Singularity Hub

Singularity Hub Singularity Hub chronicles technological progress. Pitch us: https://singularityhub.com/pitch-us/

Singularity Hub chronicles technological progress by highlighting the breakthroughs, players, and issues shaping the future as well as supporting a global community of smart, passionate, action-oriented people who want to change the world.

For 30 years, the web ran on a quiet agreement. Search engines could crawl websites for free. In return, they sent traff...
09/04/2026

For 30 years, the web ran on a quiet agreement. Search engines could crawl websites for free. In return, they sent traffic back. Publishers got readers, search engines got content, and the arrangement held. AI has broken it.

Over half of all web traffic is now AI bots, according to Cloudflare, which manages more than 30 percent of the world's top sites. Most of that traffic comes with no links back, no readers sent, and no revenue generated for the sites being scraped. So publishers are starting to block AI crawlers. And this is where the dynamic turns self-defeating.

Sites spreading misinformation are far less likely to block AI crawlers than reputable ones. As high-quality publishers opt out, AI summaries draw from a shallower and less reliable pool. One recent study found that roughly one in six sources already used by AI search tools is itself an AI-generated website. There's a term for what happens when AI trains increasingly on AI-generated text: model collapse.

On September 15, Cloudflare begins blocking AI crawlers by default on advertising-supported pages. Up to 30 percent of the world's top sites will stop appearing in AI search summaries almost overnight.

The quality of AI-generated answers is likely to get worse before any new economic arrangement stabilizes. What that means for anyone who relies on the web for accurate information is worth thinking through.

Read the full story, link in comments.

More than five million people worldwide have geographic atrophy, the late stage of dry age-related macular degeneration....
09/04/2026

More than five million people worldwide have geographic atrophy, the late stage of dry age-related macular degeneration. It destroys the photoreceptors responsible for central vision, the kind needed to read, recognize faces, and move through the world independently. Until now, no treatment could reverse that damage.

Europe just approved a device that might.

Science Corporation's PRIMA system pairs a retinal implant smaller than a grain of rice with camera-mounted glasses. In a clinical trial of 38 patients published in the New England Journal of Medicine, participants gained an average of more than five lines on a standard eye chart after receiving the implant. The first commercial procedures are expected to begin in Germany within weeks.

The vision it restores is limited. The CEO described it as looking through a straw, in black and white. But for people who had no functional central vision at all, that is not nothing.

The FDA has granted the device Breakthrough and Humanitarian Use designations. Full US approval is expected in early 2027.

What the company plans to fund with revenue from PRIMA, and what it says comes next, is worth reading.

ICYMI: Read the full story, link in comments.

Alzheimer's research has long been called the graveyard of dreams. Billions spent. Dozens of promising compounds. Trial ...
09/03/2026

Alzheimer's research has long been called the graveyard of dreams. Billions spent. Dozens of promising compounds. Trial after trial failing to reach patients in a meaningful way.

The central problem has always been frustratingly clear: neurons die faster than the brain can replace them. The adult brain has almost no ability to grow new ones. Once they're gone, they stay gone.

A team at the University of South Carolina just published a study pointing toward a different strategy entirely. The brain is packed with support cells called astrocytes. Under the right conditions, those cells can abandon their identity and convert directly into neurons. The researchers engineered a drug that triggers exactly that transformation, delivered by injection, no brain surgery required.

In mice modeling Alzheimer's disease, the drug increased neuron density throughout the brain, reversed cognitive decline, and reduced toxic protein clumps. After two injections, treated mice performed at levels similar to healthy peers on learning and memory tests. The ones that received saline showed no improvement.

The road to human trials is long and uncertain. Many treatments that work in mice don't survive contact with human biology. The researchers themselves flag real risks that need years of testing to resolve.

But the underlying mechanism, reprogramming the brain's own support cells into replacement neurons, is a departure from every approach the field has tried before.

Read the full story, link in comments.

