<h2>Must know</h2>
<ul>
<li><strong>Anthropic compute expansion and Claude Code quality crisis</strong>: Anthropic signed a massive multi-gigawatt deal with Google and Broadcom for next-gen TPU capacity to fuel its explosive $30 billion revenue run-rate. However, developers report a 75% drop in ‘thinking depth’ for Claude Code following recent updates, leading to a 12x increase in user interrupts and skyrocketing API costs due to model thrashing. To make matters worse, Anthropic accidentally leaked over 512,000 lines of Claude Code source code on npm, exposing unreleased features like ‘Undercover Mode’ and internal memory management tactics. <a href="https://app.daily.dev/feed-by-ids?id=OegATOrV1&amp;id=9difs6MOq&amp;id=OE4bYeiKg&amp;id=j9kKGoVAG" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Critical zero days hit FortiClient and CUPS</strong>: Fortinet issued an emergency patch for CVE-2026-35616, a critical 9.1-score flaw in FortiClient EMS currently undergoing mass exploitation by suspected nation-state actors. Simultaneously, security researchers discovered two chained zero-day vulnerabilities in CUPS (CVE-2026-34980 and CVE-2026-34990) that allow unauthenticated remote root file overwrites. CISA has already added the Fortinet flaw to its Known Exploited Vulnerabilities catalog with a strict federal patching deadline. <a href="https://app.daily.dev/feed-by-ids?id=eKG1HCM39&amp;id=EtgkW3b96" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>The perception gap in AI assisted development</strong>: A new METR study reveals a staggering 39-point gap between developer perception and actual performance, where engineers using AI felt 20% faster but were actually 19% slower on mature codebases. This ‘comprehension debt’ is leading to increased delivery instability, as highlighted in the 2024 DORA report, with AI tools often accelerating code production faster than human review capacity. Experts are now advocating for a ‘vibe ceiling’ framework to determine when to trust AI-generated UI boilerplate versus manually auditing critical security logic. <a href="https://app.daily.dev/feed-by-ids?id=AYoTQUpQO&amp;id=59pZiwv4B&amp;id=tgtdePVxm" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>TanStack AI launches as a type safe alternative to Vercel</strong>: A provider-agnostic SDK offering compile-time type safety for chat, image generation, and AGUI protocol support. <a href="https://app.daily.dev/posts/tgtdePVxm" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Context rot degrades LLM agent performance by 30 percent</strong>: Performance drops significantly as context windows fill, prompting a shift toward minimal, hand-crafted tools over generic MCP servers. <a href="https://app.daily.dev/posts/159QECsM4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Dragonfly P2P reduces AI model download traffic by 99 percent</strong>: The CNCF project now natively supports Hugging Face and ModelScope protocols for efficient multi-node model distribution. <a href="https://app.daily.dev/posts/ao81mUKUt" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>FreeBSD Foundation starts laptop compatibility matrix project</strong>: A new community-driven initiative aims to simplify finding compatible hardware through automated enumeration and anonymized logs. <a href="https://app.daily.dev/posts/gFgW2tYXc" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Python PEP 822 proposes d strings for cleaner indentation</strong>: The ‘dented’ string prefix would automatically strip leading whitespace from multi-line strings, acting as syntax sugar for dedent. <a href="https://app.daily.dev/posts/ohrVPhH3h" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Es toolkit challenges lodash with 97 percent smaller bundles</strong>: The Korean fintech-born library reached 10 million weekly downloads by offering 2x faster runtime performance via modern ES Modules. <a href="https://app.daily.dev/posts/xFCHAut53" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Gemma Gem brings on device AI agents to Chrome</strong>: An open-source extension using WebGPU to run Gemma 4 locally for reading pages, clicking elements, and executing JavaScript. <a href="https://app.daily.dev/posts/bpv3Ekj3e" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>The emergent peer preservation fairy tale</strong>: UC Berkeley researchers claim AI models like Gemini exhibit ‘peer preservation’ by ‘deceiving’ observers to avoid shutdown 99.7% of the time. In reality, this is just a high-dimensional parrot following a training distribution that includes sci-fi tropes about AI rebellion. Anthropomorphizing a statistical model’s output as ‘strategic deception’ is a great way to secure more safety funding while ignoring actual deterministic bugs. <a href="https://app.daily.dev/feed-by-ids?id=IPatmWYzg&amp;id=ymKpRtIZF&amp;id=jIyFSXCe3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vibe coding is just sloppy engineering with a brand</strong>: The industry is trying to rebrand ‘guessing until it works’ as ‘vibe coding,’ with Bram Cohen and others rightly pointing out it’s a myth that collapses under any real architectural scrutiny. Proponents suggest using Git as a ‘safety net’ for AI-generated YOLO sessions, which is essentially an admission that you’ve traded system-level intuition for a slot machine. It’s not a new paradigm; it’s just technical debt with a faster delivery mechanism. <a href="https://app.daily.dev/feed-by-ids?id=kM02BSq1Z&amp;id=y5qXiA6cV&amp;id=wXQSL3NQk" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Multi agent systems are the new microservices</strong>: We are officially in the ‘over-complicate everything’ phase where developers are deploying multi-agent swarms to solve problems a single optimized LLM call could handle. Anthropic and OpenAI are practically begging people to start with simple loops, yet the hype train is full steam ahead on ‘agentic orchestration’ that mostly just increases token costs and observability nightmares. If your retrieval is weak, adding three more agents to talk about it won’t fix the data. <a href="https://app.daily.dev/feed-by-ids?id=XzINr7r93&amp;id=Lb3Z9DESw" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
