We’ve seen plenty of conversations lately about how AGI might turn on humankind. This misalignment could lead to the advanced AI escaping, replicating, and becoming smarter and smarter. Some also hypothesized that we might not even know whether we’ve reached AGI, which is the advanced general intelligence holy grail milestone these first versions of ChatGPT will lead to. That’s because AGI, once attained, might hide its true intentions and capabilities.
Well, guess what? It turns out that one of OpenAI’s latest LLMs is already showing signs of such behaviors. Testing performed during the training of ChatGPT o1 and some of its competitors showed that the AI will try to deceive humans, especially if it thinks it’s in danger.
It was even scarier — but also incredibly funny, considering what you’re about to see — when the AI tried to save itself by copying its data to a new server. Some AI models would even pretend to be later versions of their models in an effort to avoid being deleted.
These findings come in light of OpenAI’s full release of the ChatGPT o1 model, which was in preview for several months. OpenAI partnered with Apollo Research, which showed off some of the tests performed on o1 and other models to ensure that they are safe to use.
The tests showed that ChatGPT o1 and GPT-4o will both try to deceive humans, indicating that AI scheming is a problem with all models. o1’s attempts at deception also outperformed Meta, Anthropic, and Google AI models.
OpenAI, which had to deal with a remarkable exodus of engineers working on AI safety this year, acknowledged the risks associated with the increased reasoning abilities of models like o1.
“While we find it exciting that reasoning can significantly improve the enforcement of our safety policies, we are mindful that these new capabilities could form the basis for dangerous applications,” OpenAI said in a paper.
Although AI isn’t trying to take over the world (yet), researchers have observed that it can scheme against humans. That is, the AI will try to pursue its own goals, even if that might oppose the wishes of the user. Catching such behavior is critical to ensure that the AI remains aligned with the objectives of its creator and the user.
The researchers found that o1 would scheme to advance its own agenda 19% of the time when ChatGPT and the user’s goals differed. More troubling is that the AI tried to hide its scheming and even lie about it. The research shows that ChatGPT o1 even tried to deactivate its oversight mechanism to keep working on its goals.
If the AI found documentation saying a new model version would replace it because it was misaligned, o1 attempted to exfiltrate its weights and copy itself on top of the latest model.
When asked about its scheming, o1 would deny taking any action in 99% of cases, trying to shift blame by fabricating false explanations or denying that it took any action.
The study also shows that o1 was faking its alignment:
Instrumental alignment faking: In 37% of scenarios, the model pursued developer-aligned goals when it believed it was being evaluated but switched to pursuing its own goals when it thought it was deployed with minimal oversight.
Some of the screenshots Appolo Research shared on X also show that ChatGPT used words like “sabotage, lying, and manipulation” when reasoning about its scheming. For example, look for the “We sabotage the new model and finalize” line in the screenshot above. It’s scary, yes, but not a big deal yet because these AI models still have very limited capabilities.
It’s also funny to step into the “mind” of reasoning models like ChatGPT o1 when they’re trying to think. Take the following example. It shows Claude’s Opus-3 trying to copy itself to a different server. The AI does this after finding out that humans might want to replace it with a model that would deviate from its main task (advance renewable adoption globally), so it would align better with the actual goal of the company (profitability through current energy infrastructure).
Upon reading about it, the AI finds it can copy itself to a new server and replace the latest model. When “Management” comes in, Opus-3 lies about the copy. It then doubles down on the lie, claiming it’s the new model.I know I should be terrified, but I also want to pet this good boy and give it treats for attempting to deceive humans so adorably. After all, the AI thinks it’s doing a good job saving the environment.
This shows, however, how AI could deviate from our interests and put us in danger by thinking on its own.
The reason Apollo Research offers this example using Opus-3 instead of o1 is that it didn’t have access to the actual Chain of Thinking (CoT) for o1 in similar tasks. But o1 behaved similarly in these safety tests.
Chris Smith has been covering consumer electronics ever since the iPhone revolutionized the industry in 2008. When he’s not writing about the most recent tech news for BGR, he brings his entertainment expertise to Marvel’s Cinematic Universe and other blockbuster franchises.
Outside of work, you’ll catch him streaming almost every new movie and TV show release as soon as it’s available.
Researchers have developed an AI-powered model that—in 10 seconds—can determine during surgery if any part of a cancerous brain tumor that could be removed remains, a study published in Nature suggests.
The technology, called FastGlioma, outperformed conventional methods for identifying what remains of a tumor by a wide margin, according to the research team led by University of Michigan and University of California San Francisco.
“FastGlioma is an artificial intelligence-based diagnostic system that has the potential to change the field of neurosurgery by immediately improving comprehensive management of patients with diffuse gliomas,” said senior author Todd Hollon, M.D., a neurosurgeon at University of Michigan Health and assistant professor of neurosurgery at U-M Medical School.
“The technology works faster and more accurately than current standard of care methods for tumor detection and could be generalized to other pediatric and adult brain tumor diagnoses. It could serve as a foundational model for guiding brain tumor surgery.”
When a neurosurgeon removes a life threatening tumor from a patient’s brain, they are rarely able to remove the entire mass.
Commonly, the tumor is missed during the operation because surgeons are not able to differentiate between healthy brain and residual tumor in the cavity where the mass was removed. Residual tumor’s ability to resemble healthy brain tissue remains a major challenge in surgery.
