AI Quick Feeds

MIT researchers have developed a security protocol that uses the quantum properties of light to ensure that data transmitted to and from a cloud server remains secure during deep learning computations. The protocol encodes data into laser light used in fiber optic communications systems, making it impossible for attackers to copy or intercept the information. The technique maintains a 96 percent accuracy rate while providing robust security measures.

1. Researchers at Carnegie Mellon University have developed Neural Motion Planning, an artificial intelligence network that allows robots to navigate unfamiliar environments without obstacles. The system uses data-driven simulations to train the robots to perform reactive and fast motion planning, allowing them to adapt and be more versatile.

2. In simulations, the robots encountered various household environments and random objects, such as shelves, dishwashers, and vases. The robots were able to avoid obstacles like lamps, plants, bookcases, and cabinet doors, demonstrating their ability to complete tasks in diverse household environments.

3. Neural Motion Planning provides a stepping stone in large-scale learning for robotics, especially in complex tasks like motion planning. This approach allows robots to learn and generalize from simulated environments to the real world, enabling them to navigate unknown settings successfully.

Boston-based company Neurable has developed a pair of smart headphones called the MW75 Neuro that use electroencephalography (EEG) and artificial intelligence to track the wearer's focus levels by reading their brain waves. The headphones capture 80 to 90% of the signals that traditional EEG technology can and can also play music. The device aims to improve mental wellness and prevent burnout, tracking focus levels and providing insights on how to improve work routines.

The headphones work best when the user is stationary, and users can earn "focus points" for every minute spent in high or medium focus. Neurable sets a goal of 100 focus points each day. The device can also detect when the user's focus is declining and can suggest breaks to avoid burnout. The device's brainwave data is converted into focus information and anonymized before being stored in a secure cloud database.

Cloudflare is launching a suite of tools that allow websites to monitor and selectively block AI data-scraping bots. These tools will be available to all Cloudflare customers, including the 33 million using its free services. The suite includes a real-time bot monitoring dashboard, expanded bot-blocking services, and the ability to pick and choose which bots to block or allow.

Cloudflare's bot-blocking measures will not be easily ignored by bad actors, as they compare it to having a physical wall patrolled by armed guards. The company has created processes to spot even the most carefully concealed AI crawlers.

Cloudflare is also planning to launch a marketplace for customers to negotiate scraping terms of use with AI companies, providing a way for content creators to receive value in return for their content. The company aims to ensure that humans get paid for their work and plans to facilitate licensing agreements and permissions arrangements between AI companies, publishers, and websites.

A committee of experts from top U.S. medical centers and research institutes is using NVIDIA-powered federated learning to train AI models for tumor segmentation in cancer detection.

Federated learning allows organizations to collaborate on AI model development without sharing sensitive data, addressing privacy and data management constraints.

The team is using NVIDIA FLARE and NVIDIA MONAI to optimize the training process and improve annotation accuracy, and plans to publish their methodology and pretrained model for future use.

Xavier Niel, a prominent figure in the French AI industry, has become a board member of TikTok's owner, ByteDance, as the company faces legal challenges in the US.

Niel, known for his disruptive approach, believes that Europe needs to invest in homegrown AI to compete with Asia and the US.

He has invested €200 million in French AI, launched a nonprofit research lab, and aims to develop AI infrastructure in France through his cloud provider, Scaleway.

A group of sex industry professionals and advocates is calling for inclusion in shaping AI regulations, stating that current discussions risk excluding their perspectives and overregulating their industry.

The group argues that policymakers need their insight to regulate in a way that protects fundamental rights and fosters a more sex-positive online environment, rather than risking censorship and misunderstandings.

The European Commission encourages adult industry representatives to participate in public consultations on AI regulations, including upcoming discussions on "unacceptable risks or prohibitions."

Researchers from the Chinese Academy of Sciences have developed a deep learning-based autofocus method for grayscale images.

The method uses a comprehensive dataset of grayscale image sequences and two focusing strategies to dynamically select regions of interest within the frame.

The deep learning-based autofocus method achieved fast and accurate focusing, demonstrating the potential of AI in enhancing traditional imaging technologies.

OpenAI has released a new series of AI models, called OpenAI o1-Preview, that are designed to improve reasoning and problem-solving capabilities.

These models have been trained to refine their thinking processes, try different methods, and recognize mistakes before providing an answer.

In tests, the models performed comparably to Ph.D. students in tasks related to physics, chemistry, biology, mathematics, and coding.

Apple's new iPhone 16 and 16 Pro models, with advanced AI capabilities, are expected to drive a boom in sales and start an iPhone "supercycle," according to some analysts. The AI features, such as improved photo editing, custom emoji generation, and a more natural Siri voice assistant, are seen as a major draw for customers. However, other analysts are skeptical of the supercycle prediction, noting that the AI features may not be compelling enough to drive large numbers of early upgrades.

This article provides a tutorial on creating a pronunciation assessment app.

The app's goal is to help users improve their pronunciation by allowing them to enter a word, record their voice, and receive a score based on the API's analysis.

Knowledge of JavaScript and Vue.js 3 is recommended to follow the tutorial.

AI researchers at Tsinghua University and Zhipu AI have developed a large language model (LLM) called LongWriter that can generate text output of up to 10,000 words.

Existing LLMs are typically limited to generating short outputs of around 2,000 words because they are trained on short documents. LongWriter overcomes this limitation by training on longer documents.

The researchers have made the code for LongWriter open-source, allowing others to build upon their work. They also posted a video demonstrating LongWriter generating a 10,000-word tourist guide.

Roeland Decorte, inspired by his experience in codebreaking, developed a smartphone app that uses AI to listen for signs of disease hidden in a person's pulse.

