Skip to content

Four Signals

I Ran Hermes Agent on the Same Task for 7 Days. The Skill File on Day 7 Looked Nothing Like Day 1.
ai/ml / Dev.to

I Ran Hermes Agent on the Same Task for 7 Days. The Skill File on Day 7 Looked Nothing Like Day 1.

Hermes Agent's persistent skill file evolved from a 12-line placeholder to a 60-line intelligent procedure after 7 daily runs of an AI news digest task. It learned to avoid low-quality sources like TechCrunch, prioritize Hacker News and GitHub Releases, and generated summaries with technical context (e.g., noting break…

Why it matters

For senior engineers building automated workflows, this demonstrates how agents can accumulate domain-specific knowledge across sessions, reducing manual tuning and improving output quality over time.
ai/ml / Hacker News (100+)

SANA-WM, a 2.6B open-source world model for 1-minute 720p video

NVIDIA Research has released SANA-WM, a 2.6 billion parameter open-source world model that can generate 1-minute long 720p videos. This appears to be a significant open-weight model for video simulation and generation, accessible via the nvlabs GitHub page. For an AI/ML engineer focusing on agent orchestration and simulation, this open-source world model could serve as a powerful, free backbone for creating training environments, testing multi-agent interactions, or generating synthetic video data at scale. Evaluate SANA-WM for integrating into your agent simulation pipelines or video generation workflows, given its size and open license.
OpenAI co-founder Greg Brockman reportedly takes charge of product strategy
ai/ml / TechCrunch

OpenAI co-founder Greg Brockman reportedly takes charge of product strategy

Greg Brockman, OpenAI co-founder and president, assumes product strategy, planning to merge ChatGPT and Codex into one platform as part of Sam Altman's "code red" refocus. The shift occurred while CEO of AGI deployment Fidji Simo is on medical leave, and Brockman aims to win in both consumer and enterprise by consolidating efforts and halting side projects like Sora. This aligns with OpenAI's push toward an agentic future and a super app. For a senior engineer building agentic systems, this signals OpenAI is betting on unified experiences and API convergence, potentially reshaping how you integrate LLMs into developer tools and enterprise workflows. Monitor the ChatGPT-Codex merger to anticipate shifts in AI agent orchestration patterns and API surface consolidation.
languages / Lobsters

My Thoughts on Bun's Rust Rewrite

Bun's production JS runtime, originally built on Zig for rapid prototyping and performance, was entirely rewritten in Rust via Claude-generated code in a 6-day merge (6,755 commits) with zero human review. The author argues this violates maintainability: AI ensures local semantic equivalence but misses global invariants that exist only in the original author's head, and the post-acquisition team now inherits risks previously borne by founder Jarred. Zig's foundational architecture—low-level memory manipulation and C interop—enabled Bun's success, but the rewrite is a pragmatic business decision, not a technical indictment of Zig. This case exposes the real-world risks of AI-driven code rewrites at scale—especially the gap between passing tests and preserving cross-module invariants, which is critical for senior engineers building or maintaining AI-generated production systems. Audit AI-generated code for global invariants and design constraints, not just test pass rates, and never merge without human review.
security / Techmeme

On Pwn2Own Berlin 2026 day 2, competitors earned $385,750 after exploiting 15 unique zero-day vulnerabilities in Windows 11, Red Hat Enterprise Linux and more (Sergiu Gatlan/BleepingComputer)

On day 2 of Pwn2Own Berlin 2026, security researchers earned $385,750 by successfully demonstrating 15 unique zero-day exploits against Windows 11, Red Hat Enterprise Linux, and other systems. This highlights critical vulnerabilities in major enterprise operating systems that could impact cloud and developer infrastructure. The discovered zero-days in Windows 11 and RHEL directly affect the enterprise and cloud environments you work with, potentially exposing your infrastructure to unpatched exploits. Monitor Pwn2Own disclosures and apply vendor patches promptly to protect your Windows 11 and RHEL deployments from these demonstrated zero-day vulnerabilities.
Microsoft Releases Aspire 13.3 with Major Deployment and Frontend Updates
cloud / InfoQ

Microsoft Releases Aspire 13.3 with Major Deployment and Frontend Updates

Microsoft released Aspire 13.3 with a new aspire destroy command for resource cleanup across Azure, Kubernetes, and Docker Compose, plus a NativeAOT CLI and standalone aspire dashboard. Kubernetes deployment is in preview with Ingress/Gateway API routing and an AKS hosting integration labeled 'Kubernetes without YAML'. Frontend gains first-class JavaScript publishing via unified PublishAs* methods for Next.js, Vite, Bun, Yarn, and pnpm, alongside TypeScript AppHost parity, browser console logging captures, and container tunnel enabled by default; breaking changes include --log-level rename and removal of the GitHub Copilot UI. For a senior engineer building cloud-native .NET apps with multi-agent orchestration, Aspire 13.3's Kubernetes preview and unified JavaScript publishing streamline deployment and frontend integration, reducing boilerplate and enabling cleaner CI/CD pipelines. Upgrade to Aspire 13.3 and adopt the new aspire destroy command for ephemeral environment teardown and the Kubernetes deployment preview with AKS integration to reduce YAML overhead.
Legible - I built an on-device document helper for immigrants using Gemma 4
general / Dev.to

Legible - I built an on-device document helper for immigrants using Gemma 4

Legible is an on-device document helper for immigrants that uses Gemma 4 E4B (via Ollama) to process photos of official documents locally—no cloud API, no data leaving the machine. The FastAPI backend streams structured XML output from the model, providing plain-language explanations, deadline countdowns, and next steps in 11 languages. Past scans are stored as Fernet-encrypted files, and the frontend is a single HTML file with no build step. This demonstrates a practical, privacy-first architecture for offline multimodal LLM applications using lightweight models, directly relevant to building secure, accessible AI tools for sensitive user data. Evaluate Gemma 4 via Ollama for on-device document understanding; its structured output capability and offline operation make it ideal for privacy-critical multilingual apps.
general / Lobsters

Git Is Not Fine

Git excels as a distributed source store but fails as a distributed workflow tool, especially for stacked PRs common in async development. The immutable commit model and poor support for commit relationships make rebasing and maintaining stacked branches painful. Tools like jj address these gaps with a different data model that preserves history relationships and simplifies pipelining work. For a senior engineer building developer tooling or optimizing team workflows, this analysis highlights a critical pain point in daily git usage that newer tools like jj aim to solve. Evaluate jj for stacked PR workflows to reduce the cognitive overhead of git rebase and merge operations.