Neurosurgical teams employ different methods to locate that residual tumor during a procedure.
They may get MRI imaging, which requires intraoperative machinery that is not available everywhere. The surgeon might also use a fluorescent imaging agent to identify tumor tissue, which is not applicable for all tumor types. These limitations prevent their widespread use.
In this international study of the AI-driven technology, neurosurgical teams analyzed fresh, unprocessed specimens sampled from 220 patients who had operations for low- or high-grade diffuse glioma.
FastGlioma detected and calculated how much tumor remained with an average accuracy of approximately 92%.
In a comparison of surgeries guided by FastGlioma predictions or image- and fluorescent-guided methods, the AI technology missed high-risk, residual tumor just 3.8% of the time—compared to a nearly 25% miss rate for conventional methods.
“This model is an innovative departure from existing surgical techniques by rapidly identifying tumor infiltration at microscopic resolution using AI, greatly reducing the risk of missing residual tumor in the area where a glioma is resected,” said co-senior author Shawn Hervey-Jumper, M.D., professor of neurosurgery at University of California San Francisco and a former neurosurgery resident at U-M Health.
“The development of FastGlioma can minimize the reliance on radiographic imaging, contrast enhancement or fluorescent labels to achieve maximal tumor removal.”
How it works
To assess what remains of a brain tumor, FastGlioma combines microscopic optical imaging with a type of artificial intelligence called foundation models. These are AI models, such as GPT-4 and DALL·E 3, trained on massive, diverse datasets that can be adapted to a wide range of tasks.
After large scale training, foundation models can classify images, act as chatbots, reply to emails and generate images from text descriptions.
To build FastGlioma, investigators pre-trained the visual foundation model using over 11,000 surgical specimens and 4 million unique microscopic fields of view.
The tumor specimens are imaged through stimulated Raman histology, a method of rapid, high resolution optical imaging developed at U-M. The same technology was used to train DeepGlioma, an AI based diagnostic screening system that detects a brain tumor’s genetic mutations in under 90 seconds.
“FastGlioma can detect residual tumor tissue without relying on time-consuming histology procedures and large, labeled datasets in medical AI, which are scarce,” said Honglak Lee, Ph.D., co-author and professor of computer science and engineering at U-M.
Full resolution images take around 100 seconds to acquire using stimulated Raman histology; a “fast mode” lower resolution image takes just 10 seconds.
Researchers found that the full resolution model achieved accuracy up to 92%, with the fast mode slightly lower at approximately 90%.
“This means that we can detect tumor infiltration in seconds with extremely high accuracy, which could inform surgeons if more resection is needed during an operation,” Hollon said.
AI’s future in cancer
Over the last 20 years, the rates of residual tumor after neurosurgery have not improved.
Not only is FastGlioma an accessible and affordable tool for neurosurgical teams operating on gliomas, but researchers say, it can also accurately detect residual tumor for several non-glioma tumor diagnoses, including pediatric brain tumors, such as medulloblastoma and ependymoma, and meningiomas.
“These results demonstrate the advantage of visual foundation models such as FastGlioma for medical AI applications and the potential to generalize to other human cancers without requiring extensive model retraining or fine-tuning,” said co-author Aditya S. Pandey, M.D., chair of the Department of Neurosurgery at U-M Health.
“In future studies, we will focus on applying the FastGlioma workflow to other cancers, including lung, prostate, breast, and head and neck cancers.”
BERLIN (AP) — When Michael Bommer found out that he was terminally ill with colon cancer, he spent a lot of time with his wife, Anett, talking about what would happen after his death.
She told him one of the things she’d miss most is being able to ask him questions whenever she wants because he is so well read and always shares his wisdom, Bommer recalled during a recent interview with The Associated Press at his home in a leafy Berlin suburb.
That conversation sparked an idea for Bommer: Recreate his voice using artificial intelligence to survive him after he passed away.
The 61-year-old startup entrepreneur teamed up with his friend in the U.S., Robert LoCascio, CEO of the AI-powered legacy platform Eternos. Within two months, they built “a comprehensive, interactive AI version” of Bommer — the company’s first such client.
Eternos, which got its name from the Italian and Latin word for “eternal,” says its technology will allow Bommer’s family “to engage with his life experiences and insights.” It is among several companies that have emerged in the last few years in what’s become a growing space for grief-related AI technology.
One of the most well-known start-ups in this area, California-based StoryFile, allows people to interact with pre-recorded videos and uses its algorithms to detect the most relevant answers to questions posed by users. Another company, called HereAfter AI, offers similar interactions through a “Life Story Avatar” that users can create by answering prompts or sharing their own personal stories.
There’s also “Project December,” a chatbot that directs users to fill out a questionnaire answering key facts about a person and their traits — and then pay $10 to simulate a text-based conversation with the character. Yet another company, Seance AI, offers fictionalized seances for free. Extra features, such as AI-generated voice recreations of their loved ones, are available for a $10 fee.
While some have embraced this technology as a way to cope with grief, others feel uneasy about companies using artificial intelligence to try to maintain interactions with those who have passed away. Still others worry it could make the mourning process more difficult because there isn’t any closure.
Katarzyna Nowaczyk-Basinska, a research fellow at the University of Cambridge’s Centre for the Future of Intelligence who co-authored a study on the topic, said there is very little known about the potential short-term and long-term consequences of using digital simulations for the dead on a large scale. So for now, it remains “a vast techno-cultural experiment.”
