shiv rossi

AI News Feed

These are AI-generated summaries I use to keep tabs on daily news.

prev
next latest

Daily Tech Newsletter - 2025-12-13

AI Regulation and Policy

President Trump signed an executive order to centralize AI regulation at the federal level, aiming to preempt state-level laws. This revives an earlier effort that failed due to bipartisan opposition. The order establishes an "AI Litigation Task Force" in the Department of Justice to challenge state laws, particularly those requiring AI models to alter truthful outputs or mandating algorithmic discrimination risk assessments. Opponents argue that the order favors Silicon Valley companies and could expose vulnerable populations to AI-related harms. The order also assigns an influential role to David Sacks, Special Adviser for AI and Crypto, in determining which state laws to challenge. In a related event, the Chandler City Council in Arizona rejected a proposal for an AI data center despite lobbying efforts that warned about federal preemption of AI data center zoning, highlighting a growing trend of local governments pushing back against energy-intensive data centers.

Relevant URLs:

AI in Healthcare Advancements

Google, in collaboration with African data science communities, hosted an Africa-wide ideathon focused on leveraging open Health AI models like MedGemma, TxGemma, and MedSigLIP to solve real-world health challenges on the continent. Winning projects included AI-powered cervical cancer screening tools, automated cervical cytology, and maternal health advice centers using TTS/STT. The Dawa Health cervical cancer screening tool is already piloting an early access program with plans for scaling to 50,000 patients. Other solutions addressed connectivity challenges with offline applications for skin condition triage and pregnancy risk assessment. The initiative aims to build capacity within African AI and health communities while fostering localized solutions addressing critical medical needs.

Relevant URLs:

Innovations in Small Language Models (SLMs)

Nanbeige LLM Lab introduced the Nanbeige4-3B family, a suite of 3-billion-parameter SLMs, demonstrating reasoning capabilities competitive with, and in some cases surpassing, models ten times its size. This performance is attributed to a specialized training recipe emphasizing data quality, curriculum scheduling, distillation, and reinforcement learning. Benchmarking results show strong performance in reasoning tasks, particularly on AIME 2024 and GPQA-Diamond. MIT CSAIL researchers developed DisCIPL, a self-steering system that improves the accuracy and efficiency of LMs on complex tasks with strict constraints. DisCIPL employs a hierarchical approach where a "boss" LLM plans the task and delegates the execution to smaller "follower" LMs, guided and corrected by the LLM. The cost savings and efficiency gains of DisCIPL are achieved by reasoning with Python code rather than text and utilizing cheaper Llama models. Finally, TimeCapsule LLM is a series of SLMs designed to emulate the voice, vocabulary, and worldview of specific historical eras by being trained exclusively on historical data from those periods, thereby reducing modern bias. Datasets are version-specific, with v2 utilizing 90GB of text from 1800-1875 London.

Relevant URLs:

Agentic Storytelling with Local LLMs

A tutorial details the construction of a fully local, API-free agentic storytelling system utilizing Griptape and a lightweight Hugging Face model. The system features an agent with tool-use capabilities and a multi-stage workflow designed for creative generation, encompassing world-generation, dynamically constructed character-generation tasks, and a final storytelling task. Griptape is used to orchestrate complex reasoning, tool interactions, and creative generation using local models, ensuring full control, reproducibility, and flexibility.

Relevant URLs:

Wider Adoption of AI "Skills" and Tooling

OpenAI has integrated "skills" into both its ChatGPT platform and Codex CLI tool, a development that follows Anthropic's earlier implementation. Skills are folder-based, containing a Markdown file and optional resources, making them easily adaptable for LLM tools capable of filesystem navigation. The author recommends a "progressive disclosure" method for using these skills. A new version of the LLM Python library and CLI tool (v0.28) has been released, adding support for new OpenAI models (gpt-5.1, gpt-5.2) and incorporating bug fixes and improved development workflows.

Relevant URLs:

AI Training Programs for Military Leaders

MIT's Departments of Mechanical Engineering (MechE) and Electrical Engineering and Computer Science (EECS) are jointly launching a new certificate program called "2N6: Applied Artificial Intelligence Program for Naval Officers." This two-year program aims to equip naval officers with advanced skills in applied AI for naval and military applications. Artificial intelligence is highlighted as a critical tool for national security enhancing decision making and improving efficiency across various military domains.

Relevant URLs:

Challenges in AI Content Generation

Amazon removed its AI-generated recap for Fallout Season 1 after fans identified significant factual errors and inaccuracies. Prime Video had begun testing AI recaps aiming to summarize plot points with AI voiceover and music. Amazon has a recent history of flawed AI features and negative feedback.

Relevant URLs:

Kindle's "Ask this Book" Feature Sparks Controversy

Amazon launched "Ask this Book," an AI-powered in-book chatbot in its Kindle iOS app, offering spoiler-free answers to reader questions about book content. Authors and publishers cannot opt out of the feature, and it was implemented without widespread prior notification to creators. The feature is surfacing amidst ongoing lawsuits about authors and unauthorized use of works, raising concerns about Amazon's legal basis and potential copyright infringement.

Relevant URLs:

Guarding Git Forges Against AI Scrapers

A self-hosted Git forge experienced an onslaught of AI scrapers, resulting in significant resource drain and impacting server performance and internet speed. The author implemented Iocaine middleware, combined with the Nam-Shub-of-Enki classifier, to effectively identify and "poison" bots by serving them endless garbage content. Scraping leads to authors radicalized opinions regarding the "corporate web."

Relevant URLs:

Microsoft's Investment and Digital Sovereignty Plan for Canada

Microsoft is investing $19 billion CAD in Canada between 2023-2027, including over $7.5 billion CAD in the next two years, for new digital and AI infrastructure. A five-point plan is being launched to protect Canada’s digital sovereignty, focusing on cybersecurity, data residency, privacy, AI developer support, and cloud service continuity. Microsoft is partnering with Canadian AI developers like Cohere, integrating their advanced language models into Azure to amplify Canadian innovation.

Relevant URLs:

AI Predictions for 2026

An annual tradition of making predictions for AI in the coming year has begun, with 20 predictions for AI in 2026, evaluating the accuracy of the predictions for 2025.

Relevant URLs:

AI Demystification through Mathematics

Secondary school teaching materials and workshop programs teach secondary students to demystify fundamental AI concepts through accessible math. Through use of hands-on activities to debunk misconceptions that AI systems "think" or "learn", mathematics are shown to provide a deeper understanding of AI and its risks and opportunities.

Relevant URLs:

Subscriptions to Financial News

Financial Times offers subscriptions to digital and print financial news.

Relevant URLs:

macOS 26.2 Release and Developer Tools

The macOS 26.2 SDK, bundled with Xcode 26.2, introduces RDMA over Thunderbolt 5, enabling low-latency communication for applications like distributed AI inference using MLX. The release also includes a new AppStore.ageRatingCode API in StoreKit and resolves several bugs related to AirDrop, Instruments, StoreKit, and Virtualization.

Relevant URLs: