<h2>Must know</h2>
<ul>
<li><strong>Flash MoE brings 400B parameter models to consumer MacBooks</strong>: A new inference engine called Flash-MoE, written in pure C and Metal, successfully runs the Qwen3.5-397B Mixture-of-Experts model on a MacBook Pro with only 48GB of RAM. By streaming the 209GB model from SSD and loading only four active experts per layer into memory, it achieves usable speeds of over 4 tokens per second. The project, detailed on GitHub, leverages Apple’s Accelerate BLAS and parallel SSD reads to bypass traditional RAM bottlenecks for massive models. <a href="https://app.daily.dev/posts/xsGaEEFUf" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cursor admits Composer 2 is built on Moonshot AI base</strong>: Cursor’s new Composer 2 model, marketed as a frontier-level coding breakthrough, was revealed by independent researchers to be built on Kimi k2.5 from Chinese firm Moonshot AI. While Cursor’s VP of education confirmed the partnership via Fireworks AI, the initial lack of disclosure has sparked geopolitical concerns regarding U.S. developer tools relying on Chinese foundational models. Cursor claims that three-quarters of the training compute was their own, resulting in distinct benchmark performance from the base model. <a href="https://app.daily.dev/posts/2O2HmWwMM" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>CERN burns anomaly detection AI directly into silicon</strong>: To manage a staggering 40,000 exabytes of annual sensor data, CERN is using a transpiler called HLS4ML to bake machine learning models directly into FPGA and ASIC hardware. Their AXOL1TL algorithm must execute in under 50 nanoseconds to decide which collision events to keep, rejecting 99.7% of data on-chip. Interestingly, CERN found that aggressively quantized tree-based models outperform deep learning for this specific high-speed tabular data task. <a href="https://app.daily.dev/posts/ibpFpiXCR" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Vercel Workflow adds automatic end to end encryption</strong>: The platform now uses per-deployment unique keys and AES-256-GCM to encrypt all step arguments and return values by default. <a href="https://app.daily.dev/posts/0r6jV9CyY" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux 7.0 rc5 stabilizes ahead of April release</strong>: Linus Torvalds noted the cycle is calming down, featuring fixes for Radeon GCN 1.0 GPUs and improved Logitech Bluetooth support. <a href="https://app.daily.dev/posts/1tWnyX7nj" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GrapheneOS refuses to comply with age verification laws</strong>: The privacy-focused Android fork will exit markets like Brazil and California rather than implement mandatory OS-level age tracking APIs. <a href="https://app.daily.dev/feed-by-ids?id=3oUfmyp1R&amp;id=f2c1Srdfh" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>MkDocs ecosystem fractures after founder return</strong>: The popular documentation tool has split into three competing forks—ProperDocs, MaterialX, and Zensical—following a controversial private rewrite by the original author. <a href="https://app.daily.dev/posts/GedRmjUpE" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Project Detroit reopens to bring Python to the JVM</strong>: The initiative aims to use the Foreign Function and Memory API to integrate JavaScript and Python directly into the Jakarta EE ecosystem. <a href="https://app.daily.dev/posts/24VxbiILg" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Logitech mouse DPI registers used for persistent data storage</strong>: A developer reverse-engineered the HID++ protocol to store 2 bytes of cross-computer persistent data inside a mouse’s DPI register. <a href="https://app.daily.dev/posts/uG0HL5oPv" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ZeroPath brings AI native exploitability analysis to RSAC</strong>: The startup uses AST-based graphs and LLMs to verify if vulnerabilities are actually reachable before flagging them to developers. <a href="https://app.daily.dev/posts/5p9F6EuUW" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>The 85 trillion dollar token compensation trap</strong>: Jensen Huang is now suggesting that AI tokens should make up half of an engineer’s salary, a convenient way to inflate ‘total comp’ with company-store credits that don’t vest or appreciate. Critics at TechCrunch rightly point out this is just a clever way to pressure staff into superhuman productivity while the company claws back the ‘salary’ as compute spend. It’s not a bonus; it’s an electricity bill you’re being asked to pay with your own labor. <a href="https://app.daily.dev/posts/Ryzsn8a9R" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vibe coding and the death of the software craft</strong>: A wave of ‘thought pieces’ is mourning the loss of coding as an art form, claiming LLMs have turned programming into a ‘fast-fashion’ assembly line. While the nostalgia for hand-rolled loops is touching, these arguments often ignore that most enterprise code was already a soul-crushing derivative of a StackOverflow snippet. The ‘grief’ isn’t about the technology; it’s about the realization that your ‘bespoke’ CRUD app was always just a statistical average. <a href="https://app.daily.dev/feed-by-ids?id=AKKujKavD&amp;id=Y5vSQxzM2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Deno CEO goes missing as the hype ship sinks</strong>: After raising $26M to ‘fix’ Node.js, Deno is facing mass layoffs and a community wondering where the leadership went. Between the failed Deno Deploy and the JSR registry that nobody asked for, the project is a masterclass in solving problems that developers didn’t actually have. Now the only thing left to ‘deploy’ is a pivot to AI or a quiet acquisition by someone with more cash than sense. <a href="https://app.daily.dev/posts/ONnCa8fkU" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
