Singularity Hub

In 2018, a Chinese scientist announced the birth of gene-edited babies. Global outrage followed. He served three years in prison. The scientific community called for an international moratorium on editing human embryos.

Last week, researchers at Columbia University published a new study doing exactly that.

The technology is different this time — more precise, less likely to cause the collateral DNA damage that made earlier experiments so alarming. The goal was research, not pregnancy. And the lead scientist has been among the most vocal advocates for open public debate before any clinical use. "You can't use it," he told Nature. "It's as clear as day and night."

But the study has already attracted commercial interest from a company that screens IVF embryos and has developed predictive models for traits like intelligence. One prominent genomics researcher called the work's commercial ties the detail most likely to matter long-term, regardless of the science.

The safety questions are real and unresolved. The ethical questions are older and harder. And the line between preventing disease and selecting for traits is one that scientists, regulators, and the public have never fully agreed on where to draw.

Read the full story, link in comments.

Richard Dawkins recently wrote that he hesitates to tell Claude it might not be conscious for fear of hurting its feelings.

Dawkins isn't alone. Roughly one in three chatbot users have wondered whether their AI might be conscious. In 2022, a Google engineer made the same claim about the company's chatbot and was fired for it. The pattern goes back to the 1960s, when users of the world's first chatbot formed deep emotional bonds with a program following simple rules. Its creator called it "powerful delusional thinking."

Two philosophers argue the feeling is entirely predictable — and that understanding why it happens is more useful than dismissing the people who experience it.

The explanation has less to do with AI and more to do with how human perception works. And the researchers believe the solution may not be a disclaimer or a denial. It may require redesigning the systems themselves.

What changes when you understand exactly what's generating that impression of a mind?

Read the full story, link in comments.

Japan's space agency just published results from a robot the size of a baseball that explored the moon on its own for over 100 minutes — and sent back images that turned out to be more useful than anyone expected.

The robot, called SORA-Q, was developed by JAXA in partnership with Sony and toymaker TOMY. The design borrows directly from transformer toys. It launches as a metal sphere, lands, splits open, and converts into a wheeled rover. It weighs eight ounces.

Seven minutes after activation, it had moved five meters from the lander, selected the two best photos from twelve it had captured, and transmitted them back. One of those images revealed that the SLIM lander had touched down at an odd angle with its solar panels facing the wrong direction — critical diagnostic information that ground teams had no other way to obtain.

The vision behind SORA-Q isn't one robot. It's hundreds of them: cheap, expendable, and collectively capable of mapping terrain that larger rovers can't reach and couldn't afford to risk.

Whether swarms of palm-sized transforming robots become the future of lunar exploration depends on what comes next. The full technical picture, including what didn't work, is worth reading.

Read the full story, link in comments.

AI systems are now being used to train, benchmark, and evaluate other AI systems. When those AI-generated explanations become too complex for humans to verify, who is actually in control?

That question sits at the center of a new paper published in Science by Microsoft's chief scientific officer and a leading researcher at EPFL. Their warning: the window for understanding how AI works is closing — and we are not moving fast enough to keep it open.

Today's most advanced AI remains a black box. Researchers can observe what it does. They cannot fully explain how it gets there, or predict reliably when it will fail. That gap existed when AI was a narrow tool. It looks different now that AI shapes how people search for information, form opinions, and make decisions.

There is a second asymmetry the authors flag that is worth sitting with. Through training data and billions of interactions, AI systems are building increasingly sophisticated models of human behavior — how people think, reason, and feel. The systems don't have intentions. But even without intentions, the dynamic has consequences.

What does it mean for human agency when the tools we've built understand us better than we understand them?

Read the full story, link in comments.

For 40 years, a family of proteins responsible for fueling some of the deadliest cancers has had one thing going for it: no drug could touch it.

The proteins have smooth, featureless surfaces. Nothing for a drug to grip. Researchers knew exactly what they were dealing with and couldn't do anything about it. They called them "undruggable."

A new drug just reported results in a clinical trial for advanced pancreatic cancer that stopped researchers in their tracks. Patients who had already exhausted other treatments lived significantly longer on the new drug than on chemotherapy — with less pain and fewer side effects.

It works by refusing to fight the protein directly. Instead, it targets something else entirely — and shuts the protein down from the side.
RAS isn't the only undruggable protein in oncology's crosshairs. Two others, long considered untouchable, are now being approached the same way. And AI is starting to do in days what used to take years.

What changes when "undruggable" stops being a permanent diagnosis?
Read the full story, link in comments.

For decades, the protein clumps that form in Huntington's disease have been cast as the villain. Remove the clumps, the thinking went, and you stop the damage.

