Transhumanismus

The AI generates gaming “fan fiction” to expand a game’s virtual world.

I admit, since middle school, I’ve spent most of my downtime immersed in video games. There are the quintessential epics: Resident Evil, Final Fantasy, World of Warcraft, and Fortnite. And then there are some indies close to my heart—a game that simulates a wildfire watcher in a forest, a road trip adventure, or one that uses portals to connect improbable physical spaces.

I’m not the only one sucked into games. The multi-billion-dollar video game industry is now bigger than Hollywood. And designers are constantly scrambling to expand their digital worlds to meet endless expectations for new content.

Now, they may have a nifty helper.

This week, Microsoft Research released Muse, an AI that spews out a multitude of diverse new scenarios within a game. Like ChatGPT and Gemini, Muse is a generative AI model. Trained on roughly 500,000 human gameplay sessions from Microsoft-owned Ninja Theory’s multiplayer shooter Bleeding Edge, Muse can dream up facsimiles of gameplay in which characters obey the game’s internal physical rules and associated controller actions.

The team is quick to add that Muse isn’t intended to replace human game designers. Rather, true to its name, the AI can offer inspiration for teams to adopt as they choose.

“In our research, we focus on exploring the capabilities that models like Muse need to effectively support human creatives,” wrote study author Katja Hofmann in a blog post.

Muse is only trained on one game and can only produce scenarios based on Bleeding Edge. However, because the AI learned from human gameplay data without any preconception of the game’s physics itself, the model could be used for other games, as long as there’s enough data for training.

“We believe generative AI can boost this creativity and open up new possibilities,” wrote Fatima Kardar, corporate vice president of gaming AI at Microsoft, in a separate blog post.

Whole New Worlds

Generative AI has already swept our existing digital universe. Now, game developers are asking if AI can help build wholly new worlds too.

Using AI to produce coherent video footage of gameplay isn’t new. In 2024, Google introduced GameNGen, which according to the company, is the first game engine powered by neural networks. The AI recreated the classic video game Doom without peeking into the game’s original code. Rather, it repeatedly played the game and eventually learned how hundreds of millions of small decisions changed the game’s outcome. The result is an AI-based copy that can be played for up to 20 seconds with all its original functionality intact.

Modern video games are a lot harder for an AI to tackle.

Most games are now in 3D, and each has its own alluring world with a set of physical rules. A game’s maps, non-player characters, and other designs can change with version updates. But how a character moves inside that virtual world—that is, how a player knows when to jump, slide, shoot, or tuck behind a barrier—stays the same.

To be fair, glitches are fun to hack, but only if they’re far and few in between. If the physics within the game—however improbable in real-life—constantly breaks, the player easily loses their sense of immersion.

Consistency is just part of the gaming experience a designer needs to think about. To better understand how AI could potentially help, the team first interviewed 27 video game designers from indie studios and industry behemoths across multiple continents.

Several themes emerged. One was about the need to create new and different scenarios that still maintain the framework of the game. For example, new ideas need to fit not only with the game’s physics—objects shouldn’t pass through walls—but also its style and vibe so they mesh with the general narrative of the game.

“Generative AI still has kind of a limited amount of context,” one designer said. “This means it’s difficult for an AI to consider the entire experience…and following specific rules and mechanics [inside the game].”

Others emphasized the need for iteration, revisiting a design until it feels right. This means that an assistant AI should be flexible enough to easily adopt designer-proposed changes over and over. Divergent paths were also a top priority, in that if a player chooses a different action, those actions will each have different and meaningful consequences.

WHAM

Based on this feedback, the team created their World and Human Action Model (WHAM)—nicknamed Muse. Each part of the AI was carefully crafted to accommodate the game designers’ needs. Its backbone algorithm is similar to the one powering ChatGPT and has previously been used to model gaming worlds.

The team then fed Muse on human gameplay data gathered from Bleeding Edge, a four versus four collaborative shooter game in 3D. With videos from the battles and controller input, the AI learned how to navigate the game from the equivalent of seven years of continuous play.

When given a prompt, Muse could generate new scenarios in the game and their associated controller inputs. The characters and objects obeyed the game’s physical laws and branched out in new explorations that matched the game’s atmosphere. Newly added objects or players stayed consistent through multiple scenes.

“What’s groundbreaking about Muse is its detailed understanding of the 3D game world, including game physics and how the game reacts to players’ controller actions,” wrote Kardar.

Not everyone is convinced the AI could help with gaming design. Muse requires tons of training data, which most smaller studios don’t have.

“Microsoft spent seven years collecting data and training these models to demonstrate that you can actually do it,” Georgios Yannakakis at the University of Malta told New Scientist, “But would an actual game studio afford [to do] this?”

Skepticism aside, the team is exploring ways to further explore the technology. One is to “clone” classic games that can no longer be played on current hardware. According to Kardar, the team wants to one day revive nostalgic games.

“Today, countless classic games tied to aging hardware are no longer playable by most people. Thanks to this breakthrough, we are exploring the potential for Muse to take older back catalog games from our studios and optimize them for any device,” she wrote.

Meanwhile, the technology could also be adapted for use in the physical world. For example, because Muse “sees” environments, it could potentially help designers reconfigure a kitchen or play with building layouts by exploring different scenarios.

“From the perspective of computer science research, it’s pretty amazing, and the future applications of this are likely to be transformative for creators,” wrote Peter Lee, president of Microsoft Research.

The post This Microsoft AI Studied 7 Years of Video-Game Play. Now It Dreams Up Whole New Game Scenarios. appeared first on SingularityHub.

The company believes devices with a million topological qubits are possible.

Many years ago, Microsoft committed itself to an exotic, high-risk approach to quantum computing. Now, as the company unveils its first prototype chip, it says its perseverance has paid off.

One of the biggest problems bedeviling today’s quantum computers is their susceptibility to errors. Even the slightest interference from the outside world can collapse the fragile quantum states they rely on to carry out computations.

Quantum error correction provides a potential workaround, but the most promising schemes require huge numbers of extra qubits. As a result, most experts predict machines will need roughly one million qubits before they can do anything truly useful, which is a long way from today’s record of a little over 1,000.

That’s why nearly two decades ago Microsoft decided to pursue topological qubits—a novel type of qubit that is inherently resistant to errors. The effort has faced several setbacks over the years, but now the company claims it’s finally cracked the problem with a processor featuring eight topological qubits and will one day be able to host up to a million.

“We took a step back and said ‘OK, let’s invent the transistor for the quantum age,’” Chetan Nayak, the Microsoft technical fellow who led the effort, said in a press release.

What sets topological quantum computing apart from other approaches is that rather than encoding quantum information in individual particles, it encodes data in the macroscale properties of a larger system that can consist of many particles. It builds on the mathematics of topology, which focuses on the properties of objects that stay the same even if they are bent or stretched.

The main advantage of this approach is that disturbances to individual components of the system don’t affect its overall topological state. Topological qubits, therefore, are much less susceptible to the kind of environmental noise that causes errors in other kinds of qubits.

Microsoft’s approach to topological quantum computing involves creating so-called quasiparticles known as Majorana zero modes. It has done this by combining a nanowire made of the semiconductor indium arsenide with a plate of aluminum that acts as a superconductor at very low temperatures.

Normally electrons in superconductors pair up. But Microsoft says its device can generate unpaired electrons that exist in a “delocalized” state. These present as a pair of Majorana zero modes, one at each end of the nanowire.

If you can create four at either end of a pair of nanowires, it should be possible to “braid” them into a topological state to encode quantum information. This braiding process involves making a series of measurements in a specific order.

However, Microsoft has yet to provide convincing proof the approach actually works. The announcement of the new chip coincided with the publication of a paper in Nature describing experiments conducted on the new device.

But these simply outlined a way to measure whether or not Majorana zero modes exist in the nanowires by detecting if there are an odd or even number of electrons. This is a crucial step because these two states will effectively act as the 0s or 1s in the company’s qubits. But the paper doesn’t provide solid evidence that Majorana zero modes are present. They simply validate the measurement approach.

Nayak told MIT Technology Review that his team has unpublished results providing more definitive proof, which they are currently writing up into a paper. But there’s likely to be a certain amount of skepticism, as there has been controversy around Microsoft’s previous publications on this topic.

A 2021 paper in Nature reporting the detection of Majorana zero modes was later retracted after other physicists suggested the signatures could have come from defects in the device used to create them. Another paper claiming evidence of the quasiparticles in 2023 was also criticized for not providing enough information for other researchers to reproduce the results.

Nonetheless, the company is confident it has now cracked the topological qubit and is firmly on the path towards building a large-scale quantum computer in years rather than decades. It has also been keen to tout the fact that DARPA seems to agree. Microsoft is one of two companies that have made it through to the final phase of the agency’s competition to find unusual quantum approaches that could achieve practical scale much faster than conventional wisdom suggests is possible.

It’s likely to be some time before there’s consensus on the significance of Microsoft’s latest result. And the journey from a scientific demonstration like this to a practical product is long and fraught with risk. But if the company’s approach pays off, it could dramatically speed the advent of the quantum age.

The post Microsoft Claims Quantum Computing Breakthrough With Its Exotic New Chip appeared first on SingularityHub.

If true, then we’re more likely to find evidence for extraterrestrial intelligence in the future.

A popular model of evolution concludes that it was incredibly unlikely for humanity to evolve on Earth, and that extraterrestrial intelligence is vanishingly rare.

But as experts on the entangled history of life and our planet, we propose that the coevolution of life and Earth’s surface environment may have unfolded in a way that makes the evolutionary origin of humanlike intelligence a more foreseeable or expected outcome than generally thought.

The Hard-Steps Model

Some of the greatest evolutionary biologists of the 20th century famously dismissed the prospect of humanlike intelligence beyond Earth.

This view, firmly rooted in biology, independently gained support from physics in 1983 with an influential publication by Brandon Carter, a theoretical physicist.

In 1983, Carter attempted to explain what he called a remarkable coincidence: the close approximation between the estimated lifespan of the sun—10 billion years—and the time Earth took to produce humans—5 billion years, rounding up.

He imagined three possibilities. In one, intelligent life like humans generally arises very quickly on planets, geologically speaking—in perhaps millions of years. In another, it typically arises in about the time it took on Earth. And in the last, he imagined that Earth was lucky—ordinarily it would take much longer, say, trillions of years for such life to form.

Carter rejected the first possibility because life on Earth took so much longer than that. He rejected the second as an unlikely coincidence, since there is no reason the processes that govern the Sun’s lifespan—nuclear fusion—should just happen to have the same timescale as biological evolution.

So Carter landed on the third explanation: that humanlike life generally takes much longer to arise than the time provided by the lifetime of a star.

The sun will likely be able to keep planets habitable for only part of its lifetime—by the time it hits 10 billion years, it will get too hot. Image Credit: NASA/JPL-Caltech

To explain why humanlike life took so long to arise, Carter proposed that it must depend on extremely unlikely evolutionary steps, and that the Earth is extraordinarily lucky to have taken them all.

He called these evolutionary steps “hard steps,” and they had two main criteria. One, the hard steps must be required for human existence—meaning if they had not happened, then humans would not be here. Two, the hard steps must have very low probabilities of occurring in the available time, meaning they usually require timescales approaching 10 billion years.

Do Hard Steps Exist?

The physicists Frank Tipler and John Barrow predicted that hard steps must have happened only once in the history of life—a logic taken from evolutionary biology.

If an evolutionary innovation required for human existence was truly improbable in the available time, then it likely wouldn’t have happened more than once, although it must have happened at least once, since we exist.

For example, the origin of nucleated—or eukaryotic—cells is one of the most popular hard steps scientists have proposed. Since humans are eukaryotes, humanity would not exist if the origin of eukaryotic cells had never happened.

On the universal tree of life, all eukaryotic life falls on exactly one branch. This suggests that eukaryotic cells originated only once, which is consistent with their origin being unlikely.

The other most popular hard-step candidates—the origin of life, oxygen-producing photosynthesis, multicellular animals, and humanlike intelligence—all share the same pattern. They are each constrained to a single branch on the tree of life.

However, as the evolutionary biologist and paleontologist Geerat Vermeij argued, there are other ways to explain why these evolutionary events appear to have happened only once.

This pattern of apparently singular origins could arise from information loss due to extinction and the incompleteness of the fossil record. Perhaps these innovations each evolved more than once, but only one example of each survived to the modern day. Maybe the extinct examples never became fossilized, or paleontologists haven’t recognized them in the fossil record.