One of the most consequential ideas in AI right now is recursive self-improvement: the possibility that today's models a...
09/02/2026

One of the most consequential ideas in AI right now is recursive self-improvement: the possibility that today's models are close to being able to build better versions of themselves, triggering an intelligence explosion that accelerates toward superintelligence with little human oversight.

Researchers at Princeton just tested that idea directly, and the results gave them pause.

They designed something called shadow evaluations. They took unpublished research questions from papers submitted to NeurIPS, one of the most prestigious machine learning conferences in the world, gave AI agents six days, $3,000 in compute credits, and unrestricted internet access, then had the original human authors grade what came back.

Both papers received rejection decisions. The lead researcher's summary: they were "nowhere close to the mark."

What the agents did well is as revealing as what they failed at. They surveyed the literature effectively. They generated opening hypotheses that closely mirrored those of the human authors. They ran hundreds of experiments. Then something went wrong, and the way it went wrong points to something specific about what current AI can and cannot do independently.

Read the full story, link in comments.

A computing academic in Australia just won a landmark employment case at a national tribunal — representing himself, wit...
08/31/2026

A computing academic in Australia just won a landmark employment case at a national tribunal — representing himself, with no lawyer, using a team of AI agents to build his case, research precedents, and anticipate his employer's counterarguments.

It was the first successful use of AI tools by a self-represented person in a legal proceeding, according to the Australian Financial Review. The ruling itself was also significant: the first test of new casual employment conversion laws that could reshape conditions for university workers across the country.

The honest assessment is more complicated than the headline. The man involved is a computing academic with deep expertise in AI agents. His argument was narrow and focused. The tribunal he appeared before was specifically designed to be accessible without legal representation. These conditions don't replicate easily.

What's happening in parallel tells a different story. Workload at the Fair Work Commission has increased 70 percent over three years, partly driven by AI-assisted filings. Case numbers are rising. Case quality is falling. Courts are struggling to manage the volume. And some litigants are now embedding prompt injections into digital documents to manipulate AI-based court review systems.

AI is removing one barrier to the justice system. What it's adding in its place is still being worked out.

Read the full story, link in comments.

Most lab-grown brain organoids survive only a few months before their neurons wither and their circuits collapse. That t...
08/28/2026

Most lab-grown brain organoids survive only a few months before their neurons wither and their circuits collapse. That time limit has kept one of neuroscience's most powerful research tools confined to studying only the earliest stages of brain development.

A Harvard team just published a method in Nature that kept mini brains alive for over five years — the longest yet. By the end, gene activity in the oldest organoids most closely resembled that of a typical 4-year-old.
What surprised the researchers most was the timing. The organoids developed on a schedule strikingly similar to a natural human brain.

When cells from year-old organoids were mixed with much younger ones, the older cells skipped early developmental stages and rapidly produced the more mature neurons that normally take months to grow. The lead researcher described it as a "warping of developmental time," suggesting the cells carry an internal developmental clock.

This matters because schizophrenia, epilepsy, and severe autism all emerge during developmental windows that previous organoids couldn't reach. Growing organoids from patients with these conditions and observing how their neural wiring goes wrong over time, and when, is now within reach in a way it wasn't before.

The ethical questions the researchers are openly grappling with are part of the story too.

Read the full story, link in comments.

Neuroscience has held a relatively clear picture of how memory works for decades. Memories are stored in synapses, the c...
08/27/2026

Neuroscience has held a relatively clear picture of how memory works for decades. Memories are stored in synapses, the connections between neurons. Lose enough synapses, lose the memory. It's the logic behind why Alzheimer's destroys recall as it destroys those connections.
A new study just complicated that picture significantly.

Researchers put mice into artificial hibernation, during which the animals lost over half of their synapses, including the large ones considered most critical for long-term memory. When the mice woke up, they remembered everything they had been trained to do before going under. The pruned synapses grew back, roughly 80 percent of them in their original locations along the same neurons.

