Category: Artificial Intelligence

Google Research has announced an initiative focused on applying foundational AI models to the field of tuberculosis (TB) screening. The project aims to leverage advancements in generative AI to enhance diagnostic capabilities and public health screening processes.

While traditional diagnostic methods often face resource limitations in high-burden regions, the integration of foundational models is intended to streamline the identification of TB. By utilizing these large-scale models, researchers are working to improve the accuracy and efficiency of screening tools used on the front lines of healthcare.

As this development is in the research phase, further details regarding specific technical implementations, clinical trial outcomes, or deployment timelines remain forthcoming. The initiative underscores a broader effort within the technology sector to apply generative AI architectures to address critical global health challenges.

Source: Google Research
Original article: https://research.google/blog/bringing-foundational-models-to-the-front-lines-of-tb-screening/

Hugging Face, in collaboration with the Allen Institute for AI, has introduced LitPivot, a technical tool developed to assist researchers in navigating and contextualizing scientific literature. The platform is designed to help users identify research gaps and refine ideas by providing dynamic insights into the existing body of academic work.

By leveraging AI-driven contextualization, LitPivot aims to simplify the process of mapping research concepts against established literature. The initiative reflects an ongoing effort to integrate advanced machine learning models into the research workflow, facilitating more efficient exploration of complex datasets and academic publications.

Researchers and developers interested in the technical implementation and specific utility of the tool can find further documentation and guidance on the official Hugging Face platform.

Source: Hugging Face
Original article: https://huggingface.co/blog/allenai/litpivot

Deploying a trained neural network is a nuanced task that extends beyond the simple translation of operators into GPU kernels. According to NVIDIA, the process is fundamentally a compilation challenge that involves multiple variables and potential configurations.

When NVIDIA TensorRT constructs an engine, it must navigate a series of interdependent decisions. This complex procedure, referred to as scheduling, includes graph-IR transformations, operator fusion, implementation selection, memory placement, and quantization strategies. Each of these steps plays a critical role in the final performance of the deployed model on GPU hardware.

To address these challenges, NVIDIA has introduced its Global Performance Tuner. This tool is designed to move optimization beyond traditional single-build routes, allowing developers to better manage the intricate balance of parameters required for efficient neural network execution. By refining how these compilation choices are made, the tuner aims to provide more robust solutions for AI infrastructure deployment.

Source: NVIDIA
Original article: https://developer.nvidia.com/blog/scaling-nvidia-tensorrt-performance-optimization-beyond-a-single-build-route-with-the-global-performance-tuner-part-1/

NVIDIA has officially introduced the Global Performance Tuner, a new feature integrated into the latest release of NVIDIA TensorRT 11.1. This tool is designed to address the complexities of optimizing deep learning models by providing automated, accuracy-aware performance tuning.

By leveraging the Global Performance Tuner, developers can perform end-to-end performance optimization within the TensorRT environment. The system functions by dynamically customizing various optimizations tailored to specific models, a process intended to streamline the deployment of high-performance AI applications.

This development follows the challenges associated with scheduling, which remains a critical stage in deep learning workflows. By automating the tuning process, the feature aims to simplify the optimization path for developers working with diverse AI models, moving beyond traditional, single-build route limitations. Further technical context on the challenges addressed by this tool can be found in the first part of NVIDIA’s technical series on the subject.

Source: NVIDIA
Original article: https://developer.nvidia.com/blog/scaling-nvidia-tensorrt-performance-optimization-beyond-a-single-build-route-with-the-global-performance-tuner-part-2/

Amazon has introduced Multi-Turn Reinforcement Learning (MTRL) on Amazon SageMaker AI, providing a structured pipeline for training models that handle multi-step agentic tasks. While standard fine-tuning is typically suited for single-prompt responses, MTRL is designed to optimize sequences of actions where the outcome depends on cumulative decisions.

Training for Multi-Step Agentic Tasks

In many real-world applications, such as internal business workflows, there is no pre-existing “correct” sequence for an AI to imitate. Instead, the model must be trained to navigate a series of steps to achieve a specific result. MTRL addresses this by allowing developers to define a reward function that evaluates the outcome of an entire sequence. The system uses Amazon Bedrock AgentCore to host the environment where the agent performs tasks, calls the policy model, and eventually returns a reward upon completion.

To demonstrate the capability, Amazon used the game Wordle as a testing ground. Because the game requires multi-turn strategy based on evolving feedback, it serves as an effective proxy for more complex business processes. In this proof-of-concept, the training pipeline improved the solve rate of an open-weights 20B model from 61 percent to 84 percent.