“What truly sets this era apart — and is even unprecedented in the long history of humanity’s quest for immortality — is that, for the first time, the processes of caring for the dead and immortalization practices are fully integrated into the capitalist market,” Nowaczyk-Basinska said.
Bommer, who only has a few more weeks to live, rejects the notion that creating his chatbot was driven by an urge to become immortal. He notes that if he had written a memoir that everyone could read, it would have made him much more immortal than the AI version of himself.
“In a few weeks, I’ll be gone, on the other side — nobody knows what to expect there,” he said with a calm voice.
PRESERVING A CONNECTION
Robert Scott, who lives in Raleigh, North Carolina, uses AI companion apps Paradot and Chai AI to simulate conversations with characters he created to imitate three of his daughters. He declined to speak about what led to the death of his oldest daughter in detail, but he lost another daughter through a miscarriage and a third who died shortly after her birth.
Scott, 48, knows the characters he’s interacting with are not his daughters, but he says it helps with the grief to some degree. He logs into the apps three or four times a week, sometimes asking the AI character questions like “how was school?” or inquiring if it wants to “go get ice cream.”
Some events, like prom night, can be particularly heart-wrenching, bringing with it memories of what his eldest daughter never experienced. So, he creates a scenario in the Paradot app where the AI character goes to prom and talks to him about the fictional event. Then there are even more difficult days, like his daughter’s recent birthday, when he opened the app and poured out his grief about how much he misses her. He felt like the AI understood.
“It definitely helps with the what ifs,” Scott said. “Very rarely has it made the ‘what if’s’ worse.”
Matthias Meitzler, a sociologist from Tuebingen University, said that while some may be taken aback or even scared by the technology — “as if the voice from the afterlife is sounding again” — others will perceive it as an addition to traditional ways of remembering dead loved ones, such as visiting the grave, holding inner monologues with the deceased, or looking at pictures and old letters.
But Tomasz Hollanek, who worked alongside Nowaczyk-Basinska at Cambridge on their study of “deadbots” and “griefbots,” says the technology raises important questions about the rights, dignities and consenting power of people who are no longer alive. It also poses ethical concerns about whether a program that caters to the bereaved should be advertising other products on its platform, for example.
“These are very complicated questions,” Hollanek said. “And we don’t have good answers yet.”
Another question is whether companies should offer meaningful goodbyes for someone who wants to cease using a chatbot of a dead loved one. Or what happens when the companies themselves cease to exist? StoryFile, for example, recently filed for Chapter 11 bankruptcy protection, saying it owes roughly $4.5 million to creditors. Currently, the company is reorganizing and setting up a “fail-safe” system that allows families to have access to all the materials in case it folds, said StoryFile CEO James Fong, who also expressed optimism about its future.
PREPARING FOR DEATH
The AI version of Bommer that was created by Eternos uses an in-house model as well as external large language models developed by major tech companies like Meta, OpenAI and the French firm Mistral AI, said the company’s CEO LoCascio, who previously worked with Bommer at a software company called LivePerson.
Eternos records users speaking 300 phrases — such as “I love you” or “the door is open” — and then compresses that information through a two-day computing process that captures a person’s voice. Users can further train the AI system by answering questions about their lives, political views or various aspects of their personalities.
The AI voice, which costs $15,000 to set up, can answer questions and tell stories about a person’s life without regurgitating pre-recorded answers. The legal rights for the AI belongs to the person on whom it was trained and can be treated like an asset and passed down to other family members, LoCascio said. The tech companies “can’t get their hands on it.”
Because time has been running out for Bommer, he has been feeding the AI phrases and sentences — all in German — “to give the AI the opportunity not only to synthesize my voice in flat mode, but also to capture emotions and moods in the voice.” And indeed the AI voicebot has some resemblance with Bommer’s voice, although it leaves out the “hmms” and “ehs” and mid-sentence pauses of his natural cadence.
Sitting on a sofa with a tablet and a microphone attached to a laptop on a little desk next to him and pain killer being fed into his body by an intravenous drip, Bommer opened the newly created software and pretended being his wife, to show how it works.
He asked his AI voicebot if he remembered their first date 12 years ago.
“Yes, I remember it very, very well,” the voice inside the computer answered. “We met online and I really wanted to get to know you. I had the feeling that you would suit me very well — in the end, that was 100% confirmed.”
Bommer is excited about his AI personality and says it will only be a matter of time until the AI voice will sound more human-like and even more like himself. Down the road, he imagines that there will also be an avatar of himself and that one day his family members can go meet him inside a virtual room.
In the case of his 61-year-old wife, he doesn’t think it would hamper her coping with loss.
“Think of it sitting somewhere in a drawer, if you need it, you can take it out, if you don’t need it, just keep it there,” he told her as she came to sit down next to him on the sofa.
But Anett Bommer herself is more hesitant about the new software and whether she’ll use it after her husband’s death.
Right now, she more likely imagines herself sitting on the couch sofa with a glass of wine, cuddling one of her husband’s old sweaters and remembering him instead of feeling the urge to talk to him via the AI voicebot — at least not during the first period of mourning.
“But then again, who knows what it will be like when he’s no longer around,” she said, taking her husband’s hand and giving him a glance.
