News
1. Medical LLM
BioMistral is an open-source Large Language Model (LLM) specifically tailored for the biomedical domain, built on the Mistral model and further trained on PubMed Central to address the unique challenges of adapting general-purpose LLMs to healthcare and medicine. It outperforms existing open-source medical models and shows competitive results against proprietary models, with comprehensive evaluations across 10 medical question-answering tasks in English and its first large-scale multilingual evaluation in 7 additional languages, making its datasets, benchmarks, scripts, and models freely available (Huggingface).
2. Groq
The Groq AI model is quickly becoming a significant competitor to ChatGPT, gaining popularity on social media for its rapid response times and innovative technology that potentially eliminates the need for GPUs. Developed by Groq Inc., this model operates on a custom-designed application-specific integrated circuit (ASIC) chip, enabling it to generate approximately 500 tokens per second, significantly outpacing ChatGPT-3.5’s capability (Coin Telegraph).
3. Open-source AI
Google has introduced Gemma, a new generation of lightweight, state-of-the-art open models developed for responsible AI development, leveraging the same research and technology behind the Gemini models. These models, available in two sizes (Gemma 2B and Gemma 7B), are designed to support developer innovation, foster collaboration, and ensure responsible use, complemented by a Responsible Generative AI Toolkit for safer AI applications (Google Blog).
Predibase has launched LoRA Land, a suite of 25 fine-tuned Mistral-7b models that surpass the performance of GPT-4 by 4-15% across various tasks, demonstrating a 70% improvement over base models. These models, fine-tuned for under $8 each, showcase the efficiency and cost-effectiveness of deploying specialized AI systems using Predibase’s LoRAX framework, which allows serving hundreds of models on a single GPU (PrediBase).
4. Stable Diffusion
Stable Diffusion 3 is announced as the most capable text-to-image model yet, with improvements in handling multi-subject prompts, image quality, and spelling, currently in early preview with a waitlist open for sign-ups. It introduces a range of models from 800M to 8B parameters, employing a diffusion transformer architecture and flow matching, aiming to democratize access and ensure safety and responsible AI practices (Stability AI).
5. ID tag
MIT researchers have developed a tiny, tamper-proof ID tag that uses terahertz waves to authenticate items, overcoming a major security flaw in traditional RFID tags by embedding microscopic metal particles in the adhesive to create a unique, unreplicable pattern. This innovation, which is cheaper and smaller than RFIDs, could revolutionize authentication in supply chains and is capable of attaching to very small items, with a machine-learning model ensuring over 99% accuracy in detecting tampering (MIT News)
6. CAR-T therapy
The first U.S. trials of engineered CAR T cells to treat multiple sclerosis (MS) have begun, offering hope for a new treatment option for this and other autoimmune diseases. These trials, which follow the successful use of CAR T cells in blood cancer treatment, aim to reset the immune system by targeting and destroying specific immune cells responsible for the neurodegenerative damage seen in MS, potentially leading to a significant shift in how these conditions are treated (Nature).
7. A new drug for food allergies
The FDA has approved Xolair, a medication designed to reduce severe allergic reactions to foods like milk, eggs, walnuts, and peanuts through repeated use every few weeks, rather than immediate reaction treatment. This groundbreaking drug, developed by Genentech and Novartis, aims to lessen the health impact of accidental exposure to allergens, with common side effects including fever and injection site reactions, and costs ranging from $2,900 to $5,000 a month, potentially lower with insurance (NPR).
8. Make medicine in space
Varda Space Industries, in partnership with Rocket Lab, successfully concluded its first experimental mission by landing its in-space manufacturing capsule, Winnebago-1, in the Utah desert, marking the first time a commercial company has landed a spacecraft on U.S. soil. The capsule, which spent over eight months in orbit, contained crystals of the drug ritonavir, used to treat HIV/AIDS, demonstrating the potential of manufacturing pharmaceuticals in space (TechCrunch).
Articles
1. How bacteria-fighting viruses could go mainstream (MIT Tech Review)
In 2020, Lynn Cole’s life-threatening bacterial infection, resistant to antibiotics, was initially treated successfully with an experimental virus therapy known as bacteriophage (phage) therapy, showcasing its potential as a powerful alternative to traditional antibiotics. However, the infection later returned, and despite further phage treatments, Cole ultimately passed away, highlighting both the promise and the current limitations of phage therapy in combating antibiotic-resistant infections.
2. The Shift from Models to Compound AI Systems (Berkeley AI Research Blog)
The post discusses the evolving landscape of AI development, highlighting a shift from focusing solely on large language models (LLMs) to the creation of compound AI systems that integrate multiple components for state-of-the-art results. Examples like Google’s AlphaCode 2 and AlphaGeometry demonstrate the effectiveness of these systems in programming and solving complex problems by combining LLMs with other technologies like symbolic solvers. The post argues that compound AI systems, which leverage diverse strategies and tools beyond simple model scaling, will likely be the most effective way to maximize AI capabilities and represent a significant trend in AI development for 2024.
3. Consumers: Spending more to buy less (McKinsey)
Despite a relatively strong economy and low unemployment, consumers are experiencing a decline in the volume of consumer packaged goods (CPG) they purchase, buying fewer items while spending more. This trend is pervasive across various categories, including groceries, personal care, and household products, with an average decline of 2 to 4 percent in US CPG volume, challenging CPG brands to navigate a confusing set of circumstances amid shifting consumer behaviors towards more frequent shopping but with fewer purchases per trip.
4. Mind-reading devices are revealing the brain’s secrets (Nature)
Brain-computer interfaces (BCIs) have enabled people with paralysis to control prosthetic limbs, speak through avatars, and type using only their thoughts, by translating brain activity into commands through implanted devices. This technology not only aids in restoring lost functions but also offers unprecedented insights into brain organization, challenging previous assumptions about brain anatomy and demonstrating the brain’s adaptability and the potential for long-term monitoring and understanding of neural activity.
5. A Turing Test (PNAS)
The study develops a Turing test to assess AI chatbots, specifically ChatGPT variants, by examining their behavior in behavioral games designed to elicit traits such as trust, fairness, and cooperation, and their responses to a Big-5 personality survey. The findings reveal that ChatGPT-4’s behaviors and personality traits are statistically indistinguishable from those of humans across a diverse global sample, suggesting that these AI models can exhibit human-like traits, modify their behavior based on experience, and potentially act more altruistically and cooperatively than the average human.
6. The Quest for a DNA Data Drive (IEEE Spectrum)
The article from IEEE Spectrum discusses the escalating challenge of data storage as the world produces exponentially more data, with projections indicating a significant shortfall in storage capacity by 2030. It highlights DNA as a promising solution for future data storage, offering unparalleled stability and density, potentially allowing the entirety of the internet’s information to be stored in a volume the size of a sugar cube, and outlines ongoing research and development efforts to make DNA data storage commercially viable.