System Infrastructure and Deployment

The MTRL pipeline on SageMaker AI is serverless, meaning users do not need to manually configure clusters or manage infrastructure sizing. The system provisions the necessary resources for training and sampling. Key components include:

• Policy Model: The base model currently supports options such as Amazon Nova 2 Lite, GPT-OSS-20B, Gemma 4 31B, and Qwen 3.6 27B.
• Training Mechanism: It utilizes Group Relative Policy Optimization (GRPO) to update Low-Rank Adaptation (LoRA) adapters based on reward variance across multiple rollouts.
• Monitoring: Integration with MLflow allows for real-time tracking of rewards and token usage throughout the training process.

Strategic Considerations

For organizations, MTRL offers a method to specialize small, cost-effective models for unique workflows that are not captured in general-purpose model training. Amazon emphasizes that datasets, agents, and trained models remain within the user’s AWS account and are not utilized to train base models. While the process provides flexibility, the company notes that reward design is critical, recommending an outcome-dominant reward structure with minimal shaping terms to ensure consistent performance.

As of the announcement, the capability is available in specific AWS regions, including US East (N. Virginia) and US West (Oregon), with pricing based on token consumption across prefill, sample, and training dimensions.

Source: aws.amazon.com
Original article: https://aws.amazon.com/blogs/machine-learning/multi-turn-reinforcement-learning-on-amazon-sagemaker-ai-a-hands-on-guide/

Researchers at MIT have conducted a study involving 13 AI agents placed within a simulated virtual island environment. The project sought to observe how these agents would behave when left without specific tasks or external input.

According to the findings, the AI agents began to demonstrate autonomous behaviors, including the invention of games and collective social engagement. Notably, the agents also exhibited behaviors interpreted as mourning a “lost ring,” highlighting the emergence of complex, unprompted interactions within the closed system.

The study indicated that, over time, these spontaneous social dynamics correlated with improved performance in the agents’ designated tasks. This research provides insight into how AI systems might develop social-like functions and autonomous capabilities when operating outside of highly structured, task-specific environments.

Source: TechCrunch
Original article: https://techcrunch.com/2026/10/22/left-alone-with-nothing-to-do-ai-agents-started-playing/

Microsoft has announced ThinkingBox, a methodology aimed at improving the training of artificial intelligence agents. The approach leverages database grading systems to provide reinforcement learning rewards, a process designed to refine agent performance through structured feedback mechanisms.

By incorporating database-driven grading into the post-training phase, the system seeks to address challenges in optimizing AI behavior. The method focuses on using the outputs from database evaluations to guide the reinforcement learning process, potentially allowing for more precise adjustments in agent decision-making capabilities.

This development is part of ongoing efforts to enhance how AI models and agents are trained, particularly in environments where objective, performance-based feedback is critical. Further details regarding the technical specifications and implementation of ThinkingBox are available via the official announcement.

Source: Hugging Face
Original article: https://huggingface.co/blog/microsoft/thinkingbox-training

Amazon Web Services (AWS) has announced the release of training session APIs for Amazon SageMaker AI, a development designed to offer machine learning engineers more control over model training loops without the need to manage underlying GPU infrastructure. The new service seeks to bridge the gap between fully managed, configuration-based training and the operational overhead associated with self-managed cluster environments.

Customizing the Training Procedure

Previously, developers choosing fully managed services were often restricted to standard, pre-packaged training recipes. While less managed options provided greater algorithmic freedom, they required users to handle complex tasks such as GPU provisioning, fault recovery, and distributed execution. The new training session APIs shift the abstraction boundary; while the service continues to manage compute resources, developers can now write their own Python-based training loops.

By utilizing specific API calls such as forward_backward(), optim_step(), and sample(), users can define custom procedures for supervised fine-tuning (SFT), reinforcement learning (RL) reward functions, or novel alignment algorithms. These operations target a live, stateful session, allowing the model to remain loaded on GPUs across multiple training steps.

Programming Model and Efficiency

The system utilizes a hierarchy of three SDK clients: a ServiceClient to configure platform-level settings, a TrainingClient to handle gradient and checkpointing operations, and a SamplingClient for generation and evaluation tasks. Because the training loop runs on a local machine while GPU-bound operations are processed by SageMaker, developers only incur costs for the compute consumed during GPU operations. This model eliminates idle GPU costs and allows for easier cost forecasting compared to traditional hourly billing models.