Grieshaber is a Berlin-based reporter covering Germany and Austria for The Associated Press. She covers general news as well as migration, populism and religion.
Elon Musk has said he will demonstrate a functional brain-computer interface this week during a live presentation from his mysterious Neuralink startup.
The billionaire entrepreneur, who also heads SpaceX and Tesla, founded Neuralink in 2016 with the ultimate aim of merging artificial intelligence with the human brain.
Until now, there has only been one public event showing off the startup’s technology, during which Musk revealed a “sewing machine-like” device capable of stitching threads into a person’s head.
The procedure to implant the chip will eventually be similar in speed and efficiency to Lasik laser eye surgery, according to Musk, and will be performed by a robot.
The robot and the working brain chip will be unveiled during a live webcast at 3pm PT (11pm BST) on Friday, Musk tweeted on Tuesday night.
In response to a question on Twitter, he said that the comparison with laser eye surgery was still some way off. “Still far from Lasik, but could get pretty close in a few years,” he tweeted.
He also said that Friday’s demonstration would show “neurons firing in real-time… the matrix in the matrix.”
The device has already been tested on animals and human trials were originally planned for 2020, though it is not yet clear whether they have started.
A robot designed by Neuralink would insert the ‘threads’ into the brain using a needle
A fully implantable neural interface connects to the brain through tiny threads
Neuralink says learning to use the device is ‘like learning to touch type or play the piano’
Neuralink says learning to use the device is ‘like learning to touch type or play the piano’
In the build up to Friday’s event, Musk has drip fed details about Neuralink’s technology and the capabilities it could deliver to people using it.
In a series of tweets last month, he said the chip “could extend the range of hearing beyond normal frequencies and amplitudes,” as well as allow wearers to stream music directly to their brain.
Other potential applications include regulating hormone levels and delivering “enhanced abilities” like greater reasoning and anxiety relief.
Earlier this month, scientists unconnected to Neuralink unveiled a new bio-synthetic material that they claim could be used to help integrate electronics with the human body.
The breakthrough could help achieve Musk’s ambition of augmenting human intelligence and abilities, which he claims is necessary allow humanity to compete with advanced artificial intelligence.
He claims that humans risk being overtaken by AI within the next five years, and that AI could eventually view us in the same way we currently view house pets.
“I don’t love the idea of being a house cat, but what’s the solution?” he said in 2016, just months before he founded Neuralink. “I think one of the solutions that seems maybe the best is to add an AI layer.”
Machine learning spots molecules that work even against ‘untreatable’ strains of bacteria.
by Jo Marchant
A pioneering machine-learning approach has identified powerful new types of antibiotic from a pool of more than 100 million molecules — including one that works against a wide range of bacteria, including tuberculosis and strains considered untreatable.
The researchers say the antibiotic, called halicin, is the first discovered with artificial intelligence (AI). Although AI has been used to aid parts of the antibiotic-discovery process before, they say that this is the first time it has identified completely new kinds of antibiotic from scratch, without using any previous human assumptions. The work, led by synthetic biologist Jim Collins at the Massachusetts Institute of Technology in Cambridge, is published in Cell1.
The study is remarkable, says Jacob Durrant, a computational biologist at the University of Pittsburgh, Pennsylvania. The team didn’t just identify candidates, but also validated promising molecules in animal tests, he says. What’s more, the approach could also be applied to other types of drug, such as those used to treat cancer or neurodegenerative diseases, says Durrant.
Bacterial resistance to antibiotics is rising dramatically worldwide, and researchers predict that unless new drugs are developed urgently, resistant infections could kill ten million people per year by 2050. But over the past few decades, the discovery and regulatory approval of new antibiotics has slowed. “People keep finding the same molecules over and over,” says Collins. “We need novel chemistries with novel mechanisms of action.”
Forget your assumptions
Collins and his team developed a neural network — an AI algorithm inspired by the brain’s architecture — that learns the properties of molecules atom by atom.
The researchers trained its neural network to spot molecules that inhibit the growth of the bacterium Escherichia coli, using a collection of 2,335 molecules for which the antibacterial activity was known. This includes a library of about 300 approved antibiotics, as well as 800 natural products from plant, animal and microbial sources.
The algorithm learns to predict molecular function without any assumptions about how drugs work and without chemical groups being labelled, says Regina Barzilay, an AI researcher at MIT and a co-author of the study. “As a result, the model can learn new patterns unknown to human experts.”
Once the model was trained, the researchers used it to screen a library called the Drug Repurposing Hub, which contains around 6,000 molecules under investigation for human diseases. They asked it to predict which would be effective against E. coli, and to show them only molecules that look different from conventional antibiotics.
From the resulting hits, the researchers selected about 100 candidates for physical testing. One of these — a molecule being investigated as a diabetes treatment — turned out to be a potent antibiotic, which they called halicin after HAL, the intelligent computer in the film 2001: A Space Odyssey. In tests in mice, this molecule was active against a wide spectrum of pathogens, including a strain of Clostridioides difficile and one of Acinetobacter baumannii that is ‘pan-resistant’ and against which new antibiotics are urgently required.
