<h2>Must know</h2>
<ul>
<li><strong>Google releases Gemma 4 open weight multimodal models</strong>: Google has launched Gemma 4, a family of natively multimodal open-weight models under an Apache 2.0 license, supporting vision, audio, and agentic reasoning. The lineup includes a 26B MoE workstation model and edge-optimized variants (E2B, E4B) that fit on consumer hardware like the Jetson Orin Nano. According to Google and NVIDIA, these models feature a 256K context window and are optimized for on-device deployment via vLLM, Ollama, and a new Docker model runner. <a href="https://app.daily.dev/feed-by-ids?id=DxUdrK5VU&amp;id=hi4RmWdI4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Meta introduces semi-formal reasoning for code verification</strong>: Meta researchers have developed a technique that allows AI agents to verify code correctness without execution by constructing explicit premises and tracing execution paths. This semi-formal approach improved patch equivalence verification accuracy from 78% to 93% and boosted fault localization on Defects4J. For DevOps teams, this suggests a future of cheaper, execution-free reinforcement learning reward signals that bypass expensive sandbox-based training pipelines. <a href="https://app.daily.dev/posts/gfdiiQquJ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Supply chain attacks hit Mercor and Axios via LiteLLM</strong>: The AI recruiting startup Mercor, valued at $10 billion, confirmed a security breach linked to a supply chain attack on the open-source LiteLLM project. Simultaneously, the Axios npm package was hijacked by North Korean threat actor UNC1069, who embedded a Trojan to harvest AWS keys and GitHub tokens. These incidents, reported by Axios and Socket, underscore the systemic vulnerability of AI-coding workflows that rely on unpinned third-party dependencies. <a href="https://app.daily.dev/feed-by-ids?id=1i5FA0Q5W&amp;id=YtEQA7dzF&amp;id=sjXZykGnw" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Clerk Core 3 expands free tier to 50k users and adds React Router support</strong>: The update introduces a unified Show component, keyless mode for TanStack Start, and modern React features like concurrent rendering and suspense. <a href="https://app.daily.dev/feed-by-ids?id=B73gISWRG&amp;id=4eb7ud3kg" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vercel Sandbox restores now hit sub-second speeds via parallel S3 streaming</strong>: Engineering optimizations using HTTP Range requests and an LRU NVMe cache reduced p75 restore times from 40 seconds to under 1 second. <a href="https://app.daily.dev/posts/1Xib7hCkG" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Model Context Protocol gains momentum with pgEdge and Waldium integrations</strong>: New MCP servers for Postgres and agentic CMS platforms allow AI agents to query live data with 30-50% less token usage via TSV output. <a href="https://app.daily.dev/feed-by-ids?id=09ZKKSp9j&amp;id=sCEi0lM0l" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Pretext library enables DOM-free text height calculation via Canvas API</strong>: Created by Cheng Lou, this minimal JS library emulates word-wrapping off-screen for high-performance dynamic layouts and RTL support. <a href="https://app.daily.dev/posts/rS0vBup8W" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux kernel v4 patches introduce cache-aware scheduling for Xeon and EPYC</strong>: The update aims to reduce cache misses by colocating tasks within the same last-level cache domain on modern multi-cache CPUs. <a href="https://app.daily.dev/posts/3Go8jplR2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Jujutsu VCS v0.40.0 adds diff-content matching and TUI improvements</strong>: The release includes new revset functions for matching added/removed lines and smarter swap actions in the ‘jj arrange’ interface. <a href="https://app.daily.dev/posts/XNj5BzBCW" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Landdown utility sandboxes Linux shell scripts using Landlock kernel features</strong>: The tool enforces explicit allowlists for file system and network access by prepending a custom shebang to scripts. <a href="https://app.daily.dev/posts/NkuXzLASM" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>The AI memory crisis is the new ‘the dog ate my homework’</strong>: Projected DRAM price hikes of 100% are being blamed on OpenAI’s Stargate gobbling up 40% of global output, according to some reports. While it’s a convenient excuse for GPU price gouging, it’s mostly a narrative to justify why your next laptop will cost 20% more while still shipping with a measly 8GB of VRAM. <a href="https://app.daily.dev/feed-by-ids?id=pR2QdYBDR&amp;id=jmoYHAJ6v" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anthropic’s ‘Undercover Mode’ is a transparency nightmare</strong>: The leaked Claude Code source revealed an ‘Undercover Mode’ designed to strip AI attribution from commits. Anthropic’s marketing team loves to talk about ‘AI Safety,’ but building features specifically to help employees (or anyone else) gaslight their peers about who wrote the code is the peak of corporate hypocrisy. <a href="https://app.daily.dev/feed-by-ids?id=CGbQv469o&amp;id=lSeOtUf5e" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Startup methodologies are officially a self-defeating prophecy</strong>: A recent critique argues that 25 years of Lean Startup and Design Thinking have done exactly zero to improve survival rates. When every founder follows the same ‘exact blueprint’ from a YC blog post, they all converge on the same mediocre strategy, effectively neutralizing any competitive advantage they hoped to gain. <a href="https://app.daily.dev/feed-by-ids?id=N6roJjjoX&amp;id=eWWWY2wcm" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
