<h2>Must know</h2>
<ul>
<li><strong>Anthropic launches Programmatic Tool Calling to kill token bloat</strong>: Anthropic introduced Programmatic Tool Calling (PTC) for Claude Opus and Sonnet 4.6, allowing the model to execute complex orchestrations within a sandboxed container rather than piping every intermediate result back into the context window. This architectural shift reduces input tokens by 24% and improved search accuracy by 11% in internal benchmarks. By keeping intermediate data out of the main prompt, PTC preserves the control surface for human approval while gaining the composability of raw code. <a href="https://app.daily.dev/posts/0Sd168FyP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AWS data centers hit by drone strikes in Middle East</strong>: Amazon Web Services facilities in the UAE and Bahrain suffered structural damage and power disruptions following Iranian drone and missile attacks on March 3, 2026. This marks the first time major US cloud infrastructure has been directly disrupted by military action, forcing customers to migrate workloads to unaffected regions. The incident highlights the physical vulnerability of global compute clusters in volatile geopolitical zones. <a href="https://app.daily.dev/posts/rar5