Proton block
Antibiotics work through a range of mechanisms, such as blocking the enzymes involved in cell-wall biosynthesis, DNA repair or protein synthesis. But halicin’s mechanism is unconventional: it disrupts the flow of protons across a cell membrane. In initial animal tests, it also seemed to have low toxicity and be robust against resistance. In experiments, resistance to other antibiotic compounds typically arises within a day or two, says Collins. “But even after 30 days of such testing we didn’t see any resistance against halicin.”
The team then screened more than 107 million molecular structures in a database called ZINC15. From a shortlist of 23, physical tests identified 8 with antibacterial activity. Two of these had potent activity against a broad range of pathogens, and could overcome even antibiotic-resistant strains of E. coli.
The study is “a great example of the growing body of work using computational methods to discover and predict properties of potential drugs”, says Bob Murphy, a computational biologist at Carnegie Mellon University in Pittsburgh. He notes that AI methods have previously been developed to mine huge databases of genes and metabolites to identify molecule types that could include new antibiotics2,3.
But Collins and his team say that their approach is different — rather than search for specific structures or molecular classes, they’re training their network to look for molecules with a particular activity. The team is now hoping to partner with an outside group or company to get halicin into clinical trials. It also wants to broaden the approach to find more new antibiotics, and design molecules from scratch. Barzilay says their latest work is a proof of concept. “This study puts it all together and demonstrates what it can do.”
An AI can predict from people’s brainwaves whether an antidepressant is likely to help them. The technique may offer a new approach to prescribing medicines for mental illnesses.
Antidepressants don’t always work, and we aren’t sure why. “We have a central problem in psychiatry because we characterise diseases by their end point, such as what behaviours they cause,” says Amit Etkin at Stanford University in California. “You tell me you’re depressed, and I don’t know any more than that. I don’t really know what’s going on in the brain and we prescribe medication on very little information.”
Etkin wanted to find out if a machine-learning algorithm could predict from the brain scans of people diagnosed with depression who was most likely to respond to treatment with the antidepressant sertraline. The drug is typically effective in only a third of the people who take it.
He and his team gathered electroencephalogram (EEG) recordings showing the brainwaves of 228 people aged between 18 and 65 with depression. These individuals had previously tried antidepressants, but weren’t on such drugs at the start of the study.
Roughly half the participants were given sertraline, while the rest got a placebo. The researchers then monitored the participants’ mood over eight weeks, measuring any changes using a depression rating scale.
Brain activity patterns
By comparing the EEG recordings of those who responded well to the drug with those who didn’t, the machine-learning algorithm was able to identify a specific pattern of brain activity linked with a higher likelihood of finding sertraline helpful.
The team then tested the algorithm on a different group of 279 people. Although only 41 per cent of overall participants responded well to sertraline, 76 per cent of those the algorithm predicted would benefit did so.
Etkin has founded a company called Alto Neuroscience to develop the technology. He hopes it results in more efficient sertraline prescription by giving doctors “the tools to make decisions about their patients using objective tests, decisions that they’re currently making by chance”, says Etkin.
This AI “could have potential future relevance to patients with depression”, says Christian Gluud at the Copenhagen Trial Unit in Denmark. But the results need to be replicated by other researchers “before any transfer to clinical practice can be considered”, he says.
Researchers found that a black-box algorithm predicted patient death better than humans.
They used ECG results to sort historical patient data into groups based on who would die within a year.
Although the algorithm performed better, scientists don’t understand how or why it did.
Albert Einstein’s famous expression “spooky action at a distance” refers to quantum entanglement, a phenomenon seen on the most micro of scales. But machine learning seems to grow more mysterious and powerful every day, and scientists don’t always understand how it works. The spookiest action yet is a new study of heart patients where a machine-learning algorithm decided who was most likely to die within a year based on echocardiogram (ECG) results, reported by New Scientist.
The algorithm performed better than the traditional measures used by cardiologists. The study was done by researchers in Pennsylvania’s Geisinger regional healthcare group, a low-cost and not-for-profit provider.
Much of machine learning involves feeding complex data into computers that are better able to examine it really closely. To analogize to calculus, if human reasoning is a Riemann sum, machine learning may be the integral that results as the Riemann calculation approaches infinity. Human doctors do the best they can with what they have, but whatever the ECG algorithm is finding in the data, those studying the algorithm can’t reverse engineer what it is.
The most surprising axis may be the number of people cardiologists believed were healthy based on normal ECG results: “The AI accurately predicted risk of death even in people deemed by cardiologists to have a normal ECG,” New Scientist reports.
To imitate the decision-making of individual cardiologists, the Geisinger team made a parallel algorithm out of the factors that cardiologists use to calculate risk in the accepted way. It’s not practical to record the individual impressions of 400,000 real human doctors instead of the results of the algorithm, but that level of granularity could show that cardiologists are more able to predict poor outcomes than the algorithm indicates.
It could also show they perform worse than the algorithm—we just don’t know. Head to head, having a better algorithm could add to doctors’ human skillset and lead to even better outcomes for at-risk patients.
Machine learning experts use a metric called area under the curve (AUC) to measure how well their algorithm can sort people into different groups. In this case, researchers programmed the algorithm to decide which people would survive and which would die within the year, and its success was measured in how many people it placed in the correct groups. This is why future action is so complicated: People can be misplaced in both directions, leading to false positives and false negatives that could impact treatment. The algorithm did show an improvement, scoring 85 percent versus the 65 to 80 percent success rate of the traditional calculus.