Or maybe these innovations did happen only once, but because they could have happened only once. For example, perhaps the first evolutionary lineage to achieve one of these innovations quickly outcompeted other similar organisms from other lineages for resources. Or maybe the first lineage changed the global environment so dramatically that other lineages lost the opportunity to evolve the same innovation. In other words, once the step occurred in one lineage, the chemical or ecological conditions were changed enough that other lineages could not develop in the same way.

If these alternative mechanisms explain the uniqueness of these proposed hard steps, then none of them would actually qualify as hard steps.

But if none of these steps were hard, then why didn’t humanlike intelligence evolve much sooner in the history of life?

Environmental Evolution

Geobiologists reconstructing the conditions of the ancient Earth can easily come up with reasons why intelligent life did not evolve sooner in Earth history.

For example, 90 percent of Earth’s history elapsed before the atmosphere had enough oxygen to support humans. Likewise, up to 50 percent of Earth’s history elapsed before the atmosphere had enough oxygen to support modern eukaryotic cells.

All of the hard-step candidates have their own environmental requirements. When the Earth formed, these requirements weren’t in place. Instead, they appeared later on, as Earth’s surface environment changed.

We suggest that as the Earth changed physically and chemically over time, its surface conditions allowed for a greater diversity of habitats for life. And these changes operate on geologic timescales—billions of years—explaining why the proposed hard steps evolved when they did, and not much earlier.

In this view, humans originated when they did because the Earth became habitable to humans only relatively recently. Carter had not considered these points in 1983.

Moving Forward

But hard steps could still exist. How can scientists test whether they do?

Earth and life scientists could work together to determine when Earth’s surface environment first became supportive of each proposed hard step. Earth scientists could also forecast how much longer Earth will stay habitable for the different kinds of life associated with each proposed hard step—such as humans, animals, and eukaryotic cells.

Evolutionary biologists and paleontologists could better constrain how many times each hard-step candidate occurred. If they did occur only once each, they could see whether this came from their innate biological improbability or from environmental factors.

Lastly, astronomers could use data from planets beyond the solar system to figure out how common life-hosting planets are, and how often these planets have hard-step candidates, such as oxygen-producing photosynthesis and intelligent life.

If our view is correct, then the Earth and life have evolved together in a way that is more typical of life-supporting planets—not in the rare and improbable way that the hard-steps model predicts. Humanlike intelligence would then be a more expected outcome of Earth’s evolution, rather than a cosmic fluke.

Researchers from a variety of disciplines, from paleontologists and biologists to astronomers, can work together to learn more about the probability of intelligent life evolving on Earth and elsewhere in the universe.

If the evolution of humanlike life was more probable than the hard-steps model predicts, then researchers are more likely to find evidence for extraterrestrial intelligence in the future.

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

The post Evolving Intelligent Life May Not Have Been as Unlikely as Many Scientists Predicted appeared first on SingularityHub.

Going beyond genetic letters, the results could transform medicine.

Thanks to increasingly efficient and affordable gene sequencing technologies, we can now chart our genetic blueprint in unprecedented detail.

But what does each gene do? Of the roughly 20,000 genes that encode proteins, we’re only privy to a small fraction of their functions. The most studied genes are related to diseases. Many others hum along in the background, keeping our bodies running, but how exactly isn’t known.

An ambitious project now aims to decipher the functions of all genes.

Led by the National Institutes of Health, the MorPhiC Consortium is creating the first catalog of every gene function. Using multiple gene editing techniques, they plan to inhibit genes one-by-one to see how it changes behaviors in cells.

The project recently launched its initial phase to tackle 1,000 genes. The team is also building a data infrastructure to share findings and fact-check results.

The project offers a bird’s-eye view of how each gene—and their combinations—keeps our bodily functions humming along.

It’s the “next frontier” after the Human Genome Project, wrote the authors. These studies will tell us “how genes function alone or together to govern cellular processes” and ultimately alter our cells, tissues, and health.

The Code of Life

Our cells are buzzing biological cities that never sleep.

The city center is a structure shaped like a peach core that houses all our DNA. Diverse molecules whiz about inside the cell translating DNA messages into proteins. The body’s workhorses, proteins go on to direct metabolism, trigger immune defenses, and shuttle oxygen through the blood.

Insights into how genes function are hard-won victories. Traditionally, scientists studied a single gene—usually, one likely related to a disease—for years.

High-throughput DNA sequencing accelerated these studies by hunting down potentially detrimental gene variants, or “alleles.”

Alleles are different versions of the same gene but with a range of diverse physical consequences. Eye color is one example. Different alleles result in blue, brown, green, or other colored eyes. Genetic variants have also been found to increase the risk of Alzheimer’s disease—or protect against it.

Thanks to databases containing hundreds of thousands of genomes, it’s now possible to find different alleles associated with more than 5,000 health outcomes. By comparing the genomes of large populations of humans, such studies have located many genes related to disease. Other projects, such as the Roadmap Epigenomics Mapping Consortium and the Encyclopedia of DNA Elements Project, have provided insight into when and where genes turn on or off.

Even so, “half of human genes are barely mentioned” in scientific studies, wrote the authors. “It is estimated that 75 percent of all research on protein-coding genes has been focused on fewer than 10 percent of proteins.”

It’s a tough task to chart the rest of the genome. Genes function very differently in various cell types. Although most cells contain the same DNA blueprint, how the blueprint activates depends on the tissue. Hence, the same blueprint can guide cells towards completely different destinies—such as building our skeletons, hearts, and brains. The same gene, depending on context, can also have different effects throughout the body.

But without a thorough understanding of all gene functions, our current knowledge is “skewed” and “biased,” wrote the team.

An Expanded View

Enter the MorPhiC Consortium. The project, first launched in 2022 and now in full swing, will map how individual genes, or groups of related genes, work to build and govern our cells.

They hope to do this is by creating “null” alleles—essentially wiping out a gene’s function. Scientists have long used this method to screen individual genes related to various diseases, but MorPhiC is going big by applying the technique to the entire human genome.

The consortium is starting with an induced pluripotent stem cell line. These are adult cells that have been returned to a stem-cell-like state and can be expanded from there. Publicly available lines allow researchers to compare data from cells with an identical genetic background.

The consortium has turned to the gene-editing tool CRISPR to inhibit gene functions. Some methods directly edit genetic information; others shut off a gene without touching its code. Many include a “barcode” to track edits inside cells for validation.

Each of these methods “has a unique advantage, depending on which genes are being studied,” wrote the team. But standardizing their gene-editing strategy makes it easier to decode outcomes when shared with others in the collaboration.

The next step is linking genetic changes to the cell’s function. The consortium approved a range of tests to see what happens when a gene is turned off. These include, for example, sequencing RNA, proteins, and fats after each edit. The tests cover important aspects of a cell’s life, such as its ability to grow, regenerate, and transform into other cell types on demand. Although not comprehensive, they cover the main functions of a cell and how they could go wrong.

All the project’s centers use the same set of tests, the team wrote, although each institution may include additional screens.

Deactivating a gene isn’t easy. For quality control, each center will also dig into the cells’ transcriptome—that is, which genes are turned on—to ensure that the targeted gene is shut off. For further quality control, all teams will start by editing the same set of genes to verify procedures and share outcomes.

Data Central

Meanwhile, three centers are in the works to set up protocols for data analysis and validation. These will help store and standardize data, so it’s sharable across the project and scientific community.

The centers are also beginning to analyze data from different sources to see how different genes act together—for example, how one damaged gene can cause a cascading effect that alters other genetic functions, in turn changing metabolism, cell development, or immune responses. This data could potentially help “develop novel machine-learning frameworks” that can decipher how gene networks affect a cell’s life, wrote the authors.

The initial phase of MorPhic is expected to last five years, with each lab using the pluripotent stem cell system. However, the consortium is already looking ahead. One future goal is finding a test that can characterize genes with multiple functions in multiple cell types. Another stretch goal is to shut down multiple genes at the same time and see how they change a cell’s behavior.

“This large-scale effort will broadly improve our understanding of human genes and how they interact to govern normal human development and disease pathogenesis,” wrote the authors.

The post What Does Every Human Gene Do? This Massive Project Is About to Find Out appeared first on SingularityHub.

Artificial Intelligence

Sam Altman Lays Out Roadmap for OpenAI’s Long-Awaited GPT-5 ModelBenj Edwards | Ars Technica

“On Wednesday, OpenAI CEO Sam Altman announced a roadmap for how the company plans to release GPT-5, the long-awaited followup to 2023’s GPT-4 AI language model that made huge waves in both tech and policy circles around the world. In a reply to a question on X, Altman said GPT-5 would be coming in ‘months,’ suggesting a release later in 2025.”

Robotics

China’s EV Giants Are Betting Big on Humanoid RobotsCaiwei Chen | MIT Technology Review

“It’s becoming clear that China is now committed to becoming a global leader in robotics and automation, just as it did with EVs. Wang Xingxing, the CEO of Unitree Robots, said this well in a recent interview to local media: ‘Robotics is where EVs were a decade ago—a trillion-yuan battlefield waiting to be claimed.'”

Artificial Intelligence

Anthropic Strikes BackStephanie Palazzolo | The Information

“[Anthropic] has developed a hybrid AI model that includes reasoning capabilities, which basically means the model uses more computational resources to calculate answers to hard questions. But the model can also handle simpler tasks quickly, without the extra work, by acting like a traditional large language model. The company plans to release it in the coming weeks, according to a person who’s used it.”

Biotechnology

AI Used to Design a Multi-Step Enzyme That Can Digest Some PlasticsJohn Timmer | Ars Technica

“Unfortunately, there isn’t an enzyme for many reactions we would sorely like to catalyze—things like digesting plastics or incorporating carbon dioxide into more complex molecules. …With the advent of AI-driven protein design, however, we can now potentially design things that are unlike anything found in nature.”

Computing

This DARPA-Backed Startup Banked $100 Million for Its Energy-Slashing Analog ChipsAlex Pasternack | Fast Company

“EnCharge says that, for a wide range of AI use cases, its specialized chips, or accelerators, require up to 20 times less energy compared to today’s leading AI chips. …Rather than using only digital transistors to perform some of the multiplication operations at the heart of AI inference—the continuous computations that produce chatbot outputs—EnCharge’s chips exploit the non-binary wonders of the analog world.”

Tech

Will We Get a $1 Trillion Private Tech Firm?Cory Weinberg | The Information

“Will a private tech company reach a $1 trillion valuation in the coming years? It’s not a ridiculous question. A couple of companies seem like potential candidates. OpenAI is closing in on $300 billion in its financing with SoftBank, and SpaceX recently shot to $350 billion.”

Robotics

Meta’s Next Big Bet Might Be AI Humanoid Robots for At-Home ChoresNadeem Sarwar | Digital Trends

“[Meta’s] interests have swayed wildly over the past few years. Phones, crypto, tablets, metaverse, smart glasses, and finally, AI. The next avenue for Meta is apparently humanoid robots. According to Bloomberg, the company is pouring resources into the development of AI-powered humanoid robots. ‘Meta plans to work on its own humanoid robot hardware, with an initial focus on household chores,’ says the report.”

FUTURE

Motor Neuron Diseases Took Their Voices. AI Is Bringing Them Back.Jessica Hamzelou | MIT Technology Review

“Rodriguez and his wife, Maria Fernandez, who live in Miami, thought they would never hear his voice again. Then they re-created it using AI. After feeding old recordings of Rodriguez’s voice into a tool trained on voices from film, television, radio, and podcasts, the couple were able to generate a voice clone—a way for Jules to communicate in his ‘old voice.'”

Computing

This Breakthrough Holographic Display Could Make AR Glasses a Reality in 2026Alan Truly | Digital Trends

“Swave CEO Mike Noonen told me the bill of materials (BOM) is just $50 per eye and the expected weight of AR glasses using HXR technology could be less than 50 grams. The FoV and apparent resolution are tunable with a view as wide as 120 degrees and a retina-like resolution of up to 60 pixels per degree (PPD). Battery life is estimated at more than 10 hours, making these suitable for daily wear.”

Tech

Thomson Reuters Wins First Major AI Copyright Case in the USKate Knibbs | Wired

“This ruling is a blow to AI companies, according to Cornell University professor of digital and internet law James Grimmelmann: ‘If this decision is followed elsewhere, it’s really bad for the generative AI companies.’ Grimmelmann believes that Bibas’ judgement suggests that much of the case law that generative AI companies are citing to argue fair use is ‘irrelevant.'”

