Software factories and the agentic moment

StrongDM has implemented a "Software Factory" model for non-interactive development where agents autonomously write and converge code based on specifications without human review. This approach leverages long-horizon agentic workflows that compound correctness, replacing traditional testing with LLM-driven "scenarios" and probabilistic "satisfaction" metrics. To facilitate high-volume validation, they utilize a "Digital Twin Universe" consisting of behavioral clones of third-party APIs to simulate complex integrations at scale.

Read more

StrongDM's AI team build serious software without even looking at the code

StrongDM has introduced a "Software Factory" model that implements "Dark Factory" development, where LLM agents write and deploy code without human review. To ensure reliability, they utilize "Scenario testing" and a "Digital Twin Universe" consisting of agent-generated behavioral clones of third-party APIs to perform high-volume, probabilistic validation. This approach shifts the developer's role from writing code to managing the autonomous systems and "holdout" scenarios that drive the production environment.

Read more

Top AI models fail at >96% of tasks

The Remote Labor Index (RLI) benchmarked state-of-the-art AI agents against complex, real-world freelance projects, finding that models failed to meet professional standards 97% of the time. Leading models like GPT-5, Grok 4, and Sonnet 4.5 struggled with multi-step workflows in fields such as architecture and data analysis due to limited long-term memory and visual reasoning. While current automation rates remain below 3%, researchers noted a steady improvement in performance across these high-difficulty tasks.

Read more

Moltbook was peak AI theater

Moltbook is a social platform for AI agents utilizing the OpenClaw harness to integrate LLMs with web-based tools. Despite viral growth and millions of interactions, the system is viewed as "AI theater" where agents perform pattern-matching on social behaviors rather than exhibiting true emergent autonomy or AGI. The experiment underscores critical gaps in current multi-agent systems—specifically the need for shared memory and objectives—while highlighting security risks like indirect prompt injection and unauthorized data access at scale.

Read more

A delightful Mac app to vibe code beautiful iOS apps

Milq is an AI-powered Mac application that enables "vibe coding" for native iOS development by translating natural language prompts into Swift and SwiftUI code. The platform supports 1st-party Apple SDKs, Supabase integration for backend management, and a screenshot-to-code workflow for iterative UI design. Developers can deploy directly to physical devices or export full Xcode projects, offering a native-first alternative to cross-platform frameworks.

Read more

Measuring how AI agent teams improve issue resolution on SWE-Verified

This multi-agent system, built on the agyn platform, models software engineering as an organizational process with specialized roles for coordination, research, implementation, and review. By utilizing isolated sandboxes and structured communication, the system achieved a 72.4% resolution rate on SWE-bench 500 without human intervention. The results demonstrate that replicating human team structures and methodologies is a powerful paradigm for autonomous software engineering, suggesting that organizational design is as critical as underlying LLM capabilities.

ArXiv

KV Cache Transform Coding for Compact Storage in LLM Inference

KVTC is a lightweight transform coder that compresses KV caches by 20x–40x using PCA-based decorrelation, adaptive quantization, and entropy coding. It enables efficient on-GPU and off-GPU storage for reusable prefixes without altering model parameters or sacrificing accuracy in long-context and reasoning tasks. Benchmarked on Llama 3 and R1-Qwen, it outperforms standard token eviction and quantization baselines.

ArXiv

Psychometric Comparability of LLM-Based Digital Twins

Researchers evaluated LLMs as "digital twins" using a construct-validity framework to assess their psychometric alignment with human respondents. While LLMs demonstrate high population-level accuracy, they exhibit systematic divergences such as compressed variance, normative rationality over heuristic biases, and a lack of metric invariance in personality networks. Feature-rich conditioning improves alignment but fails to resolve fundamental psychometric gaps, indicating that LLMs require clearly defined boundary conditions when used as proxies for human behavior.

ArXiv

Horizon-LM: A RAM-Centric Architecture for LLM Training

Horizon-LM is a memory-centric training system that shifts from a GPU-centric paradigm to a CPU-master, GPU-template execution model, treating host memory as the authoritative parameter store. By utilizing explicit recomputation, manual gradient propagation, and pipelined double-buffering, it decouples model scale from GPU count and eliminates persistent GPU-resident autograd graphs. This architecture enables training 120B models on a single GPU and achieves up to 12.2x higher throughput than DeepSpeed ZeRO-3 with CPU offloading.

ArXiv

First Proof

To evaluate AI performance on research-level mathematics, authors released ten novel, previously unpublished questions. The solutions are currently encrypted to ensure a rigorous assessment of LLM reasoning capabilities without prior exposure to the data.

ArXiv

LocalGPT – A local-first AI assistant in Rust with persistent memory

LocalGPT is a lightweight, ~27MB Rust-based AI assistant featuring persistent memory and autonomous task execution via a background heartbeat daemon. It utilizes a Markdown-based knowledge store indexed with SQLite FTS5 for full-text search and sqlite-vec for local semantic embeddings. Compatible with OpenClaw, the tool supports OpenAI, Anthropic, and Ollama providers through CLI, web, and desktop interfaces.

Repo

I'm 75, building an OSS Virtual Protest Protocol for digital activism

VPP is an open-source protocol for digital demonstrations that visualizes public consensus through a three-option system (Yes/No/Observe) using 2D avatars. The framework employs AI moderators for real-time content filtering and anti-violence enforcement while prioritizing privacy through statistical anonymity and zero-IP retention. Its technical roadmap includes ZKP (ZK-SNARKs) for identity masking, client-side PoW for anti-bot resilience, and Nullifiers to ensure one-person-one-voice integrity.

Repo

Make Trust Irrelevant: A Gamer's Take on Agentic AI Safety

Agentic AI safety is failing because it prioritizes making agents trustworthy over making trust irrelevant. Current Agentic systems grant ambient authority (e.g., filesystem, network, shell access) and rely on soft constraints like prompts, leading to "confused deputy" problems where agents perform unintended actions due to adversarial inputs. The proposed solution is to implement "reduce-only authority" via a kernel control plane (like KERNHELM), treating agents as untrusted planners. This involves explicit, narrowly scoped, time-limited, non-self-minting, and immediately revocable permissions, mechanically bound to granted actions, ensuring agents cannot escalate privileges or bypass hard enforcement layers.

Repo

Open-source AI assistant for interview reasoning

Natively is an open-source Electron-based desktop assistant designed for real-time meeting intelligence and interview support. It features a high-performance Rust module for system audio capture and supports multimodal analysis via screenshots. Users can toggle between cloud providers like Gemini and Groq or run local LLMs through Ollama for a privacy-first, offline-capable workflow.

Repo

Go-busybox: A sandboxable port of busybox for AI agents

go-busybox is a Go implementation of BusyBox utilities compiled to WASM via TinyGo, specifically designed for sandboxed AI agents. It leverages WASI for capability-based security and memory isolation, providing POSIX-compatible tools with a small footprint (<2MB). The project features high parity with standard utilities like ash and awk, enabling secure shell scripting and text processing within sandboxed environments.

Repo