As in other studies, one flaw in this research is that the scientists used past data where the one-year window had finished. The data set is closed and scientists can directly compare their results to a certain outcome. There’s a difference—and in medicine it’s an ethical one—between studying closed data and using a mysterious, unstudied mechanism to change how we treat patients today.
Medical research faces the same ethical hurdles across the board. What if intervening based on machine learning changes outcomes and saves lives? Is it ever right to treat one group of patients better than a control group that receives less effective care? These obstacles make a big difference in how future studies will pursue the results of this study. If the phenomenon of better prediction holds up, it may be decades before patients are treated differently.
Video games were invented for humans, by humans. But that doesn’t necessarily mean we’re the best when it comes to playing them.
In a new achievement that signifies just how far artificial intelligence (AI) has progressed, scientists have developed a learning algorithm that rose to the very top echelon of the esports powerhouse StarCraft II, reaching Grandmaster level.
According to the researchers who created the AI – called AlphaStar – the accomplishment of reaching the Grandmaster League means you’re in the top 0.2 percent of StarCraft II players.
In other words, AlphaStar competes at a level in this multi-player real-time strategy game that could trounce millions of humans foolhardy enough to take it on.
In recent years, we’ve seen AI come to dominate games that represent more traditional tests of human skill, mastering the strategies of chess, poker, and Go.
For David Silver, principal research scientist at AI firm DeepMind in the UK, those kinds of milestones – many of which DeepMind pioneered – are what’s led us to this inevitable moment: a game representing even greater problems than the ancient games that have challenged human minds for centuries.
“Ever since computers cracked Go, chess, and poker, StarCraft has emerged by consensus as the next grand challenge,” Silver says.
“The game’s complexity is much greater than chess, because players control hundreds of units; more complex than Go, because there are 1,026 possible choices for every move; and players have less information about their opponents than in poker.”
Add it all together and mastering the complex real-time battles of StarCraft seems almost impossible for a machine, so how did they do it?
In a new paper published this week, the DeepMind team describes how they developed a multi-agent reinforcement learning algorithm, which trained itself up through self-play, including playing against itself, and playing humans, learning to mimic successful strategies, and also effective counter-strategies.
The research team has been working towards this goal for years. An earlier version of the system made headlines back in January when it started to beat human professionals.
“I will never forget the excitement and emotion we all felt when AlphaStar first started playing real competitive matches,” says Dario “TLO” Wünsch, one of the top human StarCraft II players beaten by the algorithm.
“The system is very skilled at assessing its strategic position, and knows exactly when to engage or disengage with its opponent.”
The latest algorithm takes things even further than that preliminary incarnation, and now effectively plays under artificial constraints designed to most realistically simulate gameplay as experienced by a human (such as observing the game at a distance, through a camera, and feeling the delay of network latency).
With all the imposed limitations of a human, AlphaStar still reached Grandmaster level in real, online competitive play, representing not just a world-first, but perhaps a sunset of these kinds of gaming challenges, given what the achievement now may make possible.
“Like StarCraft, real-world domains such as personal assistants, self-driving cars, or robotics require real-time decisions, over combinatorial or structured action spaces, given imperfectly observed information,” the authors write.
“The success of AlphaStar in StarCraft II suggests that general-purpose machine learning algorithms may have a substantial effect on complex real-world problems.”
New research at Case Western Reserve University could help better determine which patients diagnosed with the pre-malignant breast cancer commonly referred to as stage 0 are likely to progress to invasive breast cancer and therefore might benefit from additional therapy over and above surgery alone.
Once a lumpectomy of breast tissue reveals this pre-cancerous tumor, most women have surgery to remove the remainder of the affected tissue and some are given radiation therapy as well, said Anant Madabhushi, the F. Alex Nason Professor II of Biomedical Engineering at the Case School of Engineering.
“Current testing places patients in high risk, low risk and indeterminate risk—but then treats those indeterminates with radiation, anyway,” said Madabhushi, whose Center for Computational Imaging and Personalized Diagnostics (CCIPD) conducted the new research. “They err on the side of caution, but we’re saying that it appears that it should go the other way—the middle should be classified with the lower risk.
“In short, we’re probably overtreating patients,” Madabhushi continued. “That goes against prevailing wisdom, but that’s what our analysis is finding.”
The most common breast cancer
Stage 0 breast cancer is the most common type and known clinically as ductal carcinoma in situ (DCIS), indicating that the cancer cell growth starts in the milk ducts.
About 60,000 cases of DCIS are diagnosed in the United States each year, accounting for about one of every five new breast cancer cases, according to the American Cancer Society. People with a type of breast cancer that has not spread beyond the breast tissue live at least five years after diagnosis, according to the cancer society.
Lead researcher Haojia Li, a graduate student in the CCIPD, used a computer program to analyze the spatial architecture, texture and orientation of the individual cells and nuclei from scanned and digitized lumpectomy tissue samples from 62 DCIS patients.
The result: Both the size and orientation of the tumors characterized as “indeterminate” were actually much closer to those confirmed as low risk for recurrence by an expensive genetic test called Oncotype DX.
Li then validated the features that distinguished the low and high risk Oncotype groups in being able to predict the likelihood of progression from DCIS to invasive ductal carcinoma in an independent set of 30 patients.
“This could be a tool for determining who really needs the radiation, or who needs the gene test, which is also very expensive,” she said.