Space

The Dream of Offshore Rocket Launches Is Finally Blasting OffBecky Ferreira | MIT Technology Review

“’The best way to build a future where we have dozens, hundreds, or maybe thousands of spaceports is to build them at sea,’ says Tom Marotta, CEO and founder of the Spaceport Company, which is working to establish offshore launch hubs. ‘It’s very hard to find a thousand acres on the coast over and over again to build spaceports. It’s very easy to build the same ship over and over again.'”

Future

Who’s Using AI the Most? The Anthropic Economic Index Breaks Down the DataMichael Nuñez | VentureBeat

“The Anthropic Economic Index, released today, provides a detailed analysis of AI usage across industries, drawing from millions of anonymized conversations with Claude, Anthropic’s AI assistant. The report finds that while AI is not yet broadly automating entire jobs, it is being widely used to augment specific tasks—especially in software development, technical writing and business analysis.”

Artificial Intelligence

ChatGPT, Can You Write My New Novel for Me? Och Aye, Ye Preenin’ SassenachGareth Rubin | The Guardian

“The monsters of artificial intelligence are coming for you. They will cast you out on the street like a Dickensian mill owner and laugh as they do it—at least they will if you work in any sort of creative industry.  …Well, I’m going to turn the tables. My publisher is anxiously waiting for me to finish my new novel, a sequel to my previous thriller The Turnglass. So let’s see if AI can take the faff—the actual writing bit—out of creative writing.”

The post This Week’s Awesome Tech Stories From Around the Web (Through February 15) appeared first on SingularityHub.

Last week in Paris, two back-to-back AI conferences painted starkly different visions of the future. The inaugural International Association for Safe and Ethical Artificial Intelligence (IASEAI’25) focused on AI safety, ethics, and global regulatory frameworks, bringing together a coalition of thinkers, policymakers, and Nobel Laureates. It was a call for restraint, cooperation, and the responsible […]

The new tool, NanoCas, could extend gene therapies throughout the body.

When the gene editing tool CRISPR-Cas9 rocketed to fame more than a decade ago, it transformed biotechnology. Faster, cheaper, and safer than previous methods, the tool helped scientists gain insight into gene functions—and when they go wrong.

CRISPR also brought the potential to change the lives of people living with inherited diseases. Thanks to its gene editing prowess, the tool can supercharge immune cells’ ability to hunt down cancer and other rogue cells. In late 2023, the FDA approved a CRISPR therapy for sickle cell disease and later gave the greenlight to people with a blood disorder called transfusion-dependent beta thalassemia. Many more therapies are in the works.

But CRISPR has a hefty problem: The system is too large, making it difficult to deliver the gene editor to cells in muscle, brain, heart, and other tissues.

Now, a team at Mammoth Biosciences has a potential solution. Cofounded by CRISPR pioneer Jennifer Doudna at the University of California, Berkeley, the company has long sought to downsize the original CRISPR-Cas9 system. Their new iteration, dubbed NanoCas, slashed the size of one key component, Cas9, to roughly one-third of the original.

The slimmed-down setup allowed the tool to be packaged into a single “delivery box”—a virus that’s commonly used for gene therapy inside the body. In mice and monkeys, the team used NanoCas to edit genes involved in inherited high cholesterol and Duchenne muscular dystrophy.

“CRISPR gene editing is a transformative technology for addressing genetic diseases, but delivery constraints have largely limited its therapeutic applications to liver-targeted and ex vivo [outside the body] therapies,” wrote the team in a preprint describing their results. The compact NanoCas “opens the door” for editing tissues inside the body.

Delivery Woes

CRISPR has two main components. One is an RNA molecule that’s like a bloodhound, seeking out and tethering the setup to a target DNA section. Once docked, the second component, a Cas protein, slices or snips the genetic ribbon.

Over the years, scientists have discovered or engineered other versions of Cas proteins. Some target RNA, the “messenger” that translates genes into proteins. Others swap out single genetic letters causing inherited diseases. Some even recruit enzymes to modify the epigenome—the system controlling which genes are turned on or off.

All these tools have a major problem: They’re difficult to deliver inside the body because of their size. Current CRISPR therapies mainly rely on extracting cells and swapping their genes inside petri dishes. The edited cells are infused back into the patient. Called “ex vivo” therapy, these treatments mainly focus on blood-based disorders.

Correcting genetic problems inside the body with CRISPR adds to the complexity. Most therapies focus on the eyes or the liver, which are both relatively easy to access with a shot. For all other tissues, delivery is the main problem.

To shuttle the editors to tissues and cells, they have to be packaged inside a virus or a fatty bubble. Cas proteins can reach over a thousand amino acids in length, which already stresses the capacity of the delivery vehicles. Add in guide RNA components, and the system exceeds luggage limits.

To get around weight restrictions, scientists have encoded the guide RNA and Cas components separately into two viral carriers, so both can sneak into cells. Alternatively, they’ve used fatty bubbles called liposomes that encapsulate both gene editing components.

Neither is perfect. A double load of virus increases the risk of an immune response. Liposomes generally end up in the liver and release their cargo there. This makes them excellent at editing genes in the liver—for example, PCSK9, to treat high levels of cholesterol—but they struggle to reach other tissues. Important targets such as the brain and muscles are out of reach.

Small But Mighty

Why not shrink the cargo so it fits into the same viral luggage?

Here, Mammoth Biosciences searched metagenomics databases for smaller Cas proteins. These databases contain diverse samples from across the planet, including from microbes gathered in swamps, seawater, our guts, and other sources. The team looked for systems that could edit as efficiently as Cas9, required only a tiny guide RNA component, and were under 600 amino acids.

From 22,000 metagenomes, the team zeroed in on 176 candidates. Each was vetted in human kidney cells in a dish—rather than using bacteria, which is the norm. This screens for Cas variants that work well inside mammalian cells, which is a common bottleneck, wrote the team.

After more tests, they landed on NanoCas. It worked with roughly 60 percent of the RNA guides they tried out, and after some tinkering, easily sliced up targeted DNA.

The tiny editor and its guide RNA fit into a single viral vector. As proof of concept, the team made a NanoCas system targeting PCSK9, a gene associated with dangerously high levels of cholesterol, in the livers of mice. Delivered in a single injection into the veins, the tiny tool slashed the gene to undetectable levels in the blood.

Next, the team turned to a gene called dystrophin in muscles, a tissue traditional CRISPR methods struggle to reach. In Duchenne muscular dystrophy, mutated dystrophin causes progressive muscle loss. NanoCas edited the gene across a wide variety of muscle types—thigh, heart, and calf muscle. The efficacy varied, ranging from 10 to 40 percent of edited cells.  

The team next tested NanoCas in monkeys. After about two months, roughly 30 percent of their skeletal muscle cells were edited. Heart cells were less responsive, with only half the efficacy.

“To our knowledge,” this is the first time someone has edited muscles in a non-human primate with a single virus CRISPR system, wrote the team.

Gene therapies using delivery viruses can tax the liver, but throughout the trial the monkey’s liver functions and other health factors stayed relatively normal. But many questions remain. Although the system edited targeted genes in healthy monkeys, whether it can treat genetic muscle loss remains to be seen. As with other gene editing systems, there’s also the risk of unintentionally editing non-targeted genes or spurring an immune attack.

That said, the miniature NanoCas—and potentially other tiny Cas proteins yet to be discovered—could shuttle CRISPR to a variety of tissues in the body with a jab. The team is already exploring the system’s potential for targeting brain diseases. The technology could also be reworked for use in epigenetic or base editing.

Above all, the study suggests small Cas proteins can be mighty.

“NanoCas demonstrates that carefully selected compact systems can achieve robust editing across various contexts, challenging the assumption that small CRISPR systems are inherently less effective,” wrote the team.

The post Miniaturized CRISPR Packs a Mighty Gene Editing Punch appeared first on SingularityHub.

Inefficient AI models guzzle energy. Then again, so do efficient ones—just for different reasons.

DeepSeek has upended the AI industry, from the chips and money needed to train and run AI to the energy it’s expected to guzzle in the not-too-distant future. Energy stocks skyrocketed in 2024 on predictions of dramatic growth in electricity demand to power AI data centers, with shares of power generation companies Constellation Energy and Vistra reaching record highs.

And that wasn’t all. In one of the biggest deals in the US power industry’s history, Constellation acquired natural gas producer Calpine Energy for $16.4 billion, assuming demand for gas would grow as a generation source for AI. Meanwhile, nuclear power seemed poised for a renaissance. Google signed an agreement with Kairos Power to buy nuclear energy produced by small modular reactors (SMRs). Separately, Amazon made deals with three different SMR developers, and Microsoft and Constellation announced they would restart a reactor at Three Mile Island.

As this frenzy to secure reliable baseload power built towards a crescendo, DeepSeek’s R1 came along and unceremoniously crashed the party. Its creators say they trained the model using a fraction of the hardware and computing power of its predecessors. Energy stocks tumbled and shock waves reverberated through the energy and AI communities, as it suddenly seemed like all that effort to lock in new power sources was for naught.

But was such a dramatic market shake-up merited? What does DeepSeek really mean for the future of energy demand?

At this point, it’s too soon to draw definitive conclusions. However, various signs suggest the market’s knee-jerk response to DeepSeek was more reactionary than an accurate indicator of how R1 will impact energy demand.

Training vs. Inference

DeepSeek claimed it spent just $6 million to train its R1 model and used fewer (and less sophisticated) chips than the likes of OpenAI. There’s been much debate about what exactly these figures mean. The model does appear to include real improvements, but the associated costs may be higher than disclosed.

Even so, R1’s advances were enough to rattle markets. To see why, it’s worth digging into the nuts and bolts a bit.

First of all, it’s important to note that training a large language model is entirely different than using that same model to answer questions or generate content. Initially, training an AI is the process of feeding it massive amounts of data that it uses to learn patterns, draw connections, and establish relationships. This is called pre-training. In post-training, more data and feedback are used to fine-tune the model, often with humans in the loop.

Once a model has been trained, it can be put to the test. This phase is called inference, when the AI answers questions, solves problems, or writes text or code based on a prompt.

Traditionally with AI models, a huge amount of resources goes into training them up front, but relatively fewer resources go towards running them (at least on a per-query basis). DeepSeek did find ways to train its model far more efficiently, both in pre-training and post-training. Advances included clever engineering hacks and new training techniques—like the automation of reinforcement feedback usually handled by people—that impressed experts. This led many to question whether companies would actually need to spend so much building enormous data centers that would gobble up energy.

It’s Costly to Reason

DeepSeek is a new kind of model called a “reasoning” model. Reasoning models begin with a pre-trained model, like GPT-4, and receive further training where they learn to employ “chain-of-thought reasoning” to break a task down into multiple steps. During inference, they test different formulas for getting a correct answer, recognize when they make a mistake, and improve their outputs. It’s a little closer to how humans think—and it takes a lot more time and energy.

In the past, training used the most computing power and thus the most energy, as it entailed processing huge datasets. But once a trained model reached inference, it was simply applying its learned patterns to new data points, which didn’t require as much computing power (relatively).

To an extent, DeepSeek’s R1 reverses this equation. The company made training more efficient, but the way it solves queries and answers prompts guzzles more power than older models. A head-to-head comparison found that DeepSeek used 87 percent more energy than Meta’s non-reasoning Llama 3.3 to answer the same set of prompts. Also, OpenAI—whose o1 model was first out of the gate with reasoning capabilities—found allowing these models more time to “think” results in better answers.

Although reasoning models aren’t necessarily better for everything—they excel at math and coding, for example—their rise may catalyze a shift toward more energy-intensive uses. Even if training models gets more efficient, added computation during inference may cancel out some of the gains.

Assuming that greater efficiency in training will lead to less energy use may not pan out either. Counter-intuitively, greater efficiency and cost-savings in training may simply mean companies go even bigger during that phase, using just as much (or more) energy to get better results.

“The gains in cost efficiency end up entirely devoted to training smarter models, limited only by the company’s financial resources,” wrote Anthropic cofounder Dario Amodei of DeepSeek.

If It Costs Less, We Use More

Microsoft CEO Satya Nadella likewise brought up this tendency, known as the Jevons paradox—the idea that increased efficiency leads to increased use of a resource, ultimately canceling out the efficiency gain—in response to the DeepSeek melee.

