Adafruit receives demand letter from Fenwick legal counsel on behalf of Flux.ai

Adafruit has received a demand letter from Flux.AI (Defy Gravity, Inc.) alleging defamation and CFAA violations following a report on Flux’s intellectual property and commercial metrics. Adafruit maintains the data was accessed via a server misconfiguration during a responsible disclosure process but has temporarily suspended blog operations while evaluating legal next steps.

Read more

MAI-Code-1-Flash

Microsoft has introduced MAI-Code-1-Flash, a lightweight agentic model integrated into GitHub Copilot and VS Code for optimized developer workflows. The model utilizes adaptive solution length control to improve efficiency, outperforming Claude Haiku 4.5 on SWE-Bench Pro while using up to 60% fewer tokens. Trained on licensed data using production-aligned harnesses, it demonstrates superior instruction-following and reasoning capabilities across coding, math, and science benchmarks.

Read more

Trump signs downsized AI order after weeks of reversals

President Trump signed a revised executive order establishing a voluntary 30-day pre-release review period for advanced AI models, reduced from a proposed 90-day window to prioritize innovation and competition with China. The directive focuses on mitigating cybersecurity risks—highlighted by the capabilities of models like Anthropic’s Mythos—by creating a Treasury-led cybersecurity clearinghouse and an NSA-overseen classified benchmarking process. While the order explicitly avoids mandatory licensing or preclearance, it mandates federal network hardening and directs the DOJ to prosecute AI-facilitated hacking.

Read more

AI outperforms law professors in Stanford Law study

A Stanford Law School study reveals that law professors prefer LLM-generated answers to student questions over peer-written responses in 75% of blind head-to-head matchups. LLMs demonstrated superior performance in complex legal reasoning and synthesis, with a significantly lower rate of pedagogical harm (3.5%) compared to human instructors (12%). The findings suggest that LLMs can effectively navigate judgment-rich domains, performing comparably to top-tier human experts in providing on-demand educational support.

Read more

MAI-Thinking-1

Microsoft AI has introduced MAI-Thinking-1, a 35B-active parameter sparse MoE reasoning model trained from scratch on licensed data without third-party distillation. The model achieves competitive results on SWE-Bench Pro and AIME benchmarks, matching or exceeding the performance of Claude Opus 4.6 and Sonnet 4.6 in human side-by-side evaluations. Designed for enterprise use, it features a 256k context window and was developed using a "Hill-Climbing Machine" pipeline to ensure continuous, self-sufficient scaling of agentic and mathematical reasoning capabilities.

Read more

AI Agents Enable Adaptive Computer Worms

Researchers have demonstrated a self-sustaining AI worm that utilizes open-weight LLMs on compromised hosts to generate target-specific attack strategies in real time. Operating across Linux, Windows, and IoT environments, the malware leverages stolen compute to eliminate marginal costs and bypasses centralized safety controls. This shift from static exploits to autonomous reasoning marks the emergence of generative adversaries capable of real-time adaptation and synthesis.

ArXiv

Benchmarking LLM-as-a-Judge for Long-Form Output Evaluation

LongJudgeBench is a new benchmark designed to evaluate the reliability of LLM-as-a-judge for long-form generation, addressing the limitations of existing short-form meta-evaluation. It assesses judges on complex document-level criteria such as organization, coverage, and consistency across diverse real-world scenarios. Findings indicate that current LLM judges remain unstable and that while rubrics or references improve performance, they are insufficient to ensure consistent human-aligned evaluation.

ArXiv

Type-Error Ablation and AI Coding Agents

Researchers investigated whether AI coding agents benefit from more detailed error messages than the terse formats traditionally optimized for human programmers. Using a statically typed language, an ablation study demonstrated that providing detailed context, such as unification stacks, significantly improves an agent's ability to repair type errors compared to minimal or dynamic error reports. The findings suggest that static type systems are more effective for agent-led debugging than test suites alone and that agents can successfully reconstruct program semantics even when code is obfuscated.

ArXiv

The Sum-Product conjecture is false for real numbers

The sum-product and many sums and products conjectures for real numbers are disproved through the construction of large sets with sub-quadratic growth in both addition and multiplication. These results extend to $p$-adics, finite fields, and function fields, while providing new lower bounds for linear equations in multiplicative groups and unit equations.

ArXiv

The Architecture of Errors

Universal LLM reliability is unattainable in unbounded domains due to an infinite set of failure modes, but real-world applications operate in bounded "patches" where failures are sparse and repetitive. In these contexts, reliability shifts from an exponential problem to a local catalogue-discovery and intervention-coverage task. Proposition 2 demonstrates that once a patch's failure catalogue saturates, the required intervention budget becomes domain-constant, provided the number of hard decisions does not scale linearly with sequence length.

ArXiv

CLI tool that packages data science projects for LLM context windows

data2prompt is a CLI tool designed to optimize data-heavy projects for LLM context windows through intelligent sampling and truncation of CSV, SQL, and Jupyter files. It provides token-aware output in Markdown or XML formats, utilizing tiktoken for real-time estimation to ensure prompts fit within target model limits. The tool preserves semantic structure and schema while stripping non-essential data, facilitating high-performance orchestration for data science workflows.

Repo

AI Vulnerability Intelligence Agent Converts CVEs to Actionable Security Reports

CVE AI Agent is an autonomous vulnerability intelligence engine that utilizes a two-pass architecture to ingest and enrich CVE data. It employs deterministic logic for initial data extraction followed by LLM enrichment for qualitative analysis, maintaining a token-efficient prompt structure (~1k tokens) to reduce hallucinations and API costs. The system is LLM-agnostic, supporting major providers like Gemini and OpenAI, and features automated recheck workflows and integrations with SOC tools like Jira, Splunk, and Slack.

Repo

Viveka: filter LLM output against a Lean-verified Advaita Vedanta model

Viveka is a witness-centered filter layer for LLM applications that validates response claims against the Scherf Logic API, a Lean 4 formalization of Advaita Vedānta axioms. It detects linguistic patterns associated with user objectification, epistemic over-claims, and induced dependency by mapping extracted claims to machine-verified axioms. The system provides non-mutating verdicts—PASS, FLAG, CORRECT, or BLOCK—and explicitly avoids silent rewrites to maintain transparency and subject/object integrity.

Repo

Tiny GPT in Go. Optimised for Understanding. Trained on Jules Verne Books

gpt-go is a simple GPT implementation in pure Go, designed as a learning companion for Karpathy's "Neural Networks: Zero to Hero" course. It prioritizes radical simplicity, foregoing batches and external dependencies to illustrate core LLM concepts from basic neurons to self-attention using 2D matrices. The project allows training on custom datasets and includes a chat-only inference mode.

Repo

MDMA – Turn LLM Responses into Interactive UI via MCP

MDMA (Markdown Document with Mounted Applications) is an extensible framework that enables LLMs to generate interactive UI components—such as forms, tables, and approval gates—directly within Markdown using fenced code blocks. It provides a complete ecosystem including a reactive runtime for state management, a React rendering layer, and a validator designed to catch and auto-fix common LLM structural errors. To streamline integration, it offers model-tuned system prompts and an MCP server that allows AI agents to programmatically access the spec and authoring tools.

Repo