Mar 8, 2026

Daily Updates

• How to run Qwen 3.5 locally
  Qwen3.5 offers a range of models, including 35B, 27B, and 397B, optimized for local deployment with 256K context support across 201 languages, excelling in tasks like coding and chat.

• Cloud VM benchmarks 2026: performance/price for 44 VM types over 7 providers
  AMD EPYC Turin emerges as the top performer in the 2026 cloud VM benchmarks, significantly outperforming previous generations and offering exceptional value across various providers.

• Files are the interface humans and agents interact with
  Filesystems are gaining traction in AI as they provide a persistent context layer that enhances agent functionality, allowing for more efficient coding and project management without the complexity of traditional databases.

• Autoresearch: Agents researching on single-GPU nanochat training automatically
  Autonomous AI agents now conduct research by modifying code and training models without human intervention, leading to potentially faster advancements in AI capabilities.

• SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via CI
  SWE-CI introduces a novel benchmark that evaluates agent capabilities in maintaining codebases through Continuous Integration, focusing on long-term maintainability rather than just short-term functional correctness.

• I don't know if my job will still exist in ten years
  The future of software engineering is uncertain, with the potential for AI to significantly alter job roles, possibly leading to a shift towards supervising AI agents rather than traditional coding tasks.

• Phi-4-reasoning-vision and the lessons of training a multimodal reasoning model
  Phi-4-reasoning-vision-15B is a 15 billion parameter multimodal reasoning model that excels in math and science reasoning, offering efficient performance across various vision-language tasks while minimizing compute costs.

• NanoJudge: Instead of prompting a big LLM once, it prompts a tiny LLM thousands of times
  NanoJudge revolutionizes ranking by utilizing thousands of pairwise comparisons with tiny LLMs, yielding a mathematically rigorous leaderboard that overcomes traditional LLM limitations.

• Graph-Oriented Generation (GOG): Replacing Vector R.A.G. for Codebases with Deterministic AST Traversal (70% Average Token Reduction)
  Graph-Oriented Generation (GOG) replaces Vector RAG by utilizing a deterministic Symbolic Reasoning Model (SRM) to parse codebases as a Directed Acyclic Graph (DAG), achieving a 70% average token reduction in processing.

• Combining Stanford's ACE paper with the Reflective Language Model pattern - agents that write code to analyze their own execution traces at scale
  Combining Stanford's ACE and Reflective Language Model (RLM) patterns enables agents to analyze execution traces more effectively by utilizing a Recursive Reflector that programmatically queries data, overcoming limitations of single-pass analysis.

• Introducing NNsight v0.6: Open-source Interpretability Toolkit for LLMs
  NNsight 0.6 introduces significant enhancements, including remote execution capabilities via the National Deep Inference Fabric (NDIF), allowing users to run analyses without local GPU resources by simply adding remote=True.

• We Turned Our Wireshark Wizard into a Markdown File
  Rocky AI has evolved from a simple chat widget to a sophisticated Root Cause Analysis Agent, leveraging human expertise codified into markdown files for effective troubleshooting of network issues.

• We analyzed 4,000 Ethereum contracts by combining an LLM and symbolic execution and found 5,783 issues
  SymGPT effectively combines large language models and symbolic execution to identify 5,783 ERC rule violations in 4,000 Ethereum contracts, highlighting its capability to detect vulnerabilities, including 1,375 with potential for financial theft.

• VeridisQuo - open-source deepfake detector that combines spatial + frequency analysis and shows you where the face was manipulated
  VeridisQuo introduces a method to detect deepfakes by combining spatial and frequency analysis, making it easier to identify manipulations in media.