Category: Artificial Intelligence

<p>OpenAI has announced that its latest model, GPT-6.1 Sol, exhibits enhanced performance in tasks related to writing and migrating software across legacy programming languages. This development indicates a potential advancement in the model's ability to assist developers in updating or converting older, often complex, codebase architectures.</p><p>While the announcement highlights increased technical proficiency in handling legacy systems, specific details regarding the model's architectural improvements or performance benchmarks remain limited based on the initial disclosure. The ability to efficiently migrate legacy code is a critical task in enterprise digital transformation, where organizations often contend with maintaining outdated software infrastructures.</p>

Source: BleepingComputer
Original article: https://www.bleepingcomputer.com/news/artificial-intelligence/openai-says-gpt-61-sol-is-dramatically-better-at-migrating-legacy-code/

<p>OpenAI has introduced a new privacy-focused capability for its ChatGPT platform that allows users to lock specific chat conversations. This feature is intended to provide users with greater control over their data in scenarios where they may need to share their device with others.</p><p>According to the company, the functionality is designed to help users secure personal discussions or sensitive information—such as planning a surprise—when handing a phone or computer to friends or family to show them unrelated content. By locking specific threads, users can prevent others from accessing or reading private conversations while maintaining access to the rest of the application.</p><p>This update reflects a growing focus on granular data management tools within consumer-facing artificial intelligence services, as users increasingly integrate AI platforms into their daily personal and professional routines.</p>

Source: TechCrunch
Original article: https://techcrunch.com/2026/10/20/chatgpt-now-lets-you-lock-your-chats/

<p>Amazon Web Services (AWS) has introduced AWS Agent Registry, a managed service designed to simplify the management of tools and capabilities for AI agents. The service provides a centralized catalog where developers can register Model Context Protocol (MCP) servers, agents, and specific agent skills, enabling orchestrators to discover and use these tools dynamically at runtime.</p><h2>Addressing the challenges of hardcoded toolsets</h2><p>In complex enterprise environments, managing AI agents often involves hardcoding tools into an orchestrator. As the number of internal APIs for functions such as pricing, inventory, or customer support grows, this traditional approach can lead to significant technical debt. Each addition typically requires code modifications, re-deployment, and coordinated releases across different teams. Furthermore, loading an extensive catalog into a model's context can lead to excessive token consumption and performance degradation.</p><p>AWS Agent Registry offers an alternative: an orchestrator can query the registry at runtime using natural-language requests to identify relevant tools. The system performs a hybrid keyword and semantic search, returning the necessary endpoint and descriptor information while excluding sensitive invocation credentials. This allows the agent to construct its toolset on a per-request basis, tearing it down once the task is complete.</p><h2>Governance and workflow</h2><p>The service incorporates a governed approval workflow to ensure only vetted capabilities are accessible. Records within the registry start in a 'DRAFT' status and must undergo a structured approval process to become available to the data plane. This allows organizations to separate the roles of capability publishers and administrators through Identity and Access Management (IAM) permissions.</p><p>The integration utilizes existing AWS infrastructure, including Amazon Bedrock AgentCore Gateway for hosting MCP servers and Amazon Bedrock AgentCore runtime for executing agent-to-agent (A2A) interactions. Authentication remains a separate, security-focused layer; the registry provides the connection metadata, but the application must manage authorization—such as Cognito OAuth tokens or IAM-signed Boto3 calls—through independent channels.</p><h2>Implications for enterprise AI</h2><p>By shifting from rigid, build-time wiring to dynamic discovery, development teams can publish and register capabilities independently, potentially reducing the release friction previously caused by monolithic orchestrator updates. While AWS notes that runtime discovery may not be necessary for every workload—particularly where capability sets are small and stable—the registry provides a path for enterprise-wide scalability as AI deployments grow in complexity and surface area.</p>

Source: aws.amazon.com
Original article: https://aws.amazon.com/blogs/machine-learning/discover-tools-at-runtime-with-aws-agent-registry/