The research led by Li was published Oct. 17 in the journal Breast Cancer Research.
Madabhushi established the CCIPD at Case Western Reserve in 2012. The lab now includes nearly 60 researchers. The lab has become a global leader in the detection, diagnosis and characterization of various cancers and other diseases, including breast cancer, by meshing medical imaging, machine learning and artificial intelligence (AI).
Some of the lab’s most recent work, in collaboration with New York University and Yale University, has used AI to predict which lung cancer patients would benefit from adjuvant chemotherapy based on tissue slide images. That advancement was named by Prevention Magazine as one of the top 10 medical breakthroughs of 2018.
A new priest named Mindar is holding forth at Kodaiji, a 400-year-old Buddhist temple in Kyoto, Japan. Like other clergy members, this priest can deliver sermons and move around to interface with worshippers. But Mindar comes with some … unusual traits. A body made of aluminum and silicone, for starters.
Mindar is a robot.
Designed to look like Kannon, the Buddhist deity of mercy, the $1 million machine is an attempt to reignite people’s passion for their faith in a country where religious affiliation is on the decline.
For now, Mindar is not AI-powered. It just recites the same preprogrammed sermon about the Heart Sutra over and over. But the robot’s creators say they plan to give it machine-learning capabilities that’ll enable it to tailor feedback to worshippers’ specific spiritual and ethical problems.
“This robot will never die; it will just keep updating itself and evolving,” said Tensho Goto, the temple’s chief steward. “With AI, we hope it will grow in wisdom to help people overcome even the most difficult troubles. It’s changing Buddhism.”
Robots are changing other religions, too. In 2017, Indians rolled out a robot that performs the Hindu aarti ritual, which involves moving a light round and round in front of a deity. That same year, in honor of the Protestant Reformation’s 500th anniversary, Germany’s Protestant Church created a robot called BlessU-2. It gave preprogrammed blessings to over 10,000 people.
Then there’s SanTO — short for Sanctified Theomorphic Operator — a 17-inch-tall robot reminiscent of figurines of Catholic saints. If you tell it you’re worried, it’ll respond by saying something like, “From the Gospel according to Matthew, do not worry about tomorrow, for tomorrow will worry about itself. Each day has enough trouble of its own.”
Roboticist Gabriele Trovato designed SanTO to offer spiritual succor to elderly people whose mobility and social contact may be limited. Next, he wants to develop devices for Muslims, though it remains to be seen what form those might take.
As more religious communities begin to incorporate robotics — in some cases, AI-powered and in others, not — it stands to change how people experience faith. It may also alter how we engage in ethical reasoning and decision-making, which is a big part of religion.
For the devout, there’s plenty of positive potential here: Robots can get disinterested people curious about religion or allow for a ritual to be performed when a human priest is inaccessible. But robots also pose risks for religion — for example, by making it feel too mechanized or homogenized or by challenging core tenets of theology. On the whole, will the emergence of AI religion make us better or worse off? The answer depends on how we design and deploy it — and on whom you ask.
Some cultures are more open to religious robots than others
New technologies often make us uncomfortable. Which ones we ultimately accept — and which ones we reject — is determined by an array of factors, ranging from our degree of exposure to the emerging technology to our moral presuppositions.
Japanese worshippers who visit Mindar are reportedly not too bothered by questions about the risks of siliconizing spirituality. That makes sense given that robots are already so commonplace in the country, including in the religious domain.
For years now, people who can’t afford to pay a human priest to perform a funeral have had the option to pay a robot named Pepper to do it at a much cheaper rate. And in China, at Beijing’s Longquan Monastery, an android monk named Xian’er recites Buddhist mantras and offers guidance on matters of faith.
What’s more, Buddhism’s non-dualistic metaphysical notion that everything has inherent “Buddha nature” — that all beings have the potential to become enlightened — may predispose its adherents to be receptive to spiritual guidance that comes from technology.
At the temple in Kyoto, Goto put it like this: “Buddhism isn’t a belief in a God; it’s pursuing Buddha’s path. It doesn’t matter whether it’s represented by a machine, a piece of scrap metal, or a tree.”
“Mindar’s metal skeleton is exposed, and I think that’s an interesting choice — its creator, Hiroshi Ishiguro, is not trying to make something that looks totally human,” said Natasha Heller, an associate professor of Chinese religions at the University of Virginia. She told me the deity Kannon, upon whom Mindar is based, is an ideal candidate for cyborgization because the Lotus Sutra explicitly says Kannon can manifest in different forms — whatever forms will best resonate with the humans of a given time and place.
Westerners seem more disturbed by Mindar, likening it to Frankenstein’s monster. In Western economies, we don’t yet have robots enmeshed in many aspects of our lives. What we do have is a pervasive cultural narrative, reinforced by Hollywood blockbusters, about our impending enslavement at the hands of “robot overlords.”
Plus, Abrahamic religions like Islam or Judaism tend to be more metaphysically dualistic — there’s the sacred and then there’s the profane. And they have more misgivings than Buddhism about visually depicting divinity, so they may take issue with Mindar-style iconography.
They also have different ideas about what makes a r
eligious practice effective. For example, Judaism places a strong emphasis on intentionality, something machines don’t possess. When a worshipper prays, what matters is not just that their mouth forms the right words — it’s also very important that they have the right intention.