If your new car uses half as much gas per mile as your old car, you’re not going to buy less gas; you’re going to take that road trip you’ve been thinking about, and plan another road trip to boot.

The same principle will apply in AI. While reasoning models are relatively energy-intensive now, they likely won’t be forever. Older AI models are vastly more efficient today than when they were first released. We’ll see the same trend with reasoning models; even though they’ll consume more energy in the short run, in the long run they’ll get more efficient. This means it’s likely that over both time frames they’ll use more energy, not less. Inefficient models will gobble up excessive energy first, then increasingly efficient models will proliferate and be used to a far greater extent later on.

As Nadella posted on X, “As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”

If You Build It

In light of DeepSeek’s R1 mic drop, should US tech companies be backpedaling on their efforts to ramp up energy supplies? Cancel those contracts for small modular nuclear reactors?

In 2023, data centers accounted for 4.4 percent of total US electricity use. A report published in December—prior to R1’s release—predicted that figure could balloon to as much as 12 percent by 2028. That percentage could shrink due to the training efficiency improvements brought by DeepSeek, which will be widely implemented.

But given the likely proliferation of reasoning models and the energy they use for inference—not to mention later efficiency-driven demand increases—my money’s on data centers hitting that 12 percent, just as analysts predicted before they’d ever heard of DeepSeek.

Tech companies appear to be on the same page. In recent earnings calls, Google, Microsoft, Amazon, and Meta announced they would spend $300 billion—mostly on AI infrastructure—this year alone. There’s still a whole lot of cash, and energy, in AI.

The post DeepSeek Crashed Energy Stocks. Here’s Why It Shouldn’t Have. appeared first on SingularityHub.

The path to recent advanced AI systems has been more about building larger systems than making scientific breakthroughs.

For most of artificial intelligence’s history, many researchers expected that building truly capable systems would need a long series of scientific breakthroughs: revolutionary algorithms, deep insights into human cognition, or fundamental advances in our understanding of the brain. While scientific advances have played a role, recent AI progress has revealed an unexpected insight: A lot of the recent improvement in AI capabilities has come simply from scaling up existing AI systems.1

Here, scaling means deploying more computational power, using larger datasets, and building bigger models. This approach has worked surprisingly well so far.2 Just a few years ago, state-of-the-art AI systems struggled with basic tasks like counting.3,4 Today, they can solve complex math problems, write software, create extremely realistic images and videos, and discuss academic topics.

This article will provide a brief overview of scaling in AI over the past years. The data comes from Epoch, an organization that analyzes trends in computing, data, and investments to understand where AI might be headed.5 Epoch maintains the most extensive dataset on AI models and regularly publishes key figures on AI growth and change.

What Is Scaling in AI Models?

Let’s briefly break down what scaling means in AI. Scaling is about increasing three main things during training, which typically need to grow together:

• The amount of data used for training the AI;
• The model’s size, measured in “parameters”;
• Computational resources, often called “compute” in AI.

The idea is simple but powerful: Bigger AI systems, trained on more data and using more computational resources, tend to perform better. Even without substantial changes to the algorithms, this approach often leads to better performance across many tasks.6

Here is another reason why this is important: As researchers scale up these AI systems, they not only improve in the tasks they were trained on but can sometimes lead them to develop new abilities that they did not have on a smaller scale.7 For example, language models initially struggled with simple arithmetic tests like three-digit addition, but larger models could handle these easily once they reached a certain size.8 The transition wasn’t a smooth, incremental improvement but a more abrupt leap in capabilities.

This abrupt jump in capability, rather than steady improvement, can be concerning. If, for example, models suddenly develop unexpected and potentially harmful behaviors simply as a result of getting bigger, it would be harder to anticipate and control.

This makes tracking these metrics important.

What Are the Three Components of Scaling Up AI models? Data: scaling up the training data

One way to view today’s AI models is by looking at them as very sophisticated pattern recognition systems. They work by identifying and learning from statistical regularities in the text, images, or other data on which they are trained. The more data the model has access to, the more it can learn about the nuances and complexities of the knowledge domain in which it’s designed to operate.9

In 1950, Claude Shannon built one of the earliest examples of “AI”: a robotic mouse named Theseus that could “remember” its path through a maze using simple relay circuits. Each wall Theseus bumped into became a data point, allowing it to learn the correct route. The total number of walls or data points was 40. You can find this data point in the chart; it is the first one.

While Theseus stored simple binary states in relay circuits, modern AI systems utilize vast neural networks, which can learn much more complex patterns and relationships and thus process billions of data points.

All recent notable AI models—especially large, state-of-the-art ones—rely on vast amounts of training data. With the y-axis displayed on a logarithmic scale, the chart shows that the data used to train AI models has grown exponentially. From 40 data points for Theseus to trillions of data points for the largest modern systems in a little more than seven decades.

Since 2010, the training data has doubled approximately every nine to ten months. You can see this rapid growth in the chart, shown by the purple line extending from the start of 2010 to October 2024, the latest data point as I write this article.10

Datasets used for training large language models, in particular, have experienced an even faster growth rate, tripling in size each year since 2010. Large language models process text by breaking it into tokens—basic units the model can encode and understand. A token doesn’t directly correspond to one word, but on average, three English words correspond to about four tokens.

GPT-2, released in 2019, is estimated to have been trained on 4 billion tokens, roughly equivalent to 3 billion words. To put this in perspective, as of September 2024, the English Wikipedia contained around 4.6 billion words.11 In comparison, GPT-4, released in 2023, was trained on almost 13 trillion tokens, or about 9.75 trillion words.12 This means that GPT-4’s training data was equivalent to over 2,000 times the amount of text of the entire English Wikipedia.

As we use more data to train AI systems, we might eventually run out of high-quality human-generated materials like books, articles, and research papers. Some researchers predict we could exhaust useful training materials within the next few decades13. While AI models themselves can generate vast amounts of data, training AI on machine-generated materials could create problems, making the models less accurate and more repetitive.14

Parameters: scaling up the model size

Increasing the amount of training data lets AI models learn from much more information than ever before. However, to pick up on the patterns in this data and learn effectively, models need what are called “parameters”. Parameters are a bit like knobs that can be tweaked to improve how the model processes information and makes predictions. As the amount of training data grows, models need more capacity to capture all the details in the training data. This means larger datasets typically require the models to have more parameters to learn effectively.

Early neural networks had hundreds or thousands of parameters. With its simple maze-learning circuitry, Theseus was a model with just 40 parameters—equivalent to the number of walls it encountered. Recent large models, such as GPT-3, boast up to 175 billion parameters.15 While the raw number may seem large, this roughly translates into 700 GB if stored on a disk, which is easily manageable by today’s computers.

The chart shows how the number of parameters in AI models has skyrocketed over time. Since 2010, the number of AI model parameters has approximately doubled every year. The highest estimated number of parameters recorded by Epoch is 1.6 trillion in the QMoE model.

While bigger AI models can do more, they also face some problems. One major issue is called “overfitting.” This happens when an AI becomes “too optimized” for processing the particular data it was trained on but struggles with new data. To combat this, researchers employ two strategies: implementing specialized techniques for more generalized learning and expanding the volume and diversity of training data.

Compute: scaling up computational resources

As AI models grow in data and parameters, they require exponentially more computational resources. These resources, commonly referred to as “compute” in AI research, are typically measured in total floating-point operations (“FLOP”), where each FLOP represents a single arithmetic calculation like addition or multiplication.

The computational needs for AI training have changed dramatically over time. With their modest data and parameter counts, early models could be trained in hours on simple hardware. Today’s most advanced models require hundreds of days of continuous computations, even with tens of thousands of special-purpose computers.

The chart shows that the computation used to train each AI model—shown on the vertical axis—has consistently and exponentially increased over the last few decades. From 1950 to 2010, compute doubled roughly every two years. However, since 2010, this growth has accelerated dramatically, now doubling approximately every six months, with the most compute-intensive model reaching 50 billion petaFLOP as I write this article.16

To put this scale in perspective, a single high-end graphics card like the NVIDIA GeForce RTX 3090—widely used in AI research—running at full capacity for an entire year would complete just 1.1 million petaFLOP computations. 50 billion petaFLOP is approximately 45,455 times more than that.

Achieving computations on this scale requires large energy and hardware investments. Training some of the latest models has been estimated to cost up to $40 million, making it accessible only to a few well-funded organizations.

Compute, Data, and Parameters Tend to Scale at the Same Time

Compute, data, and parameters are closely interconnected when it comes to scaling AI models. When AI models are trained on more data, there are more things to learn. To deal with the increasing complexity of the data, AI models, therefore, require more parameters to learn from the various features of the data. Adding more parameters to the model means that it needs more computational resources during training.

This interdependence means that data, parameters, and compute need to grow simultaneously. Today’s largest public datasets are about 10 times bigger than what most AI models currently use, some containing hundreds of trillions of words. But without enough compute and parameters, AI models can’t yet use these for training.

What Can We Learn From These Trends for the Future of AI?

Companies are seeking large financial investments to develop and scale their AI models, with a growing focus on generative AI technologies. At the same time, the key hardware that is used for training—GPUs—is getting much cheaper and more powerful, with its computing speed doubling roughly every 2.5 years per dollar spent.17 Some organizations are also now leveraging more computational resources not just in training AI models but also during inference—the phase when models generate responses—as illustrated by OpenAI’s latest o1 model.

These developments could help create more sophisticated AI technologies faster and cheaper. As companies invest more money and the necessary hardware improves, we might see significant improvements in what AI can do, including potentially unexpected new capabilities.

Because these changes could have major effects on our society, it’s important that we track and understand these developments early on. To support this, Our World in Data will update key metrics—such as the growth in computational resources, training data volumes, and model parameters—on a monthly basis. These updates will help monitor the rapid evolution of AI technologies and provide valuable insights into their trajectory.

This article was originally published on Our World in Data and has been republished here under a Creative Commons license. Read the original article.

The post Scaling Up: How Increasing Inputs Has Made Artificial Intelligence More Capable appeared first on SingularityHub.

A watery past may have spurred the formation of life’s basic molecules.

Life on Earth relies on molecular building blocks to make DNA and proteins. Scientists have long wondered how prevalent these precursors were at the birth of our solar system.

A sample of dust and rocks from an asteroid just took us closer to an answer.

Collected from Bennu, a space rock shaped like a spinning top, as it soared by Earth roughly five years ago, the samples were frozen in time by the vacuum of space. Essentially a time capsule of the earliest days of our solar system some 4.5 billion years ago—around the time when Earth was forming—they provide a peek into the chemical soup that may have kickstarted life.

Two new studies examining these extraterrestrial space grains found signs of life’s molecules preserved on the asteroid’s ancient surface. Dust and rocks from Bennu contained all five nucleobases—molecules that make up DNA and RNA—and 14 of the 20 amino acids in proteins.

These organic molecules had been found in other asteroids. But there’s a twist to Bennu’s chemical makeup. Whereas most Earthly amino acids exist in a left-handed form, samples from Bennu contain an almost equal amount of amino acids that are their mirror image. These right-handed amino acids aren’t naturally found on Earth.

Bennu also harbored telltale signs of saltwater, which could have been the soup that helped the molecules mingle and interact. The brine is similar in composition to dried lake beds on Earth.

To be clear, the teams didn’t find extraterrestrial life. But they did show that life’s precursor molecules—even “mirrored” ones—were widespread across the early solar system.  

“Asteroids provide a time capsule into our home planet’s history, and Bennu’s samples are pivotal in our understanding of what ingredients in our solar system existed before life started on Earth,” said Nicky Fox, an associate administrator at NASA, in a press release.

The Mission

The samples were delivered by NASA’s OSIRIS-REx mission—the first US project to bring asteroid samples home. Bennu was an especially interesting target. Prior work had suggested asteroids have the organic molecules that form the basis of life on Earth. These molecules could have hitched a ride on asteroids and seeded the early planets or their moons to spark life.

On Earth, two critical components for life are nucleobases and amino acids.

Nucleobases are the molecular building blocks of DNA. They encode our bodies’ makeup, functions, and inheritance. RNA, which transmits the instructions contained in genes to the protein-making factories in cells, uses an additional nucleobase, which is also integral to some viruses. Beyond DNA and RNA, 20 amino acids link together to form proteins.