Decorte's company, Decorte Future Industries, is at the forefront of an audio-powered revolution in healthcare, using algorithms to interpret the body's faint signals and diagnose various conditions, including heart problems and stomach cancer.

By using just a microphone, Decorte's technology can provide accurate readings at home, eliminating the need for multiple apps and hardware solutions for different conditions.

Researchers have ranked AI models based on their risk level, revealing a wide range of behaviors and rule-breaking tendencies among these models.

Regulations may need to be tightened to address the legal, ethical, and regulatory compliance issues associated with AI models.

Government rules and guidelines on AI are found to be less comprehensive than companies' policies, suggesting room for improvement and the need for stricter regulations.

1. MIT researchers have developed SigLLM, a framework that uses large language models (LLMs) to detect anomalies in time-series data without the need for training or fine-tuning. LLMs have the potential to be more efficient and cost-effective than deep-learning models for anomaly detection tasks.

2. The researchers found that LLMs can convert time-series data into text-based inputs that the models can process. They developed two approaches, Prompter and Detector, that use LLMs to locate anomalous values and predict future values, respectively.

3. While LLMs did not outperform state-of-the-art deep learning models, they showed promise in anomaly detection and could be used to flag potential problems in equipment such as wind turbines or satellites before they occur. Future work will focus on improving LLM performance and understanding their performance in anomaly detection tasks.

Scientists from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a method to train robots in simulated environments created using "digital twins" of real-world spaces. Users can scan a physical environment using a smartphone app and make necessary adjustments to create a digital replica. Robots can then train in the simulated environment, which accelerates the learning process compared to training in the real world.

The approach eliminates the need for extensive reward engineering and can create strong policies for various tasks, improving performance even in environments with disturbances and distractions. The researchers tested the system's performance in controlled lab settings and real-world environments and found that it outperformed traditional imitation-learning methods, especially in situations with visual distractions or physical disruptions.

RialTo, the robot training system developed by the researchers, currently takes three days to be fully trained, but the team is working on improving the training process and the model's adaptability to new environments. The researchers presented their work at the Robotics Science and Systems (RSS) conference.

MIT researchers are using machine learning to accurately measure the atomic patterns in metals, known as short-range order (SRO), which is crucial for designing custom materials. The goal is to use SRO as a tool to tailor material properties in high-entropy alloys, which have complex compositions and superior properties. Machine learning models are used to identify chemical motifs and quantify SRO, providing a more comprehensive understanding of these materials.

Researchers from MIT and other institutions have developed a new machine-learning framework that can predict phonon dispersion relations, which are key to understanding how heat moves through materials, up to 1,000 times faster than other AI-based techniques. The method, called the virtual node graph neural network (VGNN), uses flexible virtual nodes to represent phonons in a crystal structure, allowing for more efficient and accurate predictions. This technology could help engineers design more efficient energy-conversion systems and faster microelectronic devices by reducing waste heat.

The VGNN method is capable of predicting phonon dispersion relations for a few thousand materials in just a few seconds using a personal computer, significantly improving the efficiency and speed of these calculations compared to traditional methods. The researchers propose three different versions of VGNNs with increasing complexity, which can be used to predict phonons directly from a material's atomic coordinates. The technique can also be extended to predict other high-dimensional quantities, such as optical and magnetic properties.

The VGNN method offers comparable or even better accuracy compared to other AI-based techniques, making it a promising tool for predicting thermal properties of materials. By enabling faster and more efficient predictions of phonon dispersion relations, this method could help in the design of energy generation systems and microelectronics that produce more power and

A new technique has been developed by MIT researchers to assess the reliability of foundation models, which are large pretrained deep-learning models used in AI applications. The technique involves training a set of models that are slightly different from one another and assessing the consistency of their representations of the same test data point. This technique can be used to determine if a model is reliable for a specific task, without needing to test it on real-world data.

The technique outperformed state-of-the-art baseline methods in capturing the reliability of foundation models across various classification tasks. It can also be used to rank models based on their reliability scores, allowing users to select the best model for their needs.

The researchers used an ensemble approach, training multiple models with shared properties but slight differences. They used an idea called neighborhood consistency to compare the abstract representations outputted by the models and estimate their reliability. This approach aligns the models' representation spaces by using neighboring points as anchors. The technique was found to be more consistent and robust than other methods, even with challenging test points. However, training an ensemble of large foundation models can be computationally expensive, so the researchers plan to explore more efficient methods in the future.

Researchers from MIT and the MIT-IBM Watson AI Lab have developed a technique to estimate the reliability of foundation models, which are massive deep-learning models pretrained on general-purpose, unlabeled data. The technique involves training a set of foundation models that are slightly different from one another and comparing the consistency of the representations each model learns about the same test data point. The technique outperformed state-of-the-art baseline methods in capturing the reliability of foundation models on a variety of classification tasks.

The technique can be used to assess the reliability of foundation models before they are deployed to a specific task, which is particularly useful in safety-critical situations where incorrect or misleading information could have serious consequences. It also enables users to choose the most reliable model for their task and does not require testing on a real-world dataset.

One limitation of the technique is that it requires training an ensemble of large foundation models, which is computationally expensive. Future work will focus on finding more efficient ways to build multiple models.

1. OpenAI whistleblowers have filed a complaint with the SEC, alleging that the company restricted workers from speaking out about the risks of its AI technology.

2. The whistleblowers are asking the SEC to investigate OpenAI's non-disclosure agreements and enforce rules against discouraging employees from raising concerns with regulators.

3. U.S. Senator Chuck Grassley has called for changes to OpenAI's policies and practices, stating that they have a chilling effect on whistleblowers' right to speak up.