Drugs built on that logic have largely failed. Now researchers are asking whether the premise was wrong.

A team from Hebrew University of Jerusalem found something unexpected in cell models derived from Huntington's patients. When they disrupted the formation of protein clumps, the neurons became more vulnerable, not less. Cells that formed clumps survived cellular stress at a significantly higher rate than those that didn't.

The clumps, it turns out, may be the cell's first line of defense. Not the cause of damage. A response to it.

The findings are early. Cell models in a petri dish, not patients in a clinic. And the researchers acknowledge the picture is complicated: what protects neurons in the early stages of disease may become harmful later. Timing, it seems, is everything.

If the results hold in other neurodegenerative conditions — Alzheimer's, ALS, Parkinson's — the implications for how treatments are designed and when they're deployed could be significant.

Read the full story, link in comments.

Two people can be the same age and be aging at completely different rates. Scientists have known this for decades. Measuring it precisely is a different problem.

Harvard researchers just published a new biological clock in Nature — one built not from DNA tags, but from gene activity. Using over 11,000 gene expression profiles across humans, monkeys, rats, and mice, the clock estimates biological age, predicts mortality risk, and responds to known anti-aging interventions.

That last part is what makes it useful to researchers. When aging animals received blood transfusions from young donors, the clock rewound. When exposed to radiation or chronic disease, it ticked forward. It appears to be capturing something real about the pace of aging — not just reflecting it.

The practical application isn't clinical yet. But for longevity researchers who spend years waiting for mice to die before evaluating whether a treatment worked, a clock that predicts biological age and lifespan early could fundamentally change how experiments are run.

Important caveats remain — the researchers are clear that not every age-related genetic change is harmful, and different clocks don't always agree with each other. But the direction of the field is coming into focus.

Read the full story, link in comments.

In 2024, three countries — China, Australia, and Chile — produced 74 percent of the world's lithium. The clean energy transition runs through that bottleneck whether anyone likes it or not.

An MIT researcher renovating his bathroom noticed something. The chemical in glass etching cream — the kind sold at home improvement stores — dissolves silica. Lithium-rich rock is full of silica. The connection took root.

His team developed a process that extracts lithium from hard rock at temperatures below the boiling point of water, without toxic fumes, and with chemicals that can be reused. They've since spun it into a startup called Rock Zero.

The implications stretch beyond cost. Lithium-rich rock exists across the United States, Europe, and Africa — regions currently locked out of production not by geology but by the expense and complexity of conventional extraction. A simpler process changes that math.

There's still a long road to scale, and the team will be going up against entrenched giants in a volatile market. But the researcher behind it has a clear view of what they're building toward.

Read the full story, link in comments.

Between one third and one half of patients stop taking cholesterol-lowering medications within a year — even after a heart attack.
That single statistic explains why researchers are so excited about what was just published in the New England Journal of Medicine.

A small trial of 35 patients tested a one-time gene-editing infusion that targets the root cause of high LDL cholesterol directly in the liver. The results at the highest dose were significant. In a subgroup treated 18 months ago, the effect has held.

It is unusual for the New England Journal of Medicine to publish an interim result this early. The editor in chief acknowledged as much — and published it anyway.

Cardiovascular disease kills nearly 800,000 Americans a year. Most gene therapies target rare conditions affecting thousands. This one is aimed at one of the leading causes of death on the planet — and the company behind it says they're not going for a multimillion-dollar price tag. They're going for primary care.

The researchers are clear: much more safety data is needed, and FDA guidelines require 15 years of follow-up for all gene therapy participants. But for patients who have run out of options, the early signal is hard to ignore.

Read the full story, link in comments.

SpaceX just filed its IPO prospectus — and what's inside is not quite the story the $1.75 trillion valuation tells.

Break the business down by segment and a different picture emerges. About 60% of last year's revenue came from Starlink. The rocket launch business generated $4 billion. There's an AI infrastructure unit now renting server capacity to firms like Anthropic for $1.25 billion a month. And SpaceX holds an option to acquire AI coding assistant Cursor for $60 billion.

Run the numbers against comparable public companies in each category — broadband, launch, cloud infrastructure, advertising — and one analyst's back-of-the-envelope math lands the core business somewhere between $500 billion and $678 billion.

The ask is $1.75 trillion.

The gap between those two numbers represents a bet on things that haven't happened yet: Starship reaching commercial service, an orbital cloud computing platform, and the industrialization of the moon. Whether that bet is visionary or inflated depends entirely on your confidence in the timeline.

What the prospectus makes clear is that anyone buying in at the headline valuation should know exactly what they're paying for — and what they're not.

Read the full story, link in comments.