GitHub's fake star economy

A professionalized shadow economy has emerged where AI and LLM repositories purchase GitHub stars to manipulate VC sourcing signals and platform discovery algorithms. Technical analysis identifies these manipulated projects through distinct fingerprints, such as fork-to-star ratios significantly lower than organic baselines and high percentages of "ghost" accounts. While GitHub's enforcement remains largely reactive, the practice now faces significant legal exposure under new FTC rules and SEC fraud frameworks targeting inflated traction metrics.

Read more

John Ternus to become Apple CEO

Apple has announced that John Ternus, currently SVP of Hardware Engineering, will succeed Tim Cook as CEO on September 1, 2026, while Cook transitions to Executive Chairman. Ternus, a 25-year veteran who led the transition to Apple-designed silicon, will oversee the company's continued focus on high-performance hardware and vertically integrated systems. This leadership shift follows a long-term succession plan as Apple scales its ecosystem and hardware capabilities.

Read more

Qwen3.6-Max-Preview: Smarter, Sharper, Still Evolving

Qwen3.6-Max-Preview is a proprietary LLM preview featuring substantial improvements over Qwen3.6-Plus in agentic coding, world knowledge, and instruction following. It achieves top scores on benchmarks like SWE-bench Pro and SciCode while introducing a preserve_thinking feature to enhance multi-turn agentic reliability. The model is accessible via Qwen Studio and an OpenAI-compatible API on Alibaba Cloud Model Studio.

Read more

Deezer says 44% of songs uploaded to its platform daily are AI-generated

AI-generated tracks now account for 44% of daily uploads on Deezer, totaling approximately 75,000 tracks per day. Despite this volume, AI music represents only 1-3% of total streams, with 85% of those identified as fraudulent and subsequently demonetized. To manage this surge, Deezer utilizes automated detection to tag AI content, exclude it from algorithmic recommendations, and cease high-resolution storage of these files.

Read more

OpenClaw isn't fooling me. I remember MS-DOS

The author critiques current AI agent gateways like OpenClaw for lacking robust security boundaries, comparing their architectural flaws to the MS-DOS era. As an alternative, the Wirken.AI framework is presented, which implements granular isolation through per-channel process separation, Ed25519 identities, and hardened Docker sandboxes for tool execution. By enforcing security at the tool dispatch layer and utilizing hash-chained audit logs for attestation, this approach aims to mitigate risks like prompt injection and unauthorized system access inherent in broader container-wrap models.

Read more

AI Researchers' Views on Automating AI R&D and Intelligence Explosions

Leading researchers from frontier labs and academia identify the automation of AI research as a critical risk, anticipating a transition from AI assistants to autonomous developers capable of recursive improvement. While industry experts and academics disagree on growth timelines and explosive scenarios, a majority expect advanced R&D capabilities to remain internal and private. Governance consensus favors transparency-based mitigations over regulatory red lines.

ArXiv

SoK: Security of Autonomous LLM Agents in Agentic Commerce

Autonomous LLM agents are enabling machine-driven commerce, but this transition introduces new security vulnerabilities not adequately addressed by existing frameworks. This SoK develops a unified security framework for these agents, categorizing threats across five dimensions and identifying 12 cross-layer attack vectors. It proposes a layered defense architecture, emphasizing that securing agentic commerce requires coordinated controls spanning LLM safety, protocol design, identity, market structure, and regulation.

ArXiv

Anthropomorphism and Trust in Human-Large Language Model Interactions

A study of over 2,000 human-LLM interactions identifies warmth and cognitive empathy as key predictors of anthropomorphism and trust. While competence correlates with utility and trust, it does not significantly influence perceived human-likeness. Furthermore, subjective and personally relevant topics increase relational connection and anthropomorphism more than objective tasks.

ArXiv

ASI-Evolve: AI Accelerates AI

ASI-Evolve is an agentic framework for automated AI research that employs a closed-loop cycle to optimize data curation, neural architectures, and learning algorithms. By integrating a cognition base for human priors and a dedicated analyzer for insight distillation, the system discovered SOTA linear attention architectures and RL algorithms that significantly outperform GRPO. These results demonstrate the feasibility of using AI to accelerate foundational AI development through autonomous, long-horizon experimentation.

ArXiv

KV Cache Compression 900000x Beyond TurboQuant and Per-Vector Shannon Limit

Sequential KV compression improves upon per-vector quantization by treating the KV cache as a sequence of predictable samples from a formal language. The two-layer architecture utilizes probabilistic prefix deduplication to identify shared context and predictive delta coding to store residuals from the model's own KV predictions. This approach achieves a theoretical compression ratio of ~914,000x over TurboQuant, with efficiency scaling positively with context length.

ArXiv

Soul Player C64 – A real transformer running on a 1 MHz Commodore 64

Soul Player C64 is a 2-layer, 25K-parameter decoder-only transformer implemented in hand-written 6502 assembly for the Commodore 64. The architecture features real multi-head causal self-attention, RMSNorm, and int8 quantization, utilizing a 128-entry lookup table for softmax normalization. The model is trained via quantization-aware training (QAT) to handle the constraints of 8-bit fixed-point arithmetic and lacks hardware multiplication, relying on shift-and-add operations to generate tokens at a rate of roughly 60 seconds per token.

Repo

Git Push No-Mistakes

no-mistakes is a local git proxy that integrates an AI-driven validation pipeline into the development workflow. It utilizes isolated worktrees to run non-blocking checks and LLM-powered auto-fixes via agents like Claude or Codex before forwarding code to the remote. The tool automates PR creation and provides a TUI for reviewing findings, ensuring only validated code is pushed upstream.

Repo

Palmier – bridge your AI agents and your phone

Palmier is a bridge connecting local AI agents to mobile devices, enabling remote task dispatching, scheduling, and real-time interaction via a PWA or Android app. It features an MCP server that grants agents access to mobile capabilities like SMS, contacts, and GPS through a background daemon. The architecture supports local-only operation, cloud-relayed remote access, and low-latency auto-LAN routing for native mobile clients.

Repo

Self-healing browser harness via direct CDP

Browser Harness is a self-healing, CDP-based system that grants LLMs complete freedom to perform any browser task. It allows the LLM to dynamically write missing helper functions (e.g., in helpers.py) mid-task, eliminating the need for predefined frameworks. The system also offers free remote browsers for sub-agents and encourages the contribution of agent-generated "domain skills" to enhance site-specific interaction knowledge.

Repo

Aiguard-scan – Find secrets and vulnerabilities in AI-generated code

Agent Guard is a CLI security auditing tool designed to scan AI-generated code for vulnerabilities and sensitive information. It detects hardcoded secrets, SQL injection, and XSS risks introduced by LLM agents like Claude Code, Cursor, and Copilot. The tool runs locally for data privacy, supports CI/CD integration, and includes an --ai-only mode to specifically target code produced by AI agents.

Repo