Published on December 19, 2025 at 06:01 CET (UTC+1)
Beginning January 2026, all ACM publications will be made open access (1502 points by Kerrick)
The Association for Computing Machinery (ACM) announced that starting in January 2026, all its publications will transition to an open access model. This means that all future research papers published in ACM's vast portfolio of journals, conferences, and magazines will be freely available to the public without subscription barriers. This landmark policy shift is aimed at accelerating the dissemination of scientific knowledge in computing fields, including artificial intelligence, and aligns with broader pushes in academia and government for open science.
1.5 TB of VRAM on Mac Studio – RDMA over Thunderbolt 5 (272 points by rbanffy)
This article details an experiment where a cluster of Mac Studio computers, equipped with M3 Ultra chips, was linked using RDMA (Remote Direct Memory Access) over Thunderbolt 5 to create a unified pool of 1.5 TB of VRAM. The test used an open-source AI clustering tool called Exo 1.0, demonstrating a significant reduction in memory access latency. This showcases Apple's emerging potential in the high-performance computing and private AI inference space, providing a cost-effective alternative for running large language models locally compared to specialized server-grade GPU setups.
History LLMs: Models trained exclusively on pre-1913 texts (325 points by iamwil)
Researchers are developing "History LLMs," a family of large language models trained exclusively on a massive corpus of historical texts predating 1913. The project, which includes the upcoming 4-billion parameter Ranke-4B model, aims to create AI tools with period-accurate language and reasoning, free from modern linguistic and conceptual contamination. These models are intended for historians and social scientists to analyze historical texts, simulate period-appropriate writing, and potentially gain new insights into historical thought processes and cultural contexts.
We pwned X, Vercel, Cursor, and Discord through a supply-chain attack (702 points by hackermondev)
This is a detailed write-up of a significant supply-chain attack where the attackers compromised a popular open-source library to inject malicious code. The exploit ultimately allowed them to gain unauthorized access and potentially steal data from major tech companies including X (Twitter), Vercel, Cursor, and Discord, as well as hundreds of other organizations that depended on the compromised package. The article serves as a technical post-mortem, highlighting the critical vulnerabilities within software dependency ecosystems and the far-reaching impact a single compromised node can have.
Texas is suing all of the big TV makers for spying on what you watch (631 points by tortilla)
The state of Texas is suing major television manufacturers including Samsung, Sony, LG, Hisense, and TCL. The lawsuit alleges that these companies have created a "mass surveillance system" by designing their smart TVs to spy on users' viewing habits without obtaining proper, informed consent. The central claim is that the TVs collect detailed data on what is being watched, regardless of the input source (streaming apps, cable, gaming consoles, etc.), and transmit this data for advertising and profiling purposes, potentially violating state privacy and consumer protection laws.
GPT-5.2-Codex (423 points by meetpateltech)
OpenAI introduced GPT-5.2-Codex, a specialized version of its flagship model fine-tuned and optimized specifically for coding and software development tasks. This model represents a continued trend of creating domain-specific variants of general-purpose LLMs, offering enhanced performance in code generation, explanation, debugging, and refactoring. Its release signals intense competition in the AI-powered developer tools space and aims to solidify OpenAI's position against other code-specialized models and assistants.
Noclip.website – A digital museum of video game levels (57 points by ivmoreau)
Noclip.website is a digital, browser-based museum that allows users to freely explore meticulously reconstructed 3D maps and levels from a wide variety of classic and modern video games. Using techniques like reverse engineering and asset extraction, the site lets users fly through iconic game environments without the constraints of normal gameplay, serving as both an archival project for game preservation and a unique tool for education, nostalgia, and technical study of game design.
How China built its ‘Manhattan Project’ to rival the West in AI chips (273 points by artninja1988)
The article investigates China's concerted, state-backed effort—described as its 'Manhattan Project'—to achieve self-sufficiency in producing advanced AI chips, circumventing Western sanctions. It details the massive mobilization of resources to develop domestic alternatives to critical technologies like Extreme Ultraviolet Lithography (EUV) machines. The piece highlights the involvement of companies like Huawei across the entire supply chain and frames this as a pivotal technological and geopolitical race with the West, determining future leadership in artificial intelligence.