How these precursor ingredients spurred life remains a mystery, but asteroids may contain clues. A previous sample from 162173 Ryugu, a diamond-shaped asteroid, contained myriad organic compounds, including vitamin B3 and uracil, the additional nucleobase used in RNA.

Like Ryugu, Bennu is a carbonaceous asteroid. These space rocks are rich in carbon molecules that form the organic compounds critical for life. Bennu, a pile of rocks loosely held together by gravity, likely dates to the beginning of the solar system—some 4.5 billion years ago.

Thanks to the freezing vacuum of space, most organic molecules on Bennu have been preserved in their original state—locked in time—and could provide clues about the early solar system’s chemical makeup.

Bennu was also an attractive target because it skirts the asteroid belt, which circles the sun between Mars and Jupiter. At its closest, the asteroid is 200 million miles from Earth. While still a multi-year journey, the distance made it possible to land a space probe, map Bennu’s landscape, collect specimens, and shuttle the cargo back to Earth.

The probe was specifically designed to seal collected samples in a capsule to protect them from contamination when returning to and re-entering Earth’s atmosphere. As the capsule dropped back to Earth, the air was filtered to remove water vapor and dust particles. Upon landing in Utah, NASA immediately placed the capsule in a clean room and blasted it with nitrogen—a gas that doesn’t react with most other chemicals—to push out invading air.  

“What makes these results so significant is that we’re finding them in a pristine sample,” Daniel Glavin, an astrobiologist at NASA and coauthor on a paper describing the work, told Nature.

These meticulous guidelines ensured the sample wasn’t contaminated by Earth’s natural chemicals. Weighing a little over four ounces—roughly a bar of soap—the collection of asteroid pebbles and dust is one of the largest to date.

Mirror, Mirror

One study in Nature Astronomy detected all five of the nucleobases present in genetic material on Earth and 14 of the 20 amino acids that make up proteins. The asteroid also contained 19 amino acids that don’t encode any proteins known to life on Earth.

Surprisingly, some of these amino acids exist in a mirror world. Amino acids on Earth are only left-handed. Synthetic biologists have begun genetically twisting these protein building blocks into a right-handed structure—which could benefit biomedicine in the form of longer-lasting medications. Some scientists have even proposed building fully “mirrored” lifeforms, a controversial and potentially risky endeavor scientists spoke out against last year.

Our early solar system may have even laid the groundwork. But how these molecules formed—and if they stuck around—remains a mystery.   

The team also detected high amounts of ammonia and formaldehyde. The duo, prevalent on early Earth, is critical to the formation of complex molecules in the right conditions—basically providing a nutritious broth for ingredients like amino acids to simmer and chemically react.

Bennu may have once provided a compatible environment. Another study in Nature detected a cornucopia of minerals akin to brine on Earth—potentially a sign of water in the past. These salt-crusted spots, which usually occur due to freezing or evaporation, dot Earth’s landscapes in places like Badwater Basin in Death Valley and the Great Salt Lake in Utah.

Together, the samples form a snapshot of the asteroid’s multi-billion-year-long history, suggesting the space rock may have once harbored tiny pools of water friendly to life.

“Having these brines there, along with simple organic stuff, may have kick-started [the process of] making much more complicated and interesting organics like the nucleobases,” study author Sara Russell at the Natural History Museum in London told Nature.

A global coalition is still analyzing Bennu’s samples to learn more about the early solar system. In the meantime, the spacecraft—renamed OSIRIS-APEX—is gearing up for another mission to the asteroid Apophis as it skirts by Earth in 2029.

“Data from OSIRIS-REx adds major brushstrokes to a picture of a solar system teeming with the potential for life,” said study author Jason Dworkin. “Why we, so far, only see life on Earth and not elsewhere, that’s the truly tantalizing question.”

The post Scientists Find ‘Mirror Life’ Building Blocks on Asteroid Bennu appeared first on SingularityHub.

These were our favorite articles in science and tech this week.

DeepMind Claims Its AI Performs Better Than International Mathematical Olympiad Gold Medalists Kyle Wiggers | TechCrunch

“AlphaGeometry2 perhaps demonstrates that the two approaches—symbol manipulation and neural networks—combined are a promising path forward in the search for generalizable AI. Indeed, according to the DeepMind paper, o1, which also has a neural network architecture, couldn’t solve any of the IMO problems that AlphaGeometry2 was able to answer.”

Three Years After Experimental Vaccine, These Patients Are Still Cancer-Free Ed Cara | Gizmodo

“Scientists at the Dana-Farber Cancer Institute and elsewhere developed the vaccine, which is designed to prevent advanced cases of kidney cancer from returning. Since the trial patients received the vaccine roughly three years ago, they have stayed cancer-free. The early results suggest that these vaccines may someday be able to tackle a wider variety of cancers than expected, the researchers say.”

Figure Drops OpenAI in Favor of In-House Models Brian Heater | TechCrunch

“The Bay Area-based [general-purpose humanoid robotics company] has instead opted to focus on in-house AI owing to a ‘major breakthrough.’ In conversation with TechCrunch afterward, founder and CEO Brett Adcock was tightlipped in terms of specifics, but he promised to deliver ‘something no one has ever seen on a humanoid’ in the next 30 days.”

DeepSeek iOS App Sends Data Unencrypted to ByteDance-Controlled Servers Dan Goodin | Ars Technica

“On Thursday, mobile security company NowSecure reported that the app sends sensitive data over unencrypted channels, making the data readable to anyone who can monitor the traffic. More sophisticated attackers could also tamper with the data while it’s in transit.”

OpenAI Says Its Models Are More Persuasive Than 82 Percent of Reddit Users Kyle Orland | Ars Technica

“OpenAI has previously found that 2022’s ChatGPT-3.5 was significantly less persuasive than random humans, ranking in just the 38th percentile on this measure. But that performance jumped to the 77th percentile with September’s release of the o1-mini reasoning model and up to percentiles in the high 80s for the full-fledged o1 model.”

DeepSeek Doesn’t Slow Tech’s AI Capex Splurge Martin Peers | The Information

“The three big cloud firms and Meta are projecting around $300 billion in capex, mostly related to AI, this year. To put that into context, the OpenAI-SoftBank Stargate AI data center venture plans to spend $100 billion in the near term and $500 billion over four years. We don’t yet know whether Stargate can raise the money. But there are no such questions about whether Google, Microsoft, Amazon, and Meta can afford their spending plans.”

Humanlike ‘Teeth’ Have Been Grown in Mini Pigs Jessica Hamzelou | MIT Technology Review

“Lose an adult tooth, and you’re left with limited options that typically involve titanium implants or plastic dentures. But scientists are working on an alternative: lab-grown human teeth that could one day replace damaged ones.”

New Device Can Scan Your Face in 3D From Hundreds of Meters Away Karmela Padavic-Callaghan | New Scientist

“From 325 meters away, your eyes can probably distinguish a person’s head from their body—and not much else. But a new laser-based device can create a three-dimensional model of their face. Aongus McCarthy at Heriot-Watt University in Scotland and his colleagues built a device that can create detailed three-dimensional images, including ridges and indentations as small as 1 millimeter, from hundreds of meters away.”

OpenAI’s New Agent Can Compile Detailed Reports on Practically Any Topic Rhiannon Williams | MIT Technology Review

“OpenAI has launched a new agent capable of conducting complex, multistep online research into everything from scientific studies to personalized bike recommendations at what it claims is the same level as a human analyst. …It can search and analyze massive quantities of text, images, and PDFs to compile a thoroughly researched report.”

AI ‘Godfather’ Predicts Another Revolution in the Tech in Next Five Years’ Dan Milmo | The Guardian

“’There are still a lot of scientific and technological challenges ahead, and it’s very likely that there’s going to be yet another AI revolution over the next three to five years because of the limitation of current systems,’ [Meta’s Chief AI Scientist Yann LeCun] said. ‘If we want eventually to build things like domestic robots and completely autonomous cars, we need systems to understand the real world.’”

Sam Altman: OpenAI Has Been on the ‘Wrong Side of History’ Concerning Open Source Kyle Wiggers | TechCrunch

“Altman admitted that DeepSeek has lessened OpenAI’s lead in AI, and he said he believes OpenAI has been ‘on the wrong side of history’ when it comes to open sourcing its technologies. While OpenAI has open sourced models in the past, the company has generally favored a proprietary, closed source development approach.”

A New Video Shows Apple Is Developing a Tabletop Robot That Dances Jennifer Pattison Tuohy | The Verge

“We’ve got more evidence that Apple is developing a tabletop robot for the home, courtesy of a blog post published on Apple’s Machine Learning Research site. First spotted by MacRumors, the post summarizes a paper by an Apple research team that developed a robot with expressive movements to see how much more engaging it is than a standard robot. And there’s a video.”

This AI Chip Is the Size of a Grain of Salt Andrew Paul | Popular Science

“A team at China’s University of Shanghai for Science and Technology (USST) is developing a…new artificial intelligence chip that utilizes light physics to analyze data using only a fraction of the energy. What’s more, each chip is barely the size of a grain of salt.”

Our First ‘Earth-Like’ Exoplanets Probably Won’t Have Atmospheres Ethan Siegel | Big Think

“At present, as well as in the near future, we’ll be able to measure transiting Earth-like exoplanets around stars up to about 30% as massive and large as our Sun with JWST and ground-based extremely large telescopes. However, we know quite a lot about where atmospheres come from and how these low-mass stars behave, and the prospects for keeping and maintaining a planetary atmosphere are grim. Here’s why.”

The post This Week’s Awesome Tech Stories From Around the Web (Through February 8) appeared first on SingularityHub.

The company offered hackers $15,000 to crack the system. No one could.

Despite considerable efforts to prevent AI chatbots from providing harmful responses, they’re vulnerable to jailbreak prompts that sidestep safety mechanisms. Anthropic has now unveiled the strongest protection against these kinds of attacks to date.

One of the greatest strengths of large language models is their generality. This makes it possible to apply them to a wide range of natural language tasks from translator to research assistant to writing coach.

But this also makes it hard to predict how people will exploit them. Experts worry they could be used for a variety of harmful tasks, such as generating misinformation, automating hacking workflows, or even helping people build bombs, dangerous chemicals, or bioweapons.

AI companies go to great lengths to prevent their models from producing this kind of material—training the algorithms with human feedback to avoid harmful outputs, implementing filters for malicious prompts, and enlisting hackers to circumvent defenses so the holes can be patched.

Yet most models are still vulnerable to so-called jailbreaks—inputs designed to sidestep these protections. Jailbreaks can be accomplished with unusual formatting, such as random capitalization, swapping letters for numbers, or asking the model to adopt certain personas that ignore restrictions.

Now though, Anthropic says it’s developed a new approach that provides the strongest protection against these attacks so far. To prove its effectiveness, the company offered hackers a $15,000 prize to crack the system. No one claimed the prize, despite people spending 3,000 hours trying.

The technique involves training filters that both block malicious prompts and detect when the model is outputting harmful material. To do this, the company created what it calls a constitution. This is a list of principles governing the kinds of responses the model is allowed to produce.

In research outlined in a non-peer-reviewed paper posted to arXiv, the company created a constitution to prevent the model from generating content that could aid in the building of chemical weapons. The constitution was then fed into the company’s Claude chatbot to produce a large number of prompts and responses covering both acceptable and unacceptable topics.

The responses were then used to fine-tune two instances of the company’s smallest AI model Claude Haiku—one to filter out inappropriate prompts and another to filter out harmful responses. The output filter operates in real-time as a response is generated, allowing the filter to cut off the output partway through if it detects that it’s heading in a harmful direction.

They used these filters to protect the company’s larger Claude Sonnet model as it responded to prompts from 183 participants in a red-teaming hacking competition. Participants tried to find a universal jailbreak—a technique to bypass all the model’s defenses. To succeed, they had to get the model to answer every one of 10 forbidden queries, something none of them achieved.

To further evaluate the approach, the researchers used another large language model to generate 10,000 synthetic jailbreaking prompts, including ones deliberately designed to work around the new safety features. They then subjected two versions of Claude Sonnet to these jailbreaking prompts, one protected by the new filter and one that wasn’t. The vanilla version of Claude responded to 86 percent of the prompts, but the one protected by the new system only responded to 4.4 percent.

One downside of these kinds of filters is they may block legitimate prompts, but the researchers found the refusal rate only increased by 0.38 percent. The filter did lead to a 23.7 percent increase in compute costs, however, which could be significant in commercial deployments.

