Reinforcement learning is terrible — but everything else is worse.
Karpathy’s sharpest takes yet on AGI, RL, and the future of learning.
Andrej Karpathy’s vision of AGI isn’t a bang — it’s a gradient descent through human history.
Karpathy on AGI & Superintelligence.
* AGI won’t be a sudden singularity — it will blend into centuries of steady progress (~2% GDP growth).
* Superintelligence is uncertain and likely gradual, not an instant “explosion.”
A framework for building tighter security into 5G wireless communications has been created by a Ph.D. student working with the University of Portsmouth’s Artificial Intelligence and Data Center.
With its greater network capacity and ability to rapidly transmit huge amounts of information from one device to another, 5G is a critical component of intelligent systems and services—including those for health care and financial services.
However, the dynamic nature of 5G networks, the high volumes of data shared and the ever changing types of information transmitted means that these networks are extremely vulnerable to cyber threats and increasing risks of attack.
Cybersecurity researchers are calling attention to a new campaign that delivers the Astaroth banking trojan that employs GitHub as a backbone for its operations to stay resilient in the face of infrastructure takedowns.
“Instead of relying solely on traditional command-and-control (C2) servers that can be taken down, these attackers are leveraging GitHub repositories to host malware configurations,” McAfee Labs researchers Harshil Patel and Prabudh Chakravorty said in a report.
“When law enforcement or security researchers shut down their C2 infrastructure, Astaroth simply pulls fresh configurations from GitHub and keeps running.”
In the United States, significant numbers of adults with dementia require long-term care services. For example, around 750,000 people who live in nursing homes have a diagnosis of dementia. However, transportation insecurity for this population has not received sufficient attention. Although long-term care facilities provide basic medical services, residents with dementia often need external, preventative, and follow-up care such as specialist visits, diagnostics, and dental or vision services. Without reliable nonemergency medical transportation, these needs may go unmet.
To demonstrate the extent of this problem, Postdoctoral Research Scientist Soojeong Han, Ph.D., and her colleagues reviewed existing literature on non-emergency medical transportation services for individuals with dementia living in long-term care (LTC) facilities. The study, “Transportation Services in Society for Individuals Living With Dementia in Long-Term Care Facilities: A Scoping Review,” was published in the Journal of the American Medical Directors Association.
Their review found only five publications that mentioned this topic, and even then, only briefly. Across these studies, caregivers, health care professionals, and people with dementia consistently described nonemergency medical transportation as a critical need. Reported barriers included financial strain, rural-urban disparities, lack of continuity among transportation vendors, and dementia-specific challenges such as cognitive decline, stigma, and the need for caregiver accompaniment.
Spinal cord injury (SCI) imposes a significant physical, social, and economic burden on millions of patients and their families worldwide. Although medical and surgical care improvements have decreased mortality rates, sustained recovery remains constrained. Cell-based therapies offer a promising strategy for neuroprotection and neuro-regeneration post-SCI. This article reviews the most promising preclinical approaches, encompassing the transplantation of embryonic stem cells (ESCs), mesenchymal stem cells (MSCs), neural stem cells (NSCs), oligodendrocyte progenitor cells (OPCs), Schwann cells (SCs), and olfactory ensheathing cells (OECs), along with the activation of endogenous pluripotency cell banking strategies. We also outline key ancillary strategies to enhance graft cell viability and differentiation, such as trophic factor assistance, engineered biomaterials for supportive scaffolds, and innovative methods for a synergistic effect in treatment, including promoting neuronal regeneration and reducing glial scars. We highlight the key aspects of SCI pathophysiology, the fundamental biology of cell treatments, and the advantages and limitations of each approach.
There are several approaches to treating spinal cord injuries that show great promise: Cellular therapies, which utilize a range of cells such as embryonic, neural, and mesenchymal stem cells, along with astrocytes, Schwann cells, olfactory ensheathing cells, and reprogrammed cells; The use of innovative biomaterials, including hydrogels, collagen, polycaprolactone fibers, and advanced 3D-printing technologies, provides valuable support for tissue repair.
A massive phishing campaign targeted GitHub users with cryptocurrency drainers, delivered via fake invitations to the Y Combinator (YC) W2026 program.
Y Combinator is a startup accelerator that funds and mentors projects in their early stages, and connects founders with a network of alumni and venture capital firms.
The attacker abused GitHub’s notification system to deliver the fraudulent messages, by creating issues across multiple repositories and tagging targeted users.
One often-repeated example illustrates the mind-boggling potential of quantum computing: A machine with 300 quantum bits could simultaneously store more information than the number of particles in the known universe.
Now process this: Harvard scientists just unveiled a system that was 10 times bigger and the first quantum machine able to operate continuously without restarting.
In a paper published in the journal Nature, the team demonstrated a system of more than 3,000 quantum bits (or qubits) that could run for more than two hours, surmounting a series of technical challenges and representing a significant step toward building the super computers, which could revolutionize science, medicine, finance, and other fields.
Several artificial intelligence models are now advanced enough to pass the three-part chartered financial analyst exam, even the most difficult Level III test.
Previous research, particularly from two years ago, had found AI could clear Levels I and II of the exam, but it struggled with Level III, due to the essay questions.
The new study was developed by researchers from New York University Stern School of Business and Goodfin, an AI-powered wealth-management platform.
For humans to pass the prestigious, three-part chartered financial analyst exam, it typically takes around 1,000 hours of studying over the course of several years. New research found that the technology underpinning a slew of artificial intelligence models is now advanced enough to pass even the most difficult – Level III – mock exams in a matter of minutes.
The new study – developed by researchers from New York University Stern School of Business and GoodFin, an AI-powered wealth management platform – evaluated 23 large language models on their ability to answer multiple choice and essay questions on mock CFA Level III exams. They found frontier reasoning models, including o4-mini, Gemini 2.5 Pro and Claude Opus, were able to use “chain-of-thought prompting” to successfully pass.
Previous research, particularly from two years ago, had found artificial intelligence could clear Levels I and II of the exam, but it struggled with Level III, due to the essay questions. However, the technology has evolved so rapidly that the researchers wanted to know whether the models could handle, “specialized, high-stakes analytical reasoning required for professional financial decision-making.” The third CFA exam is primarily focused on portfolio management and wealth planning.