Is legal the same as legitimate: AI reimplementation and the erosion of copyleft

The maintainer of the chardet library used Claude to reimplement the project from scratch using only its API and test suite, subsequently relicensing the work from LGPL to MIT. This move has sparked a debate over whether LLM-driven "clean-room" reimplementations, while potentially legal, undermine the social compact of copyleft by stripping protections from the software commons. The case highlights a growing need for evolved licensing models, such as specification copyleft, to address the technical ease with which LLMs can bypass traditional source-code protections.

Read more

The Mog Programming Language

Mog is a statically typed, embedded programming language designed to be written by LLMs, featuring a minimal specification that fits within a 3200-token context window. It employs a capability-based security model to safely restrict I/O and propagate permissions from the host to agent-generated code, preventing the sandbox escapes common in Python or Bash. Compiled to native code via a safe Rust toolchain, Mog enables low-latency execution for agent hooks and plugins while providing native support for async/await and n-dimensional tensors.

Read more

Peter Thiel and Jeffrey Epstein Had a Yearslong Relationship

Jeffrey Epstein facilitated the strategic relationship between Peter Thiel and former Israeli Defense Minister Ehud Barak, directly leading to Palantir’s deployment of AI tools for Israeli military operations and the Pentagon. These introductions also enabled Founders Fund investments in Israeli cybersecurity firms like Carbyne and Guardicore, which were founded by Unit 8200 veterans. This network integrated advanced surveillance and AI-driven targeting systems into regional defense infrastructures.

Read more

Learnings from paying artists royalties for AI-generated art

Tess.Design was an ethical AI marketplace that fine-tuned Stable Diffusion models on consenting artists' styles, offering a 50% royalty and a novel legal framework for derivative copyright ownership. The project was shut down after 20 months due to significant artist resistance, enterprise legal concerns regarding unresolved AI litigation, and unsustainable unit economics. The retrospective highlights the challenges of supply-side recruitment in fragmented creative markets and the impact of cultural hostility toward generative AI.

Read more

Revealed: UK's multibillion AI drive is built on 'phantom investments'

A Guardian investigation reveals that the UK’s multibillion-pound AI drive is largely supported by "phantom investments" characterized by misleading infrastructure claims and minimal government oversight. Key firms CoreWeave and Nscale are accused of misrepresenting rented space in existing facilities as new datacentres and claiming undeveloped scaffolding yards as active supercomputer sites. The findings suggest that reported private investment totals and job creation figures remain largely unaudited, raising significant questions about the actual expansion of the UK's sovereign compute capacity.

Read more

Building AI Coding Agents for the Terminal

OPENDEV is an open-source, terminal-native coding agent designed for autonomous software engineering through a compound AI architecture. It employs a dual-agent system for planning and execution, utilizing workload-specialized model routing and adaptive context compaction to maintain reasoning efficiency. Key features include lazy tool discovery, automated project-specific memory, and event-driven reminders to prevent instruction fade-out during long-horizon tasks.

ArXiv

Can AI Agents Agree?

LLM agents struggle to achieve reliable agreement in Byzantine consensus settings, with performance degrading as group size increases or adversarial agents are introduced. Failures are primarily characterized by liveness issues like timeouts and stalled convergence rather than value corruption. These findings indicate that robust coordination is not yet a dependable emergent capability in current LLM-agent groups.

ArXiv

Latent Context Compilation: Distilling Long Context into Compact Portable Memory

Latent Context Compilation addresses the trade-offs between amortized compression and Test-Time Training by distilling long contexts into stateless, portable buffer tokens via a disposable LoRA module. This framework utilizes a self-aligned optimization strategy with context-agnostic random queries to maintain reasoning and fine-grained detail without synthetic data or weight modifications. Experiments on Llama-3.1-8B demonstrate effective 16x compression while decoupling memory density from model parameters.

ArXiv

Random-Bridges as Stochastic Transports for Generative Models

This paper proposes random-bridges as a framework for generative modeling, serving as stochastic transports between probability distributions with flexible Markovian or non-Markovian dynamics. By deriving learning and simulation algorithms from general probabilistic statements, the authors demonstrate that Gaussian random bridges achieve competitive FID scores with significantly fewer inference steps. The approach offers a computationally efficient alternative for high-speed generation tasks compared to traditional diffusion-based methods.

ArXiv

Paris: Causally Consistent Transactions with Partial Replication

PaRiS introduces the first Transactional Causal Consistency (TCC) system for geo-replicated data platforms that supports partial replication and non-blocking parallel reads. It employs a novel Universal Stable Time (UST) protocol, which uses lightweight gossip to identify a global stable data snapshot, enabling consistent, non-blocking reads from any server or replication site. This design requires only one timestamp for dependency tracking, enhancing resource efficiency and scalability. Evaluations demonstrate PaRiS scales effectively with DCs and partitions, handles larger datasets than full-replication solutions, and delivers substantial performance gains for non-blocking reads across diverse workloads.

ArXiv

Mcp2cli – One CLI for every API, 96-99% fewer tokens than native MCP

mcp2cli converts MCP servers and OpenAPI specs into CLIs at runtime without requiring codegen. It achieves 96–99% token savings by replacing proactive schema injection with on-demand tool discovery via --list and --help commands. The tool is provider-agnostic, supports OAuth and various MCP transports, and features a "TOON" output format designed for token-efficient LLM consumption.

Repo

DenchClaw – Local CRM on Top of OpenClaw

DenchClaw is an open-source, locally hosted AI CRM built on the OpenClaw framework. It provides a web interface featuring AI chat, object tables, and a skills store, requiring Node 22+ for deployment via npx. The platform is MIT-licensed and utilizes the OpenClaw CLI for managing its underlying gateway and configuration profiles.

Repo

Run 500B+ Parameter LLMs Locally on a Mac Mini

Graviton is an open-source inference engine that enables running 500B+ parameter LLMs on consumer hardware by streaming layers from SSD and utilizing memory-mapping. It features a QuantizedLinear module supporting INT4, mixed-precision, and 1.58-bit ternary weights, alongside speculative decoding and Top-K dynamic sparsity for optimized throughput. The framework supports MPS, CUDA, and CPU backends, providing a REST API and Python SDK for local model deployment.

Repo

Own your AI's context and memories across every model and device

To bypass vendor lock-in and data monetization, the author developed a self-hosted, model-agnostic memory system using the Model Context Protocol (MCP). The architecture leverages a Postgres database with pgvector on Supabase to maintain a persistent knowledge graph, an MCP gateway on a VPS for tool multiplexing, and TypingMind as a BYOK client. This infrastructure allows various LLMs to access a unified, private context, ensuring data ownership and compounding knowledge across different models and sessions.

Repo

Hebbs – a single-binary memory engine for AI agents – sub-10ms

HEBBS is a high-performance memory engine for AI agents that replaces fragmented RAG and vector DB setups with a single Rust-based binary. It provides sub-10ms recall across four strategies—similarity, temporal, causal, and analogical—while supporting importance-driven encoding and episodic-to-semantic consolidation. The engine features tunable scoring based on relevance, recency, and reinforcement, and can be deployed as a standalone server or an embedded library.

Repo