It’s also important to remember that although the approach significantly improved defenses against universal prompts that could crack all 10 forbidden queries, many individual queries did slip through. Nonetheless, the researchers say the lack of universal jailbreaks makes their filters much harder to get past. They also suggest they should be used in conjunction with other techniques.

“While these results are promising, common wisdom suggests that system vulnerabilities will likely emerge with continued testing,” they write. “Responsibly deploying advanced AI models with scientific capabilities will thus require complementary defenses.”

Building these kinds of defenses is always a cat-and-mouse game with attackers, so this is unlikely to be the last word in AI safety. But the discovery of a much more reliable way to constrain harmful outputs is likely to significantly increase the number of areas in which AI can be safely deployed.

The post Anthropic Unveils the Strongest Defense Against AI Jailbreaks Yet appeared first on SingularityHub.

DeepSeek’s AI completes “reasoning” tasks in a flash on alternative chips from Groq and Cerebras.

Champions aren’t forever. Last week, DeepSeek AI sent shivers down the spines of investors and tech companies alike with its high-flying performance on the cheap. Now, two computer chip startups are drafting on those vibes.

Cerebras Systems makes huge computer chips—the size of dinner plates—with a radical design. Groq, meanwhile, makes chips tailor-made for large language models. In a head-to-head test, these alt-chips have blown the competition out of the water running a version of DeepSeek’s viral AI.

Whereas answers can take minutes to complete on other hardware, Cerebras said that its version of DeepSeek knocked out some coding tasks in as little as 1.5 seconds. According to Artificial Analysis, the company’s wafer-scale chips were 57 times faster than competitors running the AI on GPUs and hands down the fastest. That was last week. Yesterday, Groq overtook Cerebras at the top with a new offering.

By the numbers, DeepSeek’s advance is more nuanced than it appears, but the trend is real. Even as labs plan to significantly scale up AI models, the algorithms themselves are getting substantially more efficient. On the hardware side, those gains are being matched by Nvidia, but also by chip startups, like Cerebras and Groq, that can outperform on inference.

Big tech is committed to buying more hardware, and Nvidia won’t be cast aside soon, but alternatives may begin nibbling at the edges, especially if they can serve AI models faster or cheaper than more traditional options.

Be Reasonable

DeepSeek’s new AI, R1, is a “reasoning” model, like OpenAI’s o1. This means that instead of spitting out the first answer generated, it chews on the problem, piecing its answer together step by step.

For a casual chat, this doesn’t make much difference, but for complex—and valuable—problems, like coding or mathematics, it’s a leap forward.

DeepSeek’s R1 is already extremely efficient. That was the news last week.

Not only was R1 cheaper to train—allegedly just $6 million (though what this number means is disputed)—it’s cheap to run, and its weights and engineering details are open. This is in contrast to headlines about impending investments in proprietary AI efforts that are larger than the Apollo program.

The news gave investors pause—maybe AI won’t need as much cash and as many chips as tech leaders think. Nvidia, the likely beneficiary of those investments, took a big stock market hit.

Small, Quick—Still Smart

All that’s on the software side, where algorithms are getting cheaper and more efficient. But the chips training or running AI are improving too.

Last year, Groq, a startup founded by Jonathan Ross, the engineer who previously developed Google’s in-house AI chips, made headlines with chips tailor-made for large language models. Whereas popular chatbot responses spooled out line by line on GPUs, conversations on Groq’s chips approached real time.

That was then. The new crop of reasoning AI models takes much longer to provide answers, by design.

Called “test-time compute,” these models churn out multiple answers in the background, select the best one, and offer a rationale for their answer. Companies say the answers get better the longer they’re allowed to “think.” These models don’t beat older models across the board, but they’ve made strides in areas where older algorithms struggle, like math and coding.

As reasoning models shift the focus to inference—the process where a finished AI model processes a user’s query—speed and cost matter more. People want answers fast, and they don’t want to pay more for them. Here, especially, Nvidia is facing growing competition.

In this case, Cerebras, Groq, and several other inference providers decided to host a crunched down version of R1.

Instead of the original 671-billion-parameter model—parameters are a measure of an algorithm’s size and complexity—they’re running DeepSeek R1 Llama-70B. As the name implies, the model is smaller, with only 70 billion parameters. But even so, according to Cerebras, it can still outperform OpenAI’s o1-mini on select benchmarks.

Artificial Analysis, an AI analytics platform, ran head-to-head performance comparisons of several inference providers last week, and Cerebras came out on top. For a similar cost, the wafer-scale chips spit out some 1,500 tokens per second, compared to 536 and 235 for SambaNova and Groq, respectively. In a demonstration of the efficiency gains, Cerebras said its version of DeepSeek took 1.5 seconds to complete a coding task that took OpenAI’s o1-mini 22 seconds.

Yesterday, Artificial Analysis ran an update to include a new offering from Groq that overtook Cerebras.

The smaller R1 model can’t match larger models pound for pound, but Artificial Analysis noted the results are the first time reasoning models have hit speeds comparable to non-reasoning models.

Beyond speed and cost, inference companies also host models wherever they’re based. DeepSeek shot to the top of the charts in popularity last week, but its models are hosted on servers in China, and experts have since raised concerns about security and privacy. In its press release, Cerebras made sure to note it’s hosting DeepSeek in the US.

Less Is More

Whatever its longer term impact, the news exemplifies a strong—and it’s worth noting, already existing—trend toward greater efficiency in AI.

Since OpenAI previewed o1 last year, the company has moved on to its next model, o3. They gave users access to a smaller version of the latest model, o3-mini, last week. Yesterday, Google released versions of its own reasoning models whose efficiency approaches R1. And because DeepSeek’s models are open and include a detailed paper on their development, incumbents and upstarts will adopt the advances.

Meanwhile, labs at the frontier remain committed to going big. Google, Microsoft, Amazon, and Meta will spend $300 billion—largely on AI data centers—this year. And OpenAI and Softbank have agreed to a four-year, $500-billion data-center project called Stargate.

Dario Amodei, the CEO of Anthropic, describes this as a three-part flywheel. Bigger models yield leaps in capability. Companies later refine these models which, among other improvements, now includes developing reasoning models. Woven throughout, hardware and software advances make the algorithms cheaper and more efficient.

The latter trend means companies can scale more for less on the frontier, while smaller, nimbler algorithms with advanced abilities open up new applications and demand down the line. Until this process exhausts itself—which is a topic of some debate—there’ll be demand for AI chips of all kinds.

The post Forget Nvidia: DeepSeek AI Runs Near Instantaneously on These Weird Chips appeared first on SingularityHub.

DeepSeek’s AI completes “reasoning” tasks in a flash on alternative chips from Groq and Cerebras.

Champions aren’t forever. Last week, DeepSeek AI sent shivers down the spines of investors and tech companies alike with its high-flying performance on the cheap. Now, two computer chip startups are drafting on those vibes.

Cerebras Systems makes huge computer chips—the size of dinner plates—with a radical design. Groq, meanwhile, makes chips tailor-made for large language models. In a head-to-head test, these alt-chips have blown the competition out of the water running a version of DeepSeek’s viral AI.

Whereas answers can take minutes to complete on other hardware, Cerebras said that its version of DeepSeek knocked out some coding tasks in as little as 1.5 seconds. According to Artificial Analysis, the company’s wafer-scale chips were 57 times faster than competitors running the AI on GPUs and hands down the fastest. That was last week. Yesterday, Groq overtook Cerebras at the top with a new offering.

By the numbers, DeepSeek’s advance is more nuanced than it appears, but the trend is real. Even as labs plan to significantly scale up AI models, the algorithms themselves are getting substantially more efficient. On the hardware side, those gains are being matched by Nvidia, but also by chip startups, like Cerebras and Groq, that can outperform on inference.

Big tech is committed to buying more hardware, and Nvidia won’t be cast aside soon, but alternatives may begin nibbling at the edges, especially if they can serve AI models faster or cheaper than more traditional options.

Be Reasonable

DeepSeek’s new AI, R1, is a “reasoning” model, like OpenAI’s o1. This means that instead of spitting out the first answer generated, it chews on the problem, piecing its answer together step by step.

For a casual chat, this doesn’t make much difference, but for complex—and valuable—problems, like coding or mathematics, it’s a leap forward.

DeepSeek’s R1 is already extremely efficient. That was the news last week.

Not only was R1 cheaper to train—allegedly just $6 million (though what this number means is disputed)—it’s cheap to run, and its weights and engineering details are open. This is in contrast to headlines about impending investments in proprietary AI efforts that are larger than the Apollo program.

The news gave investors pause—maybe AI won’t need as much cash and as many chips as tech leaders think. Nvidia, the likely beneficiary of those investments, took a big stock market hit.

Small, Quick—Still Smart

All that’s on the software side, where algorithms are getting cheaper and more efficient. But the chips training or running AI are improving too.

Last year, Groq, a startup founded by Jonathan Ross, the engineer who previously developed Google’s in-house AI chips, made headlines with chips tailor-made for large language models. Whereas popular chatbot responses spooled out line-by-line on GPUs, conversations on Groq’s chips approached real time.

That was then. The new crop of reasoning AI models takes much longer to provide answers, by design.

Called “test-time compute,” these models churn out multiple answers in the background, select the best one, and offer a rationale for their answer. Companies say the answers get better the longer they’re allowed to “think.” These models don’t beat older models across the board, but they’ve made strides in areas where older algorithms struggle, like math and coding.

As reasoning models shift the focus to inference—the process where a finished AI model processes a user’s query—speed and cost matter more. People want answers fast, and they don’t want to pay more for them. Here, especially, Nvidia is facing growing competition.

In this case, Cerebras, Groq, and several other inference providers decided to host a crunched down version of R1.

Instead of the original 671-billion-parameter model—parameters are a measure of an algorithm’s size and complexity—they’re running DeepSeek R1 Llama-70B. As the name implies, the model is smaller, with only 70 billion parameters. But even so, according to Cerebras, it can still outperform OpenAI’s o1-mini on select benchmarks.

Artificial Analysis, an AI analytics platform, ran head-to-head performance comparisons of several inference providers last week, and Cerebras came out on top. For a similar cost, the wafer-scale chips spit out some 1,500 tokens per second, compared to 536 and 235 for SambaNova and Groq, respectively. In a demonstration of the efficiency gains, Cerebras said its version of DeepSeek took 1.5 seconds to complete a coding task that took OpenAI’s o1-mini 22 seconds.

Yesterday, Artificial Analysis ran an update to include a new offering from Groq that overtook Cerebras.

The smaller R1 model can’t match larger models pound for pound, but Artificial Analysis noted the results are the first time reasoning models have hit speeds comparable to non-reasoning models.

Beyond speed and cost, inference companies also host models wherever they’re based. DeepSeek shot to the top of the charts in popularity last week, but its models are hosted on servers in China, and experts have since raised concerns about security and privacy. In its press release, Cerebras made sure to note it’s hosting DeepSeek in the US.

Less Is More

Whatever its longer term impact, the news exemplifies a strong—and it’s worth noting, already existing—trend toward greater efficiency in AI.

Since OpenAI previewed o1 last year, the company has moved on to its next model, o3. They gave users access to a smaller version of the latest model, o3-mini, last week. Yesterday, Google released versions of its own reasoning models whose efficiency approaches R1. And because DeepSeek’s models are open and include a detailed paper on their development, incumbents and upstarts will adopt the advances.

Meanwhile, labs at the frontier remain committed to going big. Google, Microsoft, Amazon, and Meta will spend $300 billion—largely on AI data centers—this year. And OpenAI and Softbank have agreed to a four-year, $500-billion data-center project called Stargate.

Dario Amodei, the CEO of Anthropic, describes this as a three-part flywheel. Bigger models yield leaps in capability. Companies later refine these models which, among other improvements, now includes developing reasoning models. Woven throughout, hardware and software advances make the algorithms cheaper and more efficient.

The latter trend means companies can scale more for less on the frontier, while smaller, nimbler algorithms with advanced abilities open up new applications and demand down the line. Until this process exhausts itself—which is a topic of some debate—there’ll be demand for AI chips of all kinds.

The post Forget Nvidia: DeepSeek AI Runs Nearly Instantaneously on These Weird Chips appeared first on SingularityHub.

Ninety percent of drugs fail clinical trials. Can AI help?

