Frontier AI has broken the open CTF format

The rapid advancement of frontier LLMs, specifically Claude Opus 4.5 and GPT-5.5, has effectively ended the competitive era of open online CTFs by making medium to "Insane" difficulty challenges solvable via agentic orchestration. This shift has transformed the format into a pay-to-win benchmark for compute and orchestration efficiency rather than a measure of human security expertise. Consequently, the traditional CTF scoreboard has lost its utility as a pedagogical ladder for beginners and a recruitment metric for professionals, necessitating a move toward controlled environments and community-focused learning.

Read more

DeepSeek-V4-Flash means LLM steering is interesting again

Steering involves manipulating LLM activations mid-inference to guide model behavior, using techniques like activation subtraction or SAEs to identify steering vectors. While projects like DwarfStar 4 are bringing these capabilities to local models like DeepSeek-V4-Flash, prompting or fine-tuning often remains more efficient for modifying complex traits. Key practical applications include "abliteration" to bypass refusals and potential context compression, though the scalability of steering for high-level concepts like intelligence remains unproven.

Read more

US is starting to see heavy job losses in roles exposed to AI

Recent BLS data shows a 0.2% decline in employment for 18 AI-exposed occupations between May 2024 and May 2025, while overall US employment grew by 0.8%. These losses, affecting approximately 10 million roles, are most concentrated in customer service, administrative, and sales positions. This trend marks the second consecutive year of contraction for roles identified as highly susceptible to AI automation.

Read more

Prepare for an AI Jobs Apocalypse

The AI boom has intensified public anxiety over a potential "jobs apocalypse," specifically impacting entry-level opportunities for computer programmers and college graduates. Although widespread unemployment has not yet materialized, the text argues that governments should proactively establish safety nets to address the disruptive potential of AI on the labor market.

Read more

A Dark Cave – Minimalistic Graphics in the Age of AI Slop

A Dark Cave is a text-based incremental survival RPG and settlement simulator inspired by titles like A Dark Room and Kittens Game. It features resource management, turn-based combat, and narrative-driven progression within a Lovecraftian setting. The HTML5-based game utilizes idle mechanics and local state persistence for browser-based play on desktop and mobile.

Read more

Ada-MK: Adaptive MegaKernel Optimization via DAG-Based Search for LLM Inference

Ada-MK optimizes LLM decode-phase latency by fusing operators into a single persistent MegaKernel, addressing the overhead of frequent kernel launches and HBM round-trips. It utilizes an MLIR-based offline DAG search to eliminate runtime branching and a 3D shared-memory constraint model to reduce peak memory usage by 50%. Integrated as a TensorRT-LLM plugin, Ada-MK improves single-batch throughput on NVIDIA L20 GPUs by up to 23.6% over vanilla TensorRT-LLM and 50.2% over vLLM.

ArXiv

δ-mem: Efficient Online Memory for Large Language Models

$\delta$-mem is a lightweight memory mechanism designed for LLMs to accumulate and reuse historical information, bypassing the costs and inefficiencies of expanding context windows. It augments a frozen full-attention backbone with a compact online associative memory state, which compresses past information into a fixed-size matrix using delta-rule learning. This state then generates low-rank corrections to the backbone's attention computation during generation. With only an $8\times8$ memory state, $\delta$-mem achieves significant performance improvements on memory-heavy benchmarks (e.g., $1.31\times$ on MemoryAgentBench) while preserving general capabilities, demonstrating effective memory through direct attention coupling without full fine-tuning or explicit context extension.

ArXiv

Self-Distillation Enables Continual Learning [pdf]

Self-Distillation Fine-Tuning (SDFT) addresses catastrophic forgetting in foundation models by enabling on-policy learning from demonstrations via in-context learning. By using the model as its own teacher to generate training signals, SDFT preserves prior capabilities better than off-policy SFT during sequential skill and knowledge acquisition. This approach allows for the accumulation of multiple skills without performance regression, offering a practical path for continual learning.

ArXiv

The Impossibility of Supersized Machines (2017)

Despite claims that machines could match or surpass human size, seven distinct arguments demonstrate that exceeding human physical dimensions is fundamentally impossible.

ArXiv

A Single Neuron Is Sufficient to Bypass Safety Alignment in LLMs

Safety alignment in LLMs is mediated by specific refusal and concept neurons rather than being robustly distributed across weights. Researchers demonstrated that suppressing a single refusal neuron can bypass safety filters, while amplifying concept neurons can induce harmful outputs from benign prompts across models up to 70B parameters. These findings suggest that safety mechanisms are fragile, as individual neurons are causally sufficient to gate refusal behavior.

ArXiv

Codiff, a local diff review tool

Codiff is a minimal local Git diff viewer designed for pre-commit code reviews. It features LLM-integrated walkthroughs using Codex to provide contextual review orders and supports inline commenting with Markdown export for follow-ups.

Repo

Strava for AI coding – analytics on your Copilot/Claude/Codex usage

AI Engineer Coach is an open-source VS Code extension that provides local analytics and insights for AI coding assistant usage. It evaluates session logs to detect anti-patterns, measure code output by model, and assess context health and agentic readiness. The tool features a rule engine for custom prompt analysis and a skill finder to optimize LLM interactions while ensuring all data remains on the user's machine.

Repo

How-to-train-your-GPT. Every line commented

"How to Train Your GPT" is a 12-chapter technical guide for building a LLaMA 3-style LLM from scratch using Python. It provides annotated implementations of modern architectural components, including RoPE, RMSNorm, SwiGLU, and causal multi-head attention. The curriculum covers the full pipeline from BPE tokenization and training with AdamW to optimized inference utilizing KV caching and various sampling strategies.

Repo

A Compression Tool for LLM Reads. Est. 60-95% Fewer Tokens

Headroom is a context compression layer for AI agents and LLMs, designed to reduce token usage by 60-95% without sacrificing accuracy. It compresses various inputs like tool outputs, logs, RAG chunks, files, and conversation history before they reach the LLM. Headroom offers multiple deployment modes, including a library, proxy, and agent wrappers, utilizing algorithms like SmartCrusher (JSON), CodeCompressor (AST), and Kompress-base (text). A key feature is its Reversible Context Compression (CCR), allowing LLMs to retrieve original content on demand, and it supports cross-agent memory.

Repo

AI Memory Reader – Native macOS app for browsing Claude Code memory files

AI Memory Reader is a native macOS/iOS application for browsing, reading, and editing AI agent memory files, compatible with systems like Claude Code, OpenClaw, Codex, Cursor, and Gemini. It features beautiful Markdown rendering, auto-discovery of AI sources, full-text search, and an integrated editor with syntax highlighting. The app also provides a URL scheme and CLI for AI agents to programmatically open and navigate specific memory files and headings.

Repo