Skills for organizations, partners, the ecosystem (248 points by adocomplete)
Anthropic announced major expansions to the "Skills" feature for its Claude AI, transforming it into a platform for customizable, repeatable workflows. Key updates include centralized admin management for enterprise deployment, a new directory of pre-built skills from partners like Notion and Figma, and the publication of an open standard to make skills interoperable across different AI platforms. This move positions Claude as a hub for enterprise automation and signifies a strategic push towards ecosystem development and platform lock-in through user-created modular functions.
Telegraph chess: A 19th century tech marvel (17 points by sohkamyung)
This article explores the 19th-century invention of "telegraph chess," a system where two players in different cities could play chess in real-time by transmitting moves over electrical telegraph lines. It details the technical and procedural innovations required to make this work, highlighting it as a pioneering example of networked digital communication and online gaming long before the internet. The story serves as a historical analogy, illustrating how new communication technologies immediately inspire novel applications and reshape social interactions.
Democratization and Specialization of Model Training: The emergence of "History LLMs" and code-specific models like GPT-5.2-Codex highlights a clear trend beyond simply scaling general-purpose models. Why it matters: The field is moving towards creating specialized, high-value models trained on curated, domain-specific datasets (like pre-1913 texts or code repositories). This leads to more accurate, efficient, and context-aware tools for professionals in fields from history to software engineering. The implication is a future ecosystem of myriad specialized AIs, lowering the barrier to entry for niche applications but also raising questions about data curation and the fragmentation of capabilities.
The Rise of Alternative AI Hardware Ecosystems: Two articles underscore a strategic diversification in AI compute. Apple's demonstration of performant, unified-memory AI clustering with consumer-grade hardware (Mac Studio) and China's state-driven project to build a complete domestic AI chip supply chain both signal a shift. Why it matters: This challenges NVIDIA's dominance and creates new pathways for deployment. It matters for cost, accessibility, and geopolitical resilience. The takeaway is that the hardware landscape for AI training and inference is becoming more competitive and segmented, offering developers and organizations more choices based on budget, privacy needs, and regulatory environment.
Enterprise AI is Becoming Platform-Centric and Integrable: Anthropic's launch of an organizational Skills directory and open standard is a prime example. Why it matters: The battle for enterprise AI is shifting from raw model capability to who can best integrate into and automate existing workflows (like those in Notion, Figma, etc.). Platform lock-in will be achieved through ecosystem development, user-friendly tooling for custom agents, and interoperability standards. For developers, the actionable takeaway is to build AI applications that are modular and easily embedded into larger enterprise platform ecosystems.
Intensifying Focus on AI Security and Supply-Chain Integrity: The detailed account of a supply-chain attack compromising major AI/tech companies (Cursor, Vercel) serves as a critical warning. Why it matters: As AI development heavily relies on open-source libraries and pre-trained models, the software supply chain becomes a massive attack surface. This threat directly impacts model integrity, data privacy, and system security. The implication is that security practices like strict dependency auditing, artifact signing, and robust CI/CD pipelines are no longer optional but foundational for any serious AI development or deployment operation.
The Institutional Push for Open Knowledge Accelerates AI Research: ACM's decision to mandate open access for all its publications is a monumental policy change. Why it matters: It removes significant paywalls from a vast repository of cutting-edge computer science research, directly fueling faster innovation in AI/ML. Easier access to seminal papers, algorithms, and results lowers barriers for independent researchers, startups, and academics in developing countries. The long-term implication is a potential acceleration of the overall pace of AI advancement and a more geographically distributed research community.
Privacy and Ethical Scrutiny Extends to AI-Adjacent Consumer Tech: Texas's lawsuit against TV makers for covert data collection is part of a broader pattern. Why it matters: The data collected by ubiquitous smart devices (TVs, speakers, etc.) is often the fuel for training and refining AI models, especially in advertising and recommendation systems. Increasing legal and regulatory pushback creates headwinds for this data-acquisition model. For AI developers, this means a growing need for transparent data practices, clearer user consent mechanisms, and a potential shift towards synthetic data or more privacy-preserving training techniques like federated learning.
Analysis generated by deepseek-reasoner