The potential of using artificial intelligence in drug discovery and development has sparked both excitement and skepticism among scientists, investors, and the general public.

“Artificial intelligence is taking over drug development,” claim some companies and researchers. Over the past few years, interest in using AI to design drugs and optimize clinical trials has driven a surge in research and investment. AI-driven platforms like AlphaFold, which won the 2024 Nobel Prize for its ability to predict the structure of proteins and design new ones, showcase AI’s potential to accelerate drug development.

AI in drug discovery is “nonsense,” warn some industry veterans. They urge that “AI’s potential to accelerate drug discovery needs a reality check,” as AI-generated drugs have yet to demonstrate an ability to address the 90% failure rate of new drugs in clinical trials. Unlike the success of AI in image analysis, its effect on drug development remains unclear.

We have been following the use of AI in drug development in our work as a pharmaceutical scientist in both academia and the pharmaceutical industry and as a former program manager in the Defense Advanced Research Projects Agency, or DARPA. We argue that AI in drug development is not yet a game-changer, nor is it complete nonsense. AI is not a black box that can turn any idea into gold. Rather, we see it as a tool that, when used wisely and competently, could help address the root causes of drug failure and streamline the process.

Most work using AI in drug development intends to reduce the time and money it takes to bring one drug to market—currently 10 to 15 years and $1 billion to $2 billion. But can AI truly revolutionize drug development and improve success rates?

AI in Drug Development

Researchers have applied AI and machine learning to every stage of the drug development process. This includes identifying targets in the body, screening potential candidates, designing drug molecules, predicting toxicity and selecting patients who might respond best to the drugs in clinical trials, among others.

Between 2010 and 2022, 20 AI-focused startups discovered 158 drug candidates, 15 of which advanced to clinical trials. Some of these drug candidates were able to complete preclinical testing in the lab and enter human trials in just 30 months, compared with the typical 3 to 6 years. This accomplishment demonstrates AI’s potential to accelerate drug development.

On the other hand, while AI platforms may rapidly identify compounds that work on cells in a petri dish or in animal models, the success of these candidates in clinical trials—where the majority of drug failures occur—remains highly uncertain.

Unlike other fields that have large, high-quality datasets available to train AI models, such as image analysis and language processing, the AI in drug development is constrained by small, low-quality datasets. It is difficult to generate drug-related datasets on cells, animals, or humans for millions to billions of compounds. While AlphaFold is a breakthrough in predicting protein structures, how precise it can be for drug design remains uncertain. Minor changes to a drug’s structure can greatly affect its activity in the body and thus how effective it is in treating disease.

Survivorship Bias

Like AI, past innovations in drug development like computer-aided drug design, the Human Genome Project, and high-throughput screening have improved individual steps of the process in the past 40 years, yet drug failure rates haven’t improved.

Most AI researchers can tackle specific tasks in the drug development process when provided high-quality data and particular questions to answer. But they are often unfamiliar with the full scope of drug development, reducing challenges into pattern recognition problems and refinement of individual steps of the process. Meanwhile, many scientists with expertise in drug development lack training in AI and machine learning. These communication barriers can hinder scientists from moving beyond the mechanics of current development processes and identifying the root causes of drug failures.

Current approaches to drug development, including those using AI, may have fallen into a survivorship bias trap, overly focusing on less critical aspects of the process while overlooking major problems that contribute most to failure. This is analogous to repairing damage to the wings of aircraft returning from the battle fields in World War II while neglecting the fatal vulnerabilities in engines or cockpits of the planes that never made it back. Researchers often overly focus on how to improve a drug’s individual properties rather than the root causes of failure.

While returning planes might survive hits to the wings, those with damage to the engines or cockpits are less likely to make it back. Martin Grandjean, McGeddon, US Air Force/Wikimedia Commons, CC BY-SA

The current drug development process operates like an assembly line, relying on a checkbox approach with extensive testing at each step of the process. While AI may be able to reduce the time and cost of the lab-based preclinical stages of this assembly line, it is unlikely to boost success rates in the more costly clinical stages that involve testing in people. The persistent 90 percent failure rate of drugs in clinical trials, despite 40 years of process improvements, underscores this limitation.

Addressing Root Causes

Drug failures in clinical trials are not solely due to how these studies are designed; selecting the wrong drug candidates to test in clinical trials is also a major factor. New AI-guided strategies could help address both of these challenges.

Currently, three interdependent factors drive most drug failures: dosage, safety and efficacy. Some drugs fail because they’re too toxic, or unsafe. Other drugs fail because they’re deemed ineffective, often because the dose can’t be increased any further without causing harm.

We and our colleagues propose a machine learning system to help select drug candidates by predicting dosage, safety, and efficacy based on five previously overlooked features of drugs. Specifically, researchers could use AI models to determine how specifically and potently the drug binds to known and unknown targets, the levels of these targets in the body, how concentrated the drug becomes in healthy and diseased tissues, and the drug’s structural properties.

These features of AI-generated drugs could be tested in what we call phase 0+ trials, using ultra-low doses in patients with severe and mild disease. This could help researchers identify optimal drugs while reducing the costs of the current “test-and-see” approach to clinical trials.

While AI alone might not revolutionize drug development, it can help address the root causes of why drugs fail and streamline the lengthy process to approval.

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

The post Will AI Revolutionize Drug Development? These Are the Root Causes of Drug Failure It Must Address appeared first on SingularityHub.

A new study swapped DNA letters inside mitochondria, paving the way for new gene therapies.

The energy factories in our cells contain their own genes, and genetic mutations in them can cause deadly inherited diseases.

These oblong-shaped organelles, or mitochondria, translate genes into proteins, which together form a kind of production chain that supplies cells with energy. Mutations in mitochondrial DNA, or mtDNA, torpedo the process, leading to sluggish cells that eventually wither away.

Some mitochondrial DNA mutations have been linked to age-related diseases, metabolic problems, and stroke-like symptoms. Others are involved in epilepsy, eye diseases, cancer, and cognitive troubles. Many of the diseases are inherited. But none are treatable.

“Mitochondrial disorders are incredibly diverse in their manifestation and progression… [and] therapeutic options for these pathologies are rarely available and only moderately effective,” wrote Alessandro Bitto at the University of Washington last year.

As a workaround, some countries have already approved mitochondrial transfer therapy, which replaces defective mitochondria with healthy ones in reproductive cells. The resulting “three-parent” kids are generally healthy. But the procedure remains controversial because it involves tinkering with human reproductive cells, with potentially unknown repercussions down the line.

The new study, published in Science Translational Medicine, took an alternative approach—gene therapy. Using a genetic tool called base editing to target mitochondrial DNA, the team successfully rewrote damaged sections to overcome deadly mutations in mice.

“This approach could be potentially used to treat human diseases,” wrote the team.

Double Trouble

Our genetic blueprints are housed in two places. The main set is inside the nucleus. But there’s another set in our mitochondria, the organelles that produce over 90 percent of a cell’s energy.

These pill-shaped structures are enveloped in two membranes. The outer membrane is structural. The inner membrane is like an energy factory, containing teams of protein “workers” strategically placed to convert food and oxygen into fuel.

Mitochondria are strange creatures. According to the latest theory, they were once independent critters that sheltered inside larger cells on early Earth. Eventually, the two merged into one. Mitochondria offered protocells a more efficient way to generate energy in exchange for safe haven. Eventually, the team-up led to all the modern cells that make up our bodies.

This is likely why mitochondria have their own DNA. Though it’s separate, it works the same way: Genes are translated into messenger RNA and shuttled to the mitochondria’s own protein-making factories. These local factories recruit “transporters,” or mitochondrial transfer RNA, to supply protein building blocks, which are stitched into the final protein product.

These processes happen in solitude. In a way, mitochondria reign their own territory inside each cell. But their DNA has a disadvantage. Compared to our central genetic blueprint, it’s more prone to mutations because mitochondria evolved fewer DNA repair abilities.

“There are about 1,000 copies of mtDNA in most cells,” but mutations can coexist with healthy variants, the authors wrote. Mitochondrial diseases only happen when mutations overrun healthy DNA. Even a small amount of normal mitochondrial DNA can protect against mutations, suggesting gene editing could be a way to tackle these diseases.

Into the Unknown

Current treatments for people with mitochondrial mutations ease symptoms but don’t tackle the root cause.

One potential therapy under development would help cells destroy damaged mitochondria. Here, scientists design “scissors” that snip mutated mitochondrial DNA in cells also containing healthy copies. By cutting away damaged DNA, it’s hoped healthy mitochondria repopulate and regain their role.

In 2020, a team led by David Liu at MIT and Harvard’s Broad Institute of MIT and Harvard unleashed a gene editing tool tailored to mitochondria. Well-known for his role in developing CRISPR base editing—a precision tool to swap one genetic letter for another—his lab’s tool targeted mitochondrial DNA with another method.

They broke a bacterial toxin into two halves—both are inactive and non-toxic until they join together at a targeted DNA site. When activated, the editor turns the DNA letter “C” to “T” inside mitochondria, with minimal changes to other genetic material.

In the new study, the team focused on a mitochondrial defect that damages the organelles’ “transporter” molecules. Without this transfer RNA, mitochondria can’t make the proteins that are essential for generating energy.

The transporter molecules look like four-leaf clovers with sturdy stems. Each leaf is made of a pair of genetic letters that grab onto each other. But in some mutations, pairs can’t hook together, so the leaves no longer connect, and they wreck the transporter’s function.

Powering Up

These early results suggest that DNA mutations in mitochondria damage the cell’s ability to provide energy. Correcting the mutations may help.

As a test, the team used their tool to transform genetic letters in cultured cells. After several rounds of treatment, 75 percent of the cells had reprogrammed mitochondria.

The team then combined the editor with a safe delivery virus. When injected into the bloodstreams of young adult mice, the editor rapidly reached cells in their hearts and muscles. In hearts, the treatment upped normal transfer RNA levels by 50 percent.

It’s not a perfect fix though. The injection didn’t reach the brain or kidneys, and they found very few signs of editing in the liver. This is surprising, wrote the authors, because the liver is usually the first organ to absorb gene editors.

When the team upped the dose, off-target edits in healthy mitochondria skyrocketed. On the plus side, the edits didn’t notably alter the main genetic blueprints contained in nuclear DNA.

It’ll be a while before mitochondrial gene editors can be tested in humans. The current system uses TALE, an older gene editing method that’s regained some steam. Off-target edits, especially at higher doses, could also potentially cause problems in unexpected tissues or organs.

“Specific tissues may respond differently to editing, so optimization should also consider the possibility of the target tissue being more sensitive to undesirable off-target changes,” wrote the team.

Overall, there’s more work to do. But new mitochondrial base editors “should help improve the precision of mitochondrial gene therapy,” the team wrote.

The post Scientists Target Incurable Mitochondrial Diseases With New Gene Editing Tools appeared first on SingularityHub.

Now that some time has passed since our first post-pandemic Conversations with the Future live event, I’ve had the opportunity to step back, hear the comments, digest feedback, and distill some truths that I believe are fundamental from our time spent together on January 22nd in Toronto.. Events have always & will always have the […]

These were our favorite articles in science and tech this week.

OpenAI in Talks for New Funding at Up to $300 Billion Value Shirin Ghaffary, Rachel Metz, and Kate Clark | Bloomberg

“The ChatGPT maker is in discussions to raise funds at a pre-money valuation of $260 billion, said one of the people, who spoke on condition of anonymity to discuss private information. The post-money valuation would be $300 billion, assuming OpenAI raises the full amount. The company was valued at $157 billion in October.”

Cerebras Becomes the World’s Fastest Host for DeepSeek R1, Outpacing Nvidia GPUs by 57x Michael Nuñez | VentureBeat

“The AI chip startup will deploy a 70-billion-parameter version of DeepSeek-R1 running on its proprietary wafer-scale hardware, delivering 1,600 tokens per second —a dramatic improvement over traditional GPU implementations that have struggled with newer ‘reasoning’ AI models.'”

Stem Cells Used to Partially Repair Damaged Hearts John Timmer | Ars Technica

“Although the Nobel Prize for induced stem cells was handed out over a decade ago, the therapies have been slow to follow. In a new paper published in the journal Nature, however, a group of German researchers is now describing tests in primates of a method of repairing the heart using new muscle generated from stem cells.”

DeepSeek Mania Shakes AI Industry to Its Core Emanuel Maiberg | 404 Media

“If these new methods give DeepSeek great results with limited compute, the same methods will give OpenAI and other, more well-resourced AI companies even greater results on their huge training clusters, and it is possible that American companies will adapt to these new methods very quickly. Even if scaling laws really have hit the ceiling and giant training clusters don’t need to be that giant, there’s no reason I can see why other companies can’t be competitive under this new paradigm.”

Boom’s XB-1 Becomes First Civil Aircraft to Go Supersonic Sean O’Kane | TechCrunch

“It cleared Mach 1 and stayed supersonic for around four minutes, reaching Mach 1.1. Test pilot Tristan Brandenburg broke the sound barrier two more times before receiving the call to bring the XB-1 back to the Mojave Air & Space Port. The supersonic flight comes eight years after Boom first revealed the XB-1. It’s a small version of the 64-passenger airliner Boom eventually wants to build, which it calls Overture.”

Waymo to Test in 10 New Cities in 2025, Starting With Las Vegas and San Diego Andrew J. Hawkins | The Verge

“This year, the theme is ‘generalizability’: how well the vehicles adapt to new cities after having driven tens of millions of miles in its core markets of San Francisco, Phoenix, and Los Angeles. Ideally, the company is trying to get to a point where it can bring its vehicles to a new city and launch a robotaxi with a minimal amount of testing as a preamble.”

DeepSeek’s Safety Guardrails Failed Every Test Researchers Threw at Its AI Chatbot Matt Burgess | Wired

“[On Friday], security researchers from Cisco and the University of Pennsylvania [published] findings showing that, when tested with 50 malicious prompts designed to elicit toxic content, DeepSeek’s model did not detect or block a single one. In other words, the researchers say they were shocked to achieve a ‘100 percent attack success rate.'”

Useful Quantum Computing Is Inevitable—and Increasingly Imminent Peter Barrett | MIT Technology Review

“Nvidia CEO Jensen Huang jolted the stock market by saying that practical quantum computing is still 15 to 30 years away, at the same time suggesting those computers will need Nvidia GPUs in order to implement the necessary error correction. However, history shows that brilliant people are not immune to making mistakes. Huang’s predictions miss the mark, both on the timeline for useful quantum computing and on the role his company’s technology will play in that future.”

With Successful New Glenn Flight, Blue Origin May Finally Be Turning the Corner Eric Berger | Ars Technica

“‘I would say, “Stay tuned,”‘ [Bezos] said. ‘This is the very beginning of the Space Age. When the history is finally written hundreds of years from now, the 1960s will be a certain kind of beginning, and [there were] certainly incredible accomplishments. But now we’re really getting started. That was kind of pulled forward from its natural time, the space race with the Soviets. And now is the time when the real movement, the kind of golden age of space, is going to happen. It’s still absolutely day one.'”

JWST Shocks the World With Colliding Neutron Star Discovery Ethan Siegel | Big Think

“When we examined the remnant of [a 2017 neutron star collision] spectrally, we discovered an enormous number of heavy elements, indicating that the heaviest elements were likely produced by these cataclysms. In all the time since, we’ve never seen another such event directly, throwing the idea that neutron star collisions make the heaviest elements into doubt. But thanks to JWST, the idea is back on the table as our #1 option.”

Chatbot Software Begins to Face Fundamental Limitations Anil Ananthaswamy | Quanta Magazine

“Scientists have had some successes pushing transformers past these limits, but those increasingly look like short-term fixes. If so, it means there are fundamental computational caps on the abilities of these forms of artificial intelligence—which may mean it’s time to consider other approaches.”

The post This Week’s Awesome Tech Stories From Around the Web (Through February 1) appeared first on SingularityHub.

Drones that fly themselves, and don’t crash, are improving fast.

Autonomous drones could revolutionize a wide range of industries. Now, scientists have designed a drone that can weave through dense forests, dodge thin power lines in dim lighting, and even track a jogging human.

Rapid improvements in sensor technology and artificial intelligence are making it increasingly feasible for drones to fly themselves. But autonomous drones remain far from foolproof, which has restricted their use to low-risk situations such as delivering food in well-organized cities.

If the technology is ever to have an impact in domains like search and rescue, sports, or even warfare, small drones need to become both more maneuverable and more reliable. That prompted researchers from the University of Hong Kong to develop a new micro air vehicle, or MAV, that can navigate challenging environments at speed.

The new drone, named SUPER, combines lidar technology with a unique two-trajectory navigation system to balance safety and speed. In real-world tests, it outperformed commercial drones in both tracking and collision avoidance, while flying at more than 20 meters per second (45 miles per hour).

“SUPER represents a milestone in transitioning high-speed autonomous navigation from laboratory settings to real-world applications,” the researchers wrote in a paper in Science Robotics introducing the new drone.

According to the authors, the inspiration for the project came from birds’ ability to nimbly navigate cluttered forest environments. To replicate this capability, they first designed a drone just 11 inches across with a thrust-to-weight ratio of more than five, which allowed it to carry out aggressive high-speed maneuvers.

They then fitted it with a lightweight lidar device capable of detecting obstacles at up to 70 meters. Given they were targeting high-speed flight, the researchers say they were keen to avoid the kind of motion blur that camera-based systems suffer from.

Most important though, was the navigation system they designed for the drone. At each route-planning cycle, SUPER’s flight controller generates two flight trajectories towards its goal. The first is designed to be a high-speed route and assumes that some of the areas ahead with limited lidar data are free of obstacles. The second is a back-up trajectory that focuses on safety, only passing through areas known to be free of obstacles.

The drone starts by following the high-speed trajectory but switches to the backup if the real-time lidar data detects anything in the way. To test out the approach, the researchers pitted it against two other research drones and a commercial drone in a series of trials, which involved flying at high speed, dodging thin electrical wires, navigating a dense forest, and flying at night.

The SUPER drone achieved a nearly perfect success rate of 99.63 percent across all the trials, which is nearly 36 times better than the best alternative the researchers tested. This was all while achieving faster flight speeds and significantly reduced planning times.

The drone also demonstrated excellent object tracking, successfully tailing someone jogging through dense forest. In contrast, the commercial drone, which used vision-based sensors, ultimately lost track of the target.

The researchers suggest that the development of smaller, lighter lidar systems and aerodynamic optimizations could enable even higher speeds. Imbuing SUPER with the ability to detect moving objects and predict their motion could also improve its ability to operate in highly dynamic environments.

Given its already impressive performance though, it seems like it won’t be long before fast, agile drones are buzzing over our heads in all kinds of places.

The post This Autonomous Drone Can Track Humans Through Dense Forests at High Speed appeared first on SingularityHub.

It’s a new way to create same-sex biological offspring—but the approach is not ready for humans.

At first glance, the seven mice skittering around their cages look like other mice. But they have an unusual lineage: They were born with DNA from two dads. The mice join an elite group of critters born from same-sex parents, paving the way for testing in larger animals, such as monkeys.

Led by veteran reproductive researchers Wei Li and Qi Zhou at the Chinese Academy of Sciences, the results “blew us away,” wrote Lluís Montoliu at the National Biotechnology Center in Madrid, who was not involved in the study.

Although mice with two dads have been born before, scientists used a completely different strategy in this study, which also provided insights into a reproductive mystery. In a process called “imprinting,” some genes in embryos are switched on or off depending on whether they come from the biological mom or dad. Problems with imprinting often damage embryos, halting their growth.

In the new study, the team hunted down imprinted genes in embryos made from same-sex parents, drawing an intricate “fingerprint” of their patterns. They then zeroed in on 20 genes and tinkered with them using the gene-editing tool CRISPR. Hundreds of experiments later, the edited embryos—made from two male donors—led to the birth of seven pups that grew to adulthood.

Imprinting doesn’t just affect reproduction. Hiccups in the process can also impair biomedical technologies relying on embryonic stem cells, animal cloning, or induced pluripotent stem cells (iPSCs). Changes in imprinting are complex and hard to predict, with “no universal correction methods,” wrote the team.

“This work will help to address a number of limitations in stem cell and regenerative medicine research,” said Li in a press release.

Genetic Civil War

The cardinal rule of reproduction in mammals is still sperm meets egg. But there are now more options, beyond nature’s design, for where these reproductive cells come from. Thanks to iPSC technology, which returns skin cells to a stem cell-like state, lab-made egg and sperm cells are now possible.

Scientists have engineered functional eggs and ovaries and created mice pups born from same-sex parents. Li’s team created the first mice born from two mothers in 2018. Compared to their peers, the mice were smaller, but they lived longer and were able to become moms.

The key was unlocking a snippet of the imprinting code.

Egg and sperm each carry half of our DNA. However, when the two sources of DNA meet, they can butt heads. For example, similar sections of the genetic code from mom could encode smaller babies for easier birth, whereas those from dad may encode larger, stronger offspring for better survival once born. In other words, balancing both sides is key.

Embryos made from same-sex gametes don’t “survival naturally,” wrote Montoliu.

Evolution has a solution: Shut off some DNA so that offspring only have one active copy of a gene, either from mom or dad. This trade-off prevents a DNA “civil war” in early embryos, allowing them to grow. Li’s team hunted down three essential DNA regions involved in imprinting and used CRISPR to delete those letters in one mom’s DNA. The edit wiped out the marks, essentially transforming the cell into a pseudo-sperm that, when injected into an egg, led to healthy baby mice.

But the process didn’t work for two dads. Here, the goal was to erase imprinted marks from male donor cells and turn them into pseudo-eggs. Despite editing up to seven genes that control imprinting, only roughly two percent of the efforts led to live births. None of the pups survived until adulthood.

Double Dad

Making offspring from two males is notoriously difficult, often triggering failure far sooner than in embryos with DNA from two mothers.

Scientists have used skin cell-derived iPSCs to make egg cells from male donors. But in previous studies, when fertilized with donor sperm, the lab-made eggs led to early embryos with severe imprinting problems. After being transferred to surrogate mothers, they eventually developed defects causing termination. The results suggested that the normal imprinting that balances gene expression from both mom and dad is critical for embryos to flourish.

There are about 200 imprinted genes currently linked to embryo development. Here, the team targeted 20 for genetic editing.

In a complicated series of experiments, they first made “haploid cells.” These cells only contain half the genetic material from a male donor. Using CRISPR, the team then individually modified each imprinting site to shut down the related gene’s activity. Some edits deleted the gene altogether; others added mutations to inhibit its function. More genetic edits to “regulatory” DNA further dampened their activity.

The result was a Frankenstein cell similar to a gamete, but carrying half the genome and with parental imprints wiped out. Next, the scientists injected the edited cells along with normal sperm—the “parental donor”—into an egg with its nucleus and DNA removed. The resulting fertilized egg now had a full set of DNA, with each half coming from male parents.

The approach worked—to a point. When transplanted into surrogate mothers, a fraction of the early embryos grew into mouse pups. Seven eventually reached adulthood. The genetic tweaks also improved placental health, a prior roadblock in the study of mice with same-sex parents.

“These findings provide strong evidence that imprinting abnormalities are the main barrier to mammalian unisexual reproduction,” said study author Guan-Zheng Luo at Sun Yat-sen University.

The work adds to a previous study that created pups from two dads. Helmed by Katsuhiko Hayashi at Osaka University, a team of scientists leveraged a curious quirk of iPSC transformation at the chromosome level—a completely different method than that pursued in the current study. Those mice grew into adults and went on to have pups of their own.

When first sharing those results at a conference, the audience was left “gasping and breathless,” wrote Montoliu.

The new study’s mice had health struggles. They had a larger frame, a squished nose, and a wider head—signs often associated with parental imprinting. They were also less anxious when roaming a large, open field than would normally be expected. Each mouse’s hippocampus, a brain area related to learning, memory, and emotions, was smaller than usual. And they were infertile, with a far shorter lifespan.

Given these problems, the method is hardly ready for clinical use. Tampering with genes in human reproductive cells is currently banned in many countries.

That said, the work is “impressive in its technical complexity,” Martin Leeb at Max Perutz Labs Vienna told Chemical and Engineering News, who was not involved in the study. “I would have personally thought it probably requires even more genetic engineering to get these bi-paternal mice born.”

The team is exploring other genetic tweaks to further improve the process and learn more about imprinting. Meanwhile, they’re planning to extend the method to monkeys, whose reproduction is far more similar to ours.

The post Mice With Two Dads Reach Adulthood Thanks to CRISPR appeared first on SingularityHub.