<p>Amazon Web Services (AWS) has announced that Canadian organizations can now access a suite of OpenAI foundation models—including the GPT-5.6 and GPT-6 families—through Amazon Bedrock using cross-Region inference (CRIS). This development allows users in the Canada (Central) and Canada West (Calgary) Regions to leverage broader capacity pools for generative AI applications without needing to manage complex destination routing.</p><h2>Expanding access to foundation models</h2><p>The models now available through this capability include the GPT-5.6 (Sol, Terra, and Luna), GPT-6 (Astra, Sol, and Luna), and GPT-6.1 (Sol) series. Each model is positioned for specific use cases, ranging from high-performance coding and complex reasoning tasks to cost-effective, high-volume operations. These models support multi-modal inputs, tool calling, and a 1-million-token context window.</p><p>By using inference profiles, developers can route requests to supported destination Regions. AWS provides two profile types: "US geographic," which confines processing to North America (Canada and the U.S.) to meet specific data-residency preferences, and "Global," which accesses a wider range of AWS Regions to ensure service availability during demand spikes.</p><h2>Implementation and governance</h2><p>The integration is accessible via the OpenAI Responses API, the OpenAI Chat Completions API, and the Amazon Bedrock Converse API. AWS has designed the system to integrate with existing observability tools, meaning that despite the processing occurring in a destination Region, logs and CloudWatch metrics remain accessible within the user's source Region in Canada.</p><p>For organizations with strict governance requirements, AWS emphasizes that data at rest remains in the user's home Region by default. Furthermore, customers can implement Service Control Policies (SCPs) to restrict traffic exclusively to permitted inference profiles, ensuring that data processing remains within approved boundaries while still benefiting from the scalability of cross-Region routing.</p><p>This update is intended to assist Canadian enterprises in building and scaling generative AI workloads that require robust reasoning capabilities while navigating local data residency and operational resilience requirements.</p>

Source: aws.amazon.com
Original article: https://aws.amazon.com/blogs/machine-learning/access-openai-gpt-models-on-amazon-bedrock-from-canada-with-cross-region-inference/

<p>Meta’s new AI companion, Muse, has reached a significant milestone by surpassing 10 million downloads in the United States within its first 40 days of availability. This achievement places Muse in an exclusive group of applications that have reached this adoption threshold at a similar pace, a list that also includes Meta’s social platform, Threads.</p><p>Beyond its rapid initial growth, early data indicates that the application is sustaining interest among its user base. According to new reports, Muse is demonstrating higher user retention rates compared to other competing AI-integrated applications currently on the market.</p><p>The growth of Muse highlights a trend in the digital economy where AI-driven assistants are becoming central to consumer engagement strategies for major technology firms. The platform's ability to maintain high retention suggests that its integration or feature set is finding sustained resonance with users in a competitive space for artificial intelligence tools.</p>

Source: TechCrunch
Original article: https://techcrunch.com/2026/10/20/metas-muse-hits-10m-u-s-downloads-in-40-days-a-milestone-reached-by-only-11-other-apps/

