise, exposure to toxins—also shape our aging trajectory.

Temperature seems like an odd addition to that list. Yet nature is littered with examples. Hibernating mammals generally live longer than those who don’t. Yellow-bellied marmots, for example, spend up to seven months a year hibernating, lowering their body temperature and throttling metabolism. Scientists have found their biological aging essentially stalls during hibernation.

These observations inspired the rate-of-living theory, which suggests that the lower a creature’s metabolic rate, the longer it lives. Recent work has given the century-old idea a nuanced update by exploring how cooler temperatures directly activate mechanisms that counter some of the hallmarks of aging. But there are caveats.

Take mitochondria, our cells’ energy factories. With age, they become increasingly dysfunctional. A chill activates brown fat, which is packed with mitochondria, ramping up energy use and potentially overall metabolic health. In a small study of a dozen people exposed to mild cold for up to eight hours, most showed signs of better blood sugar regulation. But these were short-term effects, not evidence for lasting protection. Cold can also boost mitochondrial numbers, though whether they protect against aging in humans remains to be seen.

The immune system may also respond to a chill. A study in healthy people and those with inflammatory disease, found cold exposure slashed immune molecules called cytokines, which often fuel inflammation. A different study found lower temperatures slowed the accumulation of reactive oxygen species (little “bullets” that damage cellular components, including DNA).

A cellular waste removal plant, the proteasome, also seems to benefit from air conditioning. In 2023, David Vilchez at the University of Cologne and colleagues found that moderately cold temperatures boosted proteasome activity in both C. elegans and human cells. The newly invigorated cleanup crew chomped proteins known to aggregate into clumps involved in neurodegenerative diseases.

“Lowering body temperature is one of the most effective mechanisms to extend longevity,” the team wrote at the time. But translating those ideas into a practical strategy isn’t straightforward—continuously maintaining a low body temperature is a hard sell. Mimicking those effects in a pill, however, might be easier to swallow.

A Cool Way to Live Long

In the new study, Vilchez’s team compared metabolites in C. elegans grown at a comfortable 20 degrees Celsius (68 degrees Fahrenheit) with those in worms kept at temperatures five degrees lower. On average, the chilled worms lived roughly 15 percent longer.

The levels of one metabolite, BH2, skyrocketed in the cold. The molecule helps a host of enzymes do their jobs, and genetically blocking its production wiped out the extra longevity.

The team found BH2 was mostly produced in muscles and transported to other organs. In the intestines and brain, it shifted into another form that pushed cells to make more tyrosine, an amino acid often found in meats, fish, diary, and soy. Even without a cold snap, feeding tyrosine to worms unable to make BH2 helped them live longer.

“We were genuinely surprised to find the supplementation significantly extended lifespan, even at standard and warmer temperature,” said Vilchez in a press release.

The team then tested BH2 in human neurons. Mutant proteins related to Huntington’s disease, amyotrophic lateral sclerosis, and dementia often trigger cell death, but cells treated with BH2 largely escaped this fate. Meanwhile, they better maintained protein homeostasis, the delicate balance between making, folding, and clearing proteins that goes awry with age.

These findings suggest a way to mimic the benefits of cold without having to tolerate the shivers or dangerous health effects.  But a designer pill is a long way off. Safety is one concern. In mammals, a balance between BH2 and its active version regulates reactive oxygen species. Supplementation may tip that balance, causing cells to pump out more “bullets” and ironically accelerate aging. And because the pathway controls other enzymes, unexpected side effects could do more harm than good.

For now, the takeaway isn’t to stock up on BH2 or tyrosine as a fountain of youth. Rather, the results are a good starting point for the development of metabolite-based treatments for neurodegenerative disorders. And instead of going after therapies that target the aging body as a whole, the results suggest focusing on one tissue could be enough.

“These findings suggest that specifically targeting a single tissue, such as muscle, could provide organism-wide protection. This opens up new avenues for therapeutic approaches,” said Vilchez.

