<h2>Must know</h2>
<ul>
<li><strong>Moonshot AI launches new agentic reasoning model</strong>: Moonshot AI has released Kimi K2 Thinking, an open-source AI model designed for reasoning and agentic tasks. This model, featuring a mixture-of-experts architecture, can autonomously execute hundreds of sequential tool calls, enabling extended operations without human intervention. While Moonshot AI claims state-of-the-art performance on benchmarks like Browse Comp and SWE-bench, these results lack independent verification, so take them with a grain of salt.</li>
<li><strong>AI agent development ecosystem expands rapidly</strong>: The AI agent development landscape is seeing rapid expansion with new tools and frameworks. Zed now offers one-click installation for coding agents, while <a href="http://Builder.io" target="_blank" rel="noopener nofollow">Builder.io</a>’s Fusion 1.0 aims to unify product workflows. Google’s Agent Development Kit (ADK) adds user simulation for robust testing, and Vercel has published methodologies for identifying effective AI agent use cases, emphasizing human-in-the-loop oversight for automation. This indicates a strong push towards practical, deployable AI agents.</li>
<li><strong>Keycloak 26.4 brings passwordless auth and FAPI 2.0</strong>: Keycloak 26.4 is out, introducing full Passkeys support for passwordless authentication and enhancing security with FAPI 2.0 and DPoP protocols. This release also improves multi-availability zone deployments and allows applications to authenticate using Kubernetes service account tokens or SPIFFE, eliminating the need for manual credential management. Future plans include better support for AI agents and enhanced token exchange capabilities.</li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>OpenTelemetry advances for mobile and MLflow</strong>: OpenTelemetry is tackling mobile telemetry challenges with new instrumentation for ANR errors and network calls, while MLflow 3.6.0 now supports full OpenTelemetry tracing for GenAI applications, unifying data across multiple languages.</li>
<li><strong>Rails adds Argon2 hashing and framework enhancements</strong>: Rails now supports Argon2 password hashing, removing the 72-byte password length restriction, alongside new Engines guides, improved internationalization, and schema caching for easier upgrades.</li>
<li><strong>OCaml gains traction for reliability and type safety</strong>: OCaml is being adopted for its balance of expressiveness and simplicity, with its strong type system preventing state corruption and ensuring robust error handling, making certain classes of failures impossible.</li>
<li><strong>Mesa Vulkan drivers fix HDR and Proton issues</strong>: Mesa 26.0-devel includes critical HDR bug fixes for Vulkan drivers, and Mesa 25.3-rc4 resolves rendering issues for Direct3D games on Linux via Steam Play/Proton, improving gaming experiences on AMD and Intel hardware.</li>
<li><strong>Developer workflow trends: faster, more automated</strong>: GitHub Octoverse 2025 reports a rapid shift to smaller, more frequent code updates, with widespread adoption of feature flags, CI/CD, and automated testing, while human-centered DevOps practices remain crucial.</li>
<li><strong>Historic Unix Version 4 tape discovered</strong>: A 1973 magnetic tape containing Unix Version 4, the first Unix kernel rewritten in C, was discovered at the University of Utah and is now headed to the Computer History Museum for preservation and analysis.</li>
<li><strong>Go 1.25.4-2 and 1.24.10-1 Microsoft builds available</strong>: Microsoft has released new builds of Go versions 1.25.4-2 and 1.24.10-1, which are Microsoft-specific distributions of the Go programming language with updates corresponding to upstream releases.</li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>AI benchmarking: more marketing than science</strong>: A new Oxford Internet Institute study reveals that only 16% of LLM benchmarks use rigorous scientific methods, with many failing to define basic concepts or relying on convenience sampling. Yet, AI companies like OpenAI continue to heavily market these ‘scores,’ misleading everyone about actual model capabilities. It’s almost as if they’re more interested in PR than actual progress.</li>
<li><strong>Meta’s $16 billion scam ad problem</strong>: Reuters reports Meta is raking in an estimated $16 billion annually from scam advertisements, serving 15 billion high-risk ads daily, and even allowing repeat offenders back on the platform. So, while they’re busy telling us how much they care about user safety, their bottom line seems to be saying something else entirely. Priorities, right?</li>
<li><strong>Rockstar Games lays off union members, blames leaks</strong>: Rockstar Games laid off 30-40 employees, claiming it was due to confidential information leaks. However, the Independent Workers’ Union alleges these layoffs targeted union members for organizing, not for leaking. And, oh, by the way, GTA VI got delayed. Coincidence? I’m sure it’s just a happy little accident that all the ‘leakers’ were also union organizers.</li>
</ul>