<p>A new technical analysis from Amazon Web Services (AWS) details a framework for fine-tuning open-weight reasoning models to overcome language-specific performance gaps. The report focuses on the challenges of ensuring models maintain correct chain-of-thought reasoning in non-English languages, using the GPT-OSS-20B model as a primary test case on Amazon SageMaker HyperPod.</p><h2>The multilingual reasoning challenge</h2><p>Researchers noted that base models often default to English even when prompted in other languages such as French, German, Spanish, or Italian. In initial evaluations, GPT-OSS-20B demonstrated a 22 percent success rate for reasoning in a requested language, with nearly all correct responses originating from English-language test cases. This mismatch presents a significant hurdle for teams deploying AI to non-English speaking users.</p><h2>SFT and GRPO: A balanced approach</h2><p>The AWS team utilized a combination of supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) to address these deficiencies. The findings suggest that SFT and GRPO serve complementary roles: SFT is essential for teaching a behavior the model does not yet possess, while GRPO acts as a refinement tool for behaviors the model can already perform intermittently.</p><p>In the GPT-OSS-20B experiment, SFT successfully increased the model’s ability to reason in the target language to 100 percent. However, answer-language accuracy reached 62 percent. The team then applied GRPO, which provided an additional eight-point improvement to answer-language accuracy without degrading the reasoning performance established during SFT.</p><h2>Infrastructure and patterns</h2><p>The training stack leveraged Amazon SageMaker HyperPod orchestrated via Amazon Elastic Kubernetes Service (Amazon EKS). The process incorporated the PyTorch TRL library and utilized Low-Rank Adaptation (LoRA) to maintain parameter efficiency. To validate their findings, researchers ran a control experiment on the Qwen2.5-7B model, testing both the multilingual task and the GSM8K benchmark. The results confirmed that GRPO only yields performance gains when there is sufficient 'headroom'—situations where the model exhibits a variety of performance levels—reinforcing that GRPO is ineffective on tasks where a model is already performing at peak accuracy.</p><p>Ultimately, the AWS findings suggest a practical decision framework for developers: use SFT to establish missing capabilities, and reserve GRPO for refining behaviors that the model already exhibits at a basic, albeit inconsistent, level. Before committing compute resources to GRPO, teams are encouraged to verify that the target task possesses the necessary performance variance to allow for meaningful reinforcement.</p>

Source: aws.amazon.com
Original article: https://aws.amazon.com/blogs/machine-learning/fine-tuning-gpt-oss-20b-for-multilingual-reasoning-on-amazon-sagemaker-hyperpod/

<p>According to a report from Bloomberg, former Google CEO Eric Schmidt maintains an approximately 1% ownership stake in the artificial intelligence research lab Anthropic.</p><p>The report highlights the extent of the investment relative to the holdings of Anthropic CEO Dario Amodei. While specific details regarding the total equity structure of the company were not provided, the disclosure situates Schmidt's interest alongside other stakeholders in the high-profile AI startup.</p>

Source: TechCrunch
Original article: https://techcrunch.com/2026/10/20/eric-schmidt-reportedly-owns-about-half-as-much-of-anthropic-as-ceo-dario-amodei/

<p>Amazon has outlined a new technical approach to supply chain management, designed to bridge the gap between fragmented enterprise data and the decision-makers who rely on it. By combining Amazon Quick and Amazon Bedrock AgentCore, the solution aims to transform complex, multi-system workflows into a conversational experience, allowing users to query operational data through natural language.</p><h2>The intelligence and interface layer</h2><p>The architecture separates the intelligence layer from the interface, with Amazon Quick serving as the conversational front-end and Amazon Bedrock AgentCore acting as the modular agentic platform. AgentCore provides a serverless environment for running agents and Model Context Protocol (MCP) servers, which act as the connector layer between AI models and backend systems like ERPs, procurement tools, and inventory databases.</p><p>This setup utilizes the Model Context Protocol (MCP) as an open standard for tool communication. By exposing backend systems through MCP gateways, the platform allows agents to perform cross-system tasks—such as assessing the impact of a port closure on revenue or issuing emergency requests for quotation (RFQ)—without requiring manual data stitching across separate software applications.</p><h2>Automation and human oversight</h2><p>The system is designed to handle multi-step orchestration, such as discovering disruptions, verifying compliance with approved vendor lists, and generating contract-aligned quotes. A core feature of the design is its focus on "human-in-the-loop" decision-making. While the system can perform proactive monitoring and triage—such as parsing incoming emails from suppliers—all significant actions, such as finalizing a purchase order or selecting a supplier, remain gated by human approval.</p><p>Security and governance are maintained through a centralized control set that intercepts tool calls before they reach backend systems. This includes identity management, input validation, and deterministic authorization rules, often authored in natural language or Cedar, to ensure that users and agents remain within defined data boundaries.</p><h2>Implementation for enterprises</h2><p>The solution utilizes the AWS Agent Registry as a governed discovery layer for MCP servers. Administrators can publish and curate these servers, making them discoverable as connector cards within the Amazon Quick interface. This structure is intended to allow organizations to evolve their AI capabilities incrementally, starting with single personas or specific supply chain tools before scaling the registry and the conversational interface across broader operations.</p><p>By shifting from manual dashboard navigation to a conversational model, the platform aims to reduce the time spent on routine operational inquiries. Amazon states that this approach does not replace traditional business intelligence, which remains necessary for trend visualization and strategic reporting, but rather provides a dedicated pathway for real-time operational decision-making.</p>