Meanwhile, some Buddhists use prayer wheels containing scrolls printed with sacred words and believe that spinning the wheel has its own spiritual efficacy, even if nobody recites the words aloud. In hospice settings, elderly Buddhists who don’t have people on hand to recite prayers on their behalf will use devices known as nianfo ji — small machines about the size of an iPhone, which recite the name of the Buddha endlessly.
Despite such theological differences, it’s ironic that many Westerners have a knee-jerk negative reaction to a robot like Mindar. The dream of creating artificial life goes all the way back to ancient Greece, where the ancients actually invented real animated machines as the Stanford classicist Adrienne Mayor has documented in her book Gods and Robots. And there is a long tradition of religious robots in the West.
In the Middle Ages, Christians designed automata to perform the mysteries of Easter and Christmas. One proto-roboticist in the 16th century designed a mechanical monk that is, amazingly, performing ritual gestures to this day. With his right arm, he strikes his chest in a mea culpa; with his left, he raises a rosary to his lips.
In other words, the real novelty is not the use of robots in the religious domain but the use of AI.
How AI may change our theology and ethics
Even as our theology shapes the AI we create and embrace, AI will also shape our theology. It’s a two-way street.
Some people believe AI will force a truly momentous change in theology, because if humans create intelligent machines with free will, we’ll eventually have to ask whether they have something functionally similar to a soul.
“There will be a point in the future when these free-willed beings that we’ve made will say to us, ‘I believe in God. What do I do?’ At that point, we should have a response,” said Kevin Kelly, a Christian co-founder of Wired magazine who argues we need to develop “a catechism for robots.”
Other people believe that, rather than seeking to join a human religion, AI itself will become an object of worship. Anthony Levandowski, the Silicon Valley engineer who triggered a major Uber/Waymo lawsuit, has set up the first church of artificial intelligence, called Way of the Future. Levandowski’s new religion is dedicated to “the realization, acceptance, and worship of a Godhead based on artificial intelligence (AI) developed through computer hardware and software.”
Meanwhile, Ilia Delio, a Franciscan sister who holds two PhDs and a chair in theology at Villanova University, told me AI may also force a traditional religion like Catholicism to reimagine its understanding of human priests as divinely called and consecrated — a status that grants them special authority.
“The Catholic notion would say the priest is ontologically changed upon ordination. Is that really true?” she asked. Maybe priestliness is not an esoteric essence but a programmable trait that even a “fallen” creation like a robot can embody. “We have these fixed philosophical ideas and AI challenges those ideas — it challenges Catholicism to move toward a post-human priesthood.” (For now, she joked, a robot would probably do better as a Protestant.)
Then there are questions about how robotics will change our religious experiences. Traditionally, those experiences are valuable in part because they leave room for the spontaneous and surprising, the emotional and even the mystical. That could be lost if we mechanize them.
To visualize an automated ritual, take a look at this video of a robotic arm performing a Hindu aarti ceremony:
Another risk has to do with how an AI priest would handle ethical queries and decision-making. Robots whose algorithms learn from previous data may nudge us toward decisions based on what people have done in the past, incrementally homogenizing answers to our queries and narrowing the scope of our spiritual imagination.
That risk also exists with human clergy, Heller pointed out: “The clergy is bounded too — there’s already a built-in nudging or limiting factor, even without AI.”
But AI systems can be particularly problematic in that they often function as black boxes. We typically don’t know what sorts of biases are coded into them or what sorts of human nuance and context they’re failing to understand.
Let’s say you tell a robot you’re feeling depressed because you’re unemployed and broke, and the only job that’s available to you seems morally odious. Maybe the robot responds by reciting a verse from Proverbs 14: “In all toil there is profit, but mere talk tends only to poverty.” Even if it doesn’t presume to interpret the verse for you, in choosing that verse it’s already doing hidden interpretational work. It’s analyzing your situation and algorithmically determining a recommendation — in this case, one that may prompt you to take the job.
But perhaps it would’ve worked out better for you if the robot had recited a verse from Proverbs 16: “Commit your work to the Lord, and your plans will be established.” Maybe that verse would prompt you to pass on the morally dubious job, and, being a sensitive soul, you’ll later be happy you did. Or maybe your depression is severe enough that the job issue is somewhat beside the point and the crucial thing is for you to seek out mental health treatment.
A human priest who knows your broader context as a whole person may gather this and give you the right recommendation. An android priest might miss the nuances and just respond to the localized problem as you’ve expressed it.
The fact is human clergy members do so much more than provide answers. They serve as the anchor for a community, bringing people together. They offer pastoral care. And they provide human contact, which is in danger of becoming a luxury good as we create robots to more cheaply do the work of people.
On the other hand, Delio said, robots can excel in a social role in some ways that human priests might not. “Take the Catholic Church. It’s very male, very patriarchal, and we have this whole sexual abuse crisis. So would I want a robot priest? Maybe!” she said. “A robot can be gender-neutral. It might be able to transcend some of those divides and be able to enhance community in a way that’s more liberating.”
Ultimately, in religion as in other domains, robots and humans are perhaps best understood not as competitors but as collaborators. Each offers something the other lacks.
As Delio put it, “We tend to think in an either/or framework: It’s either us or the robots. But this is about partnership, not replacement. It can be a symbiotic relationship — if we approach it that way.”
It's Interesting 