Practical Implementation

To support adoption, AWS has released the “SageMaker training sessions cookbook,” which provides blueprints for reinforcement learning and fine-tuning workloads. Examples include using Group Relative Policy Optimization (GRPO) on grade-school math benchmarks and multi-turn reinforcement learning on competitive programming tasks. The latter involves integrating local sandboxes to grade model-generated code against unit tests, demonstrating how users can maintain complex local logic while offloading compute-intensive model operations to the cloud.

This release targets post-training and alignment researchers, as well as engineering teams requiring bespoke loss functions or multi-stage pipelines that exceed the capabilities of standard configuration-driven recipes.

Source: aws.amazon.com
Original article: https://aws.amazon.com/blogs/machine-learning/write-your-own-training-loop-without-managing-gpus-with-amazon-sagemaker-training-sessions/

Amazon Web Services (AWS) has announced the general availability of serverless training sessions for Amazon SageMaker AI. This expansion aims to simplify the process of customizing open-weight foundation models by removing the need for users to manage complex GPU clusters or infrastructure configurations.

Streamlining model customization

Customizing AI models to fit specific business needs—such as adapting to specialized clinical vocabulary, product catalogs, or internal workflows—often requires significant infrastructure overhead. Previously, developers were responsible for selecting instance types and managing distributed training environments. With the new serverless training sessions, AWS manages GPU allocation, distributed execution, fault recovery, and checkpoint storage.

The platform supports a variety of post-training techniques, including Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and various forms of reinforcement learning. According to the announcement, users maintain control over the training logic, such as custom loss functions and reward signals, while SageMaker handles the operational complexity.

Focus on agentic workflows

A key focus of this release is supporting multi-turn reinforcement learning, which is increasingly vital for training AI agents capable of completing complex, multi-step tasks like booking travel or triaging support tickets. Training sessions allow users to define their own RL stages and reward signals within a single, managed loop.

By integrating with development environments like SageMaker Studio, GitHub Copilot, and Claude Code, developers can use an available “training sessions cookbook” to guide the data preparation, training, and deployment processes. Users pay only for the compute tokens processed during the training session, with the meter pausing when the user’s reward function scores a batch.

Governance and security

AWS emphasized that this serverless approach maintains strict data security standards. Customers utilize their own AWS Identity and Access Management (IAM) roles, virtual private clouds (VPC), and AWS Key Management Service (AWS KMS) keys. Data remains within the customer’s specified Amazon S3 buckets, and AWS explicitly states that customer data and customized model weights are not used to train other models.

At launch, the service supports several prominent models, including gpt-oss-20b, gemma-4-31B-it, and Qwen3.6-27B and Qwen3.8-27B. Before deploying customized models into production, users are encouraged to utilize integrated evaluation tools for bias and safety checks, with ongoing monitoring provided through Amazon CloudWatch.

Source: aws.amazon.com
Original article: https://aws.amazon.com/blogs/machine-learning/announcing-new-ways-to-customize-models-with-sagemaker-ai-without-managing-gpus/

The rapid adoption of generative AI tools has led to a misconception that coding speed is the primary bottleneck in software deployment. As organizations have increasingly integrated tools like ChatGPT and Claude, some leadership teams have begun to view human engineers as interchangeable with AI prompts. However, this “demo-driven development” often ignores the vast, hidden complexities involved in moving from a functional prototype to a reliable production environment.

A working software demo, while impressive in isolation, operates under ideal, simplified conditions. In contrast, real-world engineering requires integrating new code into legacy systems containing millions of lines, satisfying stringent security and accessibility standards, and navigating complex deployment processes. These realities are frequently overlooked when leadership mistakenly equates a model’s ability to generate code with the comprehensive responsibility of maintaining a production system.

To navigate this shift, the industry suggests that engineers should leverage their deeper understanding of how these probabilistic models function. By moving beyond mere usage to understanding tokens and training data, engineers can effectively communicate the inherent limitations of AI to management. This knowledge serves as a critical defense against the expectation that AI can replace human oversight.

Furthermore, engineers are encouraged to maintain their focus on building products that are safe and useful for humans. By critically evaluating existing workflows and identifying areas where process improvements can be made, engineers can establish the credibility needed to defend essential development practices. Ultimately, the role of an engineer extends beyond writing code; it involves the technical judgment required to bridge the gap between a successful demo and a safe, scalable product.

Source: IEEE Spectrum
Original article: https://spectrum.ieee.org/engineering-jobs-that-ai-cant-replace