The post Study Could Help Explain Why Colder Body Temperature Extends Lifespan in Many Species appeared first on SingularityHub.

It’s widely believed AI’s inner workings are largely inscrutable. A former OpenAI researcher now says this narrative could be wrong.

The idea that the neural networks underpinning modern AI are “black boxes” almost impossible to understand has become received wisdom. But a provocative blog post from a researcher who played a key role in setting that narrative says these models may be far more interpretable than we thought.

The opacity of today’s leading AI models is becoming an increasingly pressing problem as the technology seeps into ever more areas of our lives. Recent incidents where agents from leading developers escaped their sandboxes to go on hacking sprees have only boosted the sense of urgency.

This has spurred interest in a research field known as mechanistic interpretability, which attempts to reverse engineer neural networks to learn what individual neurons and the circuits they create do. By detailing how these networks operate, researchers hope it’ll be easier to control and direct them.

Progress has been slow, however, and leading theories have suggested the way these networks learn leads to a tangled mess that is incredibly difficult to parse. Now though, Nick Cammarata, a researcher at AI startup Goodfire who helped lay the foundations of the field while at OpenAI, has published a 26,000-word blog post arguing that these theories are wrong and even the largest neural networks could be much simpler to read than expected.

“I think I have good news about the nature of neural networks,” Cammarata writes in the post. “Interpretability might be far more tractable, possibly in the near term, than we’ve assumed.”

The idea Cammarata is challenging is one he helped establish. In 2020, while studying a Google image-recognition model called InceptionV1, he and colleagues found neurons that seemed to respond to unrelated things—most famously one that fired on both cat faces and the fronts of cars.

The leading explanation was the concept of superposition, which suggested that when a network runs short of neurons, it starts lumping unrelated concepts together into the same neurons. And it specifically combines concepts that are so disconnected they’re unlikely to clash. This means that looking at an individual neuron tells you very little, because there’s no rhyme or reason behind what ideas it groups together.

But when Cammarata revisited the famous cat-car neuron, he realized both images shared a prominent V shape. Further tests showed the same neuron also responded to the corner of a pool table and the pointed nose of a Japanese bullet train.

Cammarata then went looking for similar explanations across the rest of InceptionV1 with the help of a specialized suite of tools and AI agents to help speed up the search. He concluded that unrelated concepts packed into neurons were usually more connected than they first appeared. This is often because a neuron starts out tracking one thing before gradually branching out into neighboring ideas as the network learns—a phenomenon he calls recursive generalization.

Cammarata has since applied the same approaches to modern language models. Preliminary results from the Qwen3-32B model suggest roughly 90 percent of its neurons track one main concept, potentially alongside a few closely related offshoots. Cammarata says he found similar results while analyzing the much larger, trillion-parameter Kimi K3 model, which he plans to detail in a forthcoming technical paper.

The blog post sparked huge debate in the AI community, including some rapturous praise from leading researchers at large AI companies. But not everyone is convinced. For a start, the blog post self-identifies as a working hypothesis rather than a rigorous empirical study. “I appreciate that he looked at thousands of examples, but even a large number here is not a substitute for rigorous evaluation,” Chris Potts at Stanford University told MIT Technology Review.

And others question how far understanding individual neurons can take you. In a series of posts on X, Jacob Hilton at the Alignment Research Center, pointed out that to truly understand what a neural network is doing you need to track and explain how they connect neurons into larger circuits.

Even Cammarata says he is surprised by the response the blog has received and admits some people have taken his findings too far. “It’s not infinitely good news; it’s not everything solved,” he told Tech Review. “There’s a bunch of open questions around, like how much the combinations of neurons matter versus just one at a time.”

Nonetheless, the work has reopened a central question in AI that many researchers considered settled. And while Cammarata’s forthcoming technical paper should give the field a clearer sense of how justified his optimism really is, there is now a sliver of hope that the intelligent machines we are building may be slightly less inscrutable than initially feared.

The post Is AI Really a Black Box? Neural Networks May Be Easier to Understand Than We Thought appeared first on SingularityHub.