College instructor turns to typewriters to curb AI-written work

A Cornell University instructor has introduced manual typewriters for assignments to prevent students from using generative AI and online translation tools. This "analog" approach eliminates access to LLMs and digital editing features, forcing students to engage in unassisted, intentional composition. The initiative reflects a broader academic trend toward low-tech evaluation methods to ensure academic integrity in the age of AI.

Read more

Graphs that explain the state of AI in 2026

Stanford’s 2026 AI Index reveals that the US continues to lead in notable model releases, while global compute capacity has scaled 3.3x annually since 2022. LLMs are rapidly conquering complex benchmarks in agentic AI and drug discovery, yet training-related carbon emissions are skyrocketing and models still exhibit failures in basic spatial reasoning. Despite record private investment of $581 billion, the impact on technical employment remains mixed, with a notable reduction in entry-level software engineering headcounts.

Read more

Claude Opus 4.7 Intelligence, Performance and Price Analysis

The Artificial Analysis Intelligence Index v4.0 evaluates LLMs across 10 benchmarks, including GPQA Diamond and SciCode, to compare proprietary and open-weights models. It provides technical metrics on price-performance, context windows for RAG, and inference latency, specifically accounting for "thinking" time in reasoning models. The index also tracks output speed and distinguishes between active and total parameters for MoE architectures.

Read more

In the AI propaganda war, Iran is winning

Iran is leveraging generative AI to produce sophisticated propaganda, outperforming the Trump administration in digital influence operations. These AI-generated videos represent a shift from traditional, unconvincing state media to more engaging, creative content. This evolution underscores the growing role of AI in geopolitical information warfare and the democratization of high-quality media synthesis for state actors.

Read more

Headless Everything for Personal AI

Services are shifting toward headless architectures to accommodate personal AI agents, utilizing MCP and CLIs instead of traditional GUIs. CLIs offer superior composability and security, allowing agents to execute local, cross-app workflows without the vulnerabilities inherent in complex web interfaces. This transition suggests a future where front-end design prioritizes brand identity while core functionality is delivered via hardened, minimal interfaces optimized for agentic consumption.

Read more

Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems

Claude Code is an agentic coding tool built on a core execution loop supported by a seven-mode permission system, a five-layer context compaction pipeline, and extensibility via MCP and subagent delegation. A comparative analysis with OpenClaw demonstrates how deployment contexts—CLI-based tools versus gateway-integrated assistants—influence architectural decisions regarding safety classification, runtime environments, and capability registration. The study identifies key design principles and future directions for scaling agentic LLM systems.

ArXiv

Generative artificial intelligence for computational chemistry

Generative AI methods, including LLMs, GANs, and flow models, are being applied to computational chemistry for molecular sampling, force field development, and structure prediction. Despite progress, these models face challenges in predicting emergent chemical phenomena. To reach true predictive utility, future AI architectures must integrate fundamental chemical principles and statistical mechanics.

ArXiv

Huoziime: An On-Device LLM-Enhanced Input Method for Deep Personalization

HUOZIIME is a personalized on-device IME that utilizes a post-trained LLM and a hierarchical memory mechanism to capture and leverage user-specific input history. The system incorporates mobile-specific optimizations to ensure real-time responsiveness and privacy-preserving performance while delivering high-fidelity generative text predictions.

ArXiv

Productivity gains with an AI-based IDE at Google

Google details the development and optimization of internal AI-powered IDE features, specifically code completion and natural-language code transformation. By addressing latency, UX, and suggestion quality across the UI, backend, and model layers, they achieved significant enterprise productivity gains through rigorous experimentation.

ArXiv

Generating Hierarchical JSON Representations of Scientific Sentences Using LLMs

Researchers fine-tuned a lightweight LLM using a novel structural loss function to map scientific sentences into hierarchical JSON representations. By reconstructing the original text from these structures, the study demonstrates through semantic and lexical similarity that structured formats effectively preserve the information density of scientific content.

ArXiv

I can't write Python. It works anyway

Garmin Local Archive is a local-first Python pipeline designed to archive and normalize Garmin Connect data into structured JSON and Excel formats, bypassing cloud-side data degradation. It features a modular architecture for data validation and AES-256-GCM encrypted credential management. The system is specifically optimized for local AI workflows, providing pre-processed datasets and system prompts for RAG implementations using tools like Ollama, Open WebUI, and AnythingLLM.

Repo

Open-source isolated runtime for AI agents

SmolVM provides secure, isolated microVMs powered by Firecracker for AI agents to execute code, browse the web, and manage state. It features sub-second boot times, hardware-level isolation, and network egress controls to safely run untrusted code. The platform supports host directory mounting and integrates with major agent frameworks like LangChain and PydanticAI.

Repo

LingBot-Map: Geometric Context Transformer for Streaming 3D Reconstruction

LingBot-Map is a feed-forward 3D foundation model for streaming 3D reconstruction based on a Geometric Context Transformer. The architecture unifies coordinate grounding, dense geometric cues, and long-range drift correction through anchor context, pose-reference windows, and trajectory memory. It achieves high-efficiency inference at ~20 FPS for sequences exceeding 10,000 frames by leveraging paged KV cache attention and FlashInfer.

Repo

Scopeon – AI Observability – token breakdown, cache ROI, cost tracking, CI gates

Scopeon is a local-first observability tool designed for AI coding agents like Claude Code, Aider, and Cursor. It provides granular tracking of token usage, prompt cache hit rates, and real-time USD costs across sessions and projects. Key features include context window fill-rate predictions, CI cost gating, and an MCP server that enables agents to monitor their own resource consumption and provenance.

Repo

Don't just quit. Leave with leverage. An open-source employment attorney via .md

ai-quit-job is an open-source repository featuring 14 markdown-based system prompts that turn LLMs into specialized employment attorneys and career coaches. The project uses a modular agent architecture to help users identify legal leverage, such as wage theft or misclassification, and generate structured frameworks for severance negotiations and resignations. Designed for portability, these prompts can be integrated into any LLM interface or API to provide diagnostic screeners and transition strategies.

Repo