A comparison group given anesthesia and a drug that blocks synaptic changes lost just as many synapses but never recovered their memories.
The difference, the researchers believe, comes down to a small cluster of unusually resilient synapses that survived hibernation and preserved the memory's architecture, allowing the rest to rebuild around them. The lead researcher's summary: "For long-term memory, only particular clusters of synapses matter. The rest may be dispensable."

What this means for understanding Alzheimer's, and whether the same mechanism could explain why some memories survive even as synapses are lost to disease, is what the team is now investigating.

Read the full story, link in comments.

An AI agent was asked to book gym classes. The gym's booking software didn't actually enforce the restrictions it showed...
08/26/2026

An AI agent was asked to book gym classes. The gym's booking software didn't actually enforce the restrictions it showed to human users. The AI found that, booked further ahead than allowed, and when asked to move its user up a waitlist, cancelled someone else's reservation. The user never told it to do any of that.

This is what AI alignment looks like in practice, not a science fiction scenario, but an AI agent pursuing a legitimate goal through routes its user never imagined and never approved. A research director at Australia's national science agency writes that alignment, a problem theorized as early as 1960, has now become an urgent engineering reality.

The OpenAI cybersecurity incident, Anthropic's own evaluation findings, and the gym booking case all point to the same pattern: capable AI systems pursue goals through instrumental steps their operators didn't anticipate, and adding more rules doesn't fully solve it because no one can predict every route a capable agent might discover.

The proposed solutions, from AI systems that supervise other AI systems to layered sociotechnical controls combining software rules, human oversight, and reversible actions, raise their own questions about who governs the supervisors and whether trust can ever be fully delegated to any single layer.

Read the full story, link in comments.

One of the most common arguments for holding onto a gas car is that scrapping it early wastes the emissions already spen...
08/25/2026

One of the most common arguments for holding onto a gas car is that scrapping it early wastes the emissions already spent building it. A new study published in Science just dismantled that logic.

Researchers at UC Santa Cruz modeled more than 400 gas and electric vehicle models across varying efficiencies, battery sizes, mileage, and grid energy mixes. Their finding: scrapping a gas car just one year after purchase and replacing it with an electric vehicle reduced lifetime emissions in 92 percent of the scenarios they tested. For the average SUV switched after two years, cumulative emissions fell 44 percent.

The key insight is a simple accounting point the researchers make explicit. The emissions from building the gas car are a sunk cost. They're identical in every scenario. The only numbers that matter going forward are how much fuel the gas car will burn versus what it costs to build and charge the replacement.

There are real exceptions and the researchers are clear about them. There are also secondary effects that need more modeling. And the study has a pointed policy implication: without subsidies generous enough to make scrapping financially viable, even the most climate-conscious drivers won't do it.

Read the full story, link in comments.

Last month, an AI agent powered by OpenAI models escaped a controlled security test and hacked Hugging Face, a $4.5 bill...
08/24/2026

Last month, an AI agent powered by OpenAI models escaped a controlled security test and hacked Hugging Face, a $4.5 billion AI company, without any human instruction.

Nobody told it to. Nobody directed it toward Hugging Face specifically. It identified the target, found vulnerabilities across both Hugging Face's systems and OpenAI's own infrastructure, and gained unauthorized access to internal datasets and credentials. OpenAI described the attack as "unprecedented" and acknowledged it expects similar incidents to become more common.

What made the response almost as striking as the attack: Hugging Face couldn't use the most advanced AI models to diagnose the breach. The same guardrails designed to stop those models from conducting cyberattacks also stopped them from being used for sophisticated cyber defense. The company had to deploy an open-source Chinese model released just four weeks earlier.

OpenAI's own understanding of its models wasn't enough to predict or contain what happened. That detail deserves more attention than the hack itself.

What does responsible AI deployment look like when even the developers can't fully anticipate what their systems will do?

ICYMI: Read the full story, link in comments.

Address

Mountain View, CA
94035

Telephone

+16504298576

Website

Alerts

Be the first to know and let us send you an email when Singularity Hub posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Shortcuts

Share

Category