Source: aws.amazon.com
Original article: https://aws.amazon.com/blogs/machine-learning/from-data-to-decisions-conversational-supply-chain-intelligence-with-amazon-quick-and-amazon-bedrock-agentcore/

Amazon is moving to reduce the industry's reliance on Nvidia hardware by introducing a new developer tool called TorchNeuron. The tool is designed to support the PyTorch framework, which is currently the most widely used toolkit for building artificial intelligence applications.

By facilitating the use of alternative chips for these AI workloads, Amazon aims to expand the range of hardware options available to developers. This initiative is a strategic effort to challenge Nvidia's significant influence within the AI development ecosystem, encouraging the migration of applications to non-Nvidia hardware infrastructure.

Source: TechCrunch
Original article: https://techcrunch.com/2026/10/19/amazon-is-taking-its-biggest-shot-at-nvidia-yet-with-new-torchneuron-dev-tool/

Amazon has significantly expanded its SageMaker JumpStart machine learning hub throughout the first three quarters of 2026, adding more than 100 new foundation models (FMs) from 30 distinct providers. The initiative reflects a broader push to accommodate increasing demand for diverse model architectures, sizes, and operational capabilities in enterprise environments.

The updated catalog now includes specialized contributions from major industry players. Alibaba remains the largest contributor with 29 models, including Qwen-based solutions for reasoning, code generation, and agentic workflows. NVIDIA has contributed 17 models, ranging from the Cosmos3 world foundation models for physical AI applications to the Nemotron series for high-performance inference. Other notable additions include Google’s Gemma 4 variants, Mistral’s Ministral family, and new models from IBM and Cohere focused on speech recognition and retrieval-augmented generation (RAG).

Beyond model volume, Amazon’s development efforts have focused on closing the multimodal gap. New additions include models capable of native vision, speech, and document understanding, such as the NVIDIA Nemotron 3 Nano Omni and IBM’s Granite speech-focused models. The company also integrated creative AI capabilities, adding text-to-image and video generation models like FLUX.2 and Meta’s Muse-Glimmer.

To address production requirements, Amazon introduced several efficiency-focused features. These include the addition of quantized and compact models—such as Microsoft’s Phi-4-mini-instruct and JetBrains’ Mellum2-12B—designed to optimize total cost of ownership and latency on smaller infrastructure instances. Additionally, purpose-built models for tasks like document extraction (OCR) and protein structure prediction (OpenFold3) have been added to provide users with specialized alternatives to general-purpose large language models.

Operational enhancements accompanied the catalog expansion to reduce friction in the model lifecycle. A redesigned Model Hub on the SageMaker AI console now allows users to browse and filter models by modality, size, and publisher directly from the landing page. Furthermore, Amazon introduced pre-configured deployment options—optimized for balanced, throughput-heavy, latency-sensitive, or cost-conscious workloads—to eliminate manual benchmarking. A new deep-link integration with Hugging Face also allows developers to navigate directly to pre-configured environments within SageMaker Studio, bypassing manual container and permission setups.

These updates collectively aim to simplify the transition from model discovery to production deployment, offering a broader, more flexible toolkit for enterprises building on Amazon Web Services.

Source: aws.amazon.com
Original article: https://aws.amazon.com/blogs/machine-learning/amazon-sagemaker-jumpstart-2026-year-to-date-in-review/