AI News
1. SIMA
Google DeepMind unveiled SIMA, an AI agent that is being trained to play video games like a human player rather than trying to win or optimize gameplay. SIMA is learning gaming skills by playing open-world and open-ended games like No Man’s Sky, Valheim, and Goat Simulator, with the goal of eventually being able to play any video game while following instructions from human players. The AI is not intended to replace existing game AI, but rather to serve as an intelligent, collaborative partner for human players in games (The Verge).
2. WSE-3
Cerebras unveiled the WSE-3, the world’s largest semiconductor chip designed for training large AI models, doubling performance from its previous generation while maintaining the same power consumption and price. The chip, which is the size of an entire semiconductor wafer, can handle a hypothetical language model with 24 trillion parameters on a single machine, vastly exceeding current state-of-the-art models like GPT-4. Cerebras also announced a partnership with Qualcomm to optimize inference, or the deployment of trained AI models, by combining techniques like sparsity and network architecture search to reduce computational requirements and costs (ZDNet).
3. Humanoid robot
A video showcases “Figure 01,” a humanoid robot powered by OpenAI’s technology that can converse, recognize objects, and perform tasks like handing someone an apple when asked for something to eat. Developed by startup Figure with contributions from experts at companies like Boston Dynamics and Tesla, Figure 01 utilizes visual language models and onboard cameras to understand its surroundings and interact with humans in a conversational, multitasking manner. The startup’s founder aims to ultimately control billions of such robots with advanced AI systems, presenting both exciting potential and concerns about an uncannily human-like robot workforce (Mashable).
4. AI software engineer
A startup called Cognition Labs has unveiled “Devin,” claimed to be the first AI software engineer capable of coding, debugging, and deploying apps, raising concerns among software engineers about AI potentially replacing their jobs. Devin demonstrated an ability to fix real-world coding issues at a higher rate than previous AI models, with the startup positioning it as a “tireless, skilled teammate” to augment human engineers. While some engineers remain skeptical of Devin’s current capabilities, the development has reignited fears about AI disrupting a profession long considered secure from automation (Business Insider).
5. RFM-1
Covariant, a UC Berkeley AI spinout, has launched RFM-1 (Robotics Foundation Model 1), an AI platform designed to give robots “human-like” reasoning abilities by processing real-world data and determining the best course of action to execute tasks. Described as a “ChatGPT for robots,” RFM-1 allows users to input text commands, and the system uses its training data to identify objects and generate simulated video outcomes to determine the optimal way to complete the task. With its ability to understand language and the physical world, RFM-1 aims to provide a more natural method for interacting with robots, enabling them to adapt to new tasks and environments more efficiently than traditional programmed systems (TechCrunch).
6. Grok
Elon Musk’s platform xAI has launched Grok, an AI chatbot available to premium X (formerly Twitter) subscribers that aims to be an edgier, more opinionated alternative to ChatGPT and Google’s Gemini. Grok can tap into real-time data from X, provide sarcastic and rebellious responses, and Musk plans to make it open source, positioning it as a “maximum truth-seeking AI” in contrast to the perceived restraints of rivals like OpenAI. While still in early stages, Grok represents Musk’s entry into the generative AI race with a purportedly unfiltered model backed by claims of prioritizing transparency over profitability (Yahoo).
7. C-Transformer
Scientists from KAIST claim to have developed an AI chip called the “Complementary-Transformer” using 28nm Samsung manufacturing technology that can match the speed of Nvidia’s A100 GPU while being smaller and consuming far less power. This neuromorphic computing system mimics the human brain and demonstrated impressive performance running OpenAI’s GPT-2 language model much faster than a regular laptop, but questions remain about how it truly compares to leading GPU accelerators in real-world generative AI workloads. If the claimed capabilities prove valid, the low-power, compact design could enable advanced AI processing even on mobile devices (TechRadar).
Other Innovations
1. 3D printing
A new high-throughput 3D printing technique has been developed, enabling the mass production of micrometre-sized components with intricate shapes, which could significantly benefit fields like electronics and biotechnology. This innovative method, presented by Kronenfeld et al. in Nature, surpasses traditional fabrication techniques by allowing rapid production of complex 3D microparticles, offering enhanced control over their geometry (Nature).
Icon, a construction-tech startup, announced a suite of new products aimed at overcoming roadblocks in 3D printed home construction, including a multi-story 3D printer called Phoenix, a low-carbon concrete printing mix called CarbonX, a digital catalog of ready-to-print home designs called Codex, and an AI home designer called Vitruvius. These products are intended to reduce construction time, costs, and environmental impact while enabling greater design flexibility and accessibility for developers, builders, and homebuyers. With this new technology, Icon aims to accelerate the adoption of 3D printed housing as a solution to the ongoing housing crisis (Business Insider).
SpaceX’s Starship has undergone three test launches so far, with the first two resulting in explosions shortly after liftoff. In the third test, the Starship successfully reached orbit and conducted various maneuvers before being lost during atmospheric re-entry, though SpaceX still deemed the test a success as it achieved several key milestones (USA Today).
Brazil is battling a severe dengue outbreak that has infected over a million people this year by employing innovative strategies like releasing Wolbachia-infected mosquitoes that cannot transmit the virus, as well as using genetically modified and sterile male mosquitoes to suppress wild populations. One promising approach involves the World Mosquito Program breeding Wolbachia-carrying mosquitoes and releasing them to spread the bacteria, which has led to significant reductions in dengue, chikungunya, and Zika cases in areas like Niterói. While no single solution provides a quick fix, Brazil is combining multiple mosquito control methods with vaccine development in an effort to curb the alarming rise of this mosquito-borne disease exacerbated by climate change (MIT Tech Review).
Articles
1. Why are so many young people getting cancer? What the data say (Nature)
The article explores the rising incidence of cancer among young people, highlighting cases and research worldwide. Despite advancements in cancer treatment and detection, early-onset cancer rates are increasing, particularly in cancers affecting the digestive system. Researchers are investigating various factors, including genetics, lifestyle, and environmental influences, to understand this trend. They emphasize the complexity of cancer causation, suggesting that multiple factors contribute to the rise in early-onset cancer cases.
2. How to train your large language model (Economist)
The article discusses a new method called Direct Preference Optimization (DPO) for training large language models (LLMs), which is more efficient than the traditional Reinforcement Learning from Human Feedback (RLHF). DPO directly adjusts the model based on data, eliminating the need for a separate reward model, making the process three to six times more efficient. This advancement is democratizing AI, allowing smaller companies to improve their LLMs for tasks like text summarization.
