AI is just unauthorised plagiarism at a bigger scale

The author criticizes the AI industry for non-consensual data scraping and the monetization of LLM-generated content that plagiarizes original work. They highlight a specific instance where ChatGPT-generated copies of their tutorials outranked the original source in Google search results, even while retaining the original's internal links. This illustrates the ongoing tension between original content creators and the SEO impact of AI-driven derivative works.

Read more

Throwing AI-generated walls of text into conversations

A "slop grenade" is the practice of pasting verbose LLM-generated responses into human communication channels where concise judgment is expected. This behavior degrades signal-to-noise ratios and stifles dialogue by forcing recipients to extract meaning from generic walls of text. AI should be used to sharpen thinking and improve clarity rather than to generate low-value filler.

Read more

Shunning AI is the human choice

The "AI Rebellion" reflects growing public and professional backlash against the forced integration of LLMs, driven by high-profile instances of hallucinations and fabricated content in literature and media. While tech leaders frame AI as an inevitable tool for hyper-optimization, critics argue that over-reliance on these models undermines human agency and creative legitimacy. This tension highlights a widening divide between those viewing AI as a necessary "rocket ship" and those who see it as a liability that produces low-quality "slop."

Read more

The memory shortage is causing a repricing of consumer electronics

AI-driven demand for HBM is causing a structural reset in the semiconductor market by cannibalizing wafer capacity previously allocated to DDR and LPDDR. Because HBM production is approximately three times more wafer-intensive than commodity DRAM, memory makers are prioritizing high-margin AI workloads, leading to 200-400% price surges in consumer-grade memory. This shift has already decimated the budget smartphone market in developing regions and is beginning to impact premium OEMs through increased BOM costs and supply delays. As next-generation AI platforms like Nvidia’s Vera Rubin further scale LPDDR consumption for inference, the era of cheap consumer computing is being reversed by the infrastructure requirements of the AI buildout.

Read more

Flipper One – we need your help

Flipper One is an open-source Linux cyberdeck powered by a Rockchip RK3576 SoC featuring a dedicated NPU for local LLM inference and an RP2350 co-processor for low-level management. The platform aims for a blob-free mainline Linux experience and introduces Flipper OS, a snapshot-based system designed for portable tactical use. Hardware capabilities include M.2 expansion for 5G or SDR, dual Gigabit Ethernet, Wi-Fi 6E, and a full-size HDMI 2.1 port for desktop and media applications.

Read more

CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs

CODA is a GPU kernel abstraction that addresses memory-bound bottlenecks in Transformer training by fusing auxiliary operators into GEMM epilogues. By reparameterizing computations like normalization and activations to execute while GEMM output tiles remain on-chip, it minimizes global memory traffic across the forward and backward passes. This approach provides a high-performance, composable framework for non-attention computations that maintains the efficiency of expert-written kernels.

ArXiv

Lecture Notes on Statistical Physics and Neural Networks

This text introduces statistical physics concepts, including phase transitions and the renormalization group, as relevant to neural networks and deep learning. It details the Boltzmann-Gibbs distribution, Ising spins, and spin-glass models, connecting them to Hopfield networks and Boltzmann machines. The discussion covers learning algorithms for Restricted Boltzmann Machines (RBMs), noting their link to the renormalization group and their influence on early deep learning, before concluding with a description of LLMs.

ArXiv

PopPy: Opportunistically Exploiting Parallelism in Python Compound AI Apps

PopPy optimizes compound AI applications by automatically parallelizing Python code that invokes external ML models. By combining an ahead-of-time compiler with a specialized runtime, it addresses challenges like dynamic dispatch and variable mutation to achieve up to 6.4x speedups while preserving sequential program semantics.

ArXiv

Multi-Stream LLMs: new paper on parallelizing/separating prompts, thinking, I/O

Traditional LLM agents are bottlenecked by sequential message exchange formats that prevent simultaneous reading, thinking, and acting. This work proposes instruction-tuning for parallel streams of computation, enabling models to process multiple input and output streams concurrently in a single forward pass. This architectural shift improves agent efficiency, security, and real-time responsiveness by decoupling core functions into separate, parallelized data streams.

ArXiv

DashAttention: Differentiable and Adaptable Sparse Hierarchical Attention

DashAttention is a differentiable hierarchical attention mechanism that replaces fixed top-k block selection with an adaptive $\alpha$-entmax transformation. By enabling variable-length block selection and maintaining gradient flow between sparse and dense stages, it achieves comparable accuracy to full attention at 75% sparsity while outperforming NSA and InfLLMv2 in long-context modeling. A GPU-aware Triton implementation further provides significant inference speedups over FlashAttention-3.

ArXiv

I Made a Claude Skill for Spec-Driven Development (SDD)

Spec-driven-development is a Claude skill designed to eliminate agent drift by establishing a shared source of truth across tools like Cursor, Copilot, and Claude Code. It enforces a workflow where requirements.md, design.md, and tasks.md are generated via a conversational interview before any implementation begins. The skill synchronizes these specs using a Universal Instruction Block in tool-specific configuration files, ensuring all LLM-based agents adhere to the same architectural constraints and divergence protocols.

Repo

LoongForge-A high-performance training framework for LLM, VLM, VLA, Wan

LoongForge is a modular, high-performance training framework built on Megatron-LM for LLMs, VLMs, diffusion, and embodied models. It supports native NVIDIA GPU and Kunlun XPU hardware, delivering up to 5.0x speedups through heterogeneous parallelism, MoE-native optimizations, and adaptive FP8 training. Key technical features include decoupled encoder-decoder training, DP load balancing, and bidirectional checkpoint conversion between Megatron and HuggingFace formats.

Repo

Assay – validation layer for AI agents that touch money

Assay is an open-source safety and validation library designed for AI agentic workflows in finance. It provides four layers of guardrails—output validation, tool-call gating, trajectory monitoring, and entity resolution—to audit agent decisions before executing downstream actions like trades or wire transfers. The library features Pydantic-based schema enforcement, regulatory rule packs (e.g., SEC, FINRA), and optional LLM-powered semantic consistency checks while keeping sensitive firm data local.

Repo

The three layers: browser, index, AI – what happens when you own all three

Information control is governed by a closed feedback loop consisting of a browser (attention signal), a search index, and an AI layer. Google dominates this stack, while LLM providers like OpenAI and Anthropic lack the browser-level session data necessary to measure real-world utility and ground their models. The transition to AI-mediated answers threatens to trigger model collapse by eliminating the human click signal that historically refined search indices. To counter this, the author proposes regulating Google as a utility and supporting Brave to maintain an independent, human-grounded alternative to centralized AI monopolies.

Repo

Linki – open-source AI SDR for LinkedIn sequences and cold email

Linki is an open-source, self-hosted AI SDR designed for multichannel B2B outreach across LinkedIn and email. It features an AI agent integrated with OpenRouter to leverage various models, utilizing a three-layer prompt hierarchy to generate personalized messages based on lead enrichment data. The platform includes automated cost tracking for token usage, Sales Navigator importing, and Apollo.io integration, ensuring data privacy by running entirely on the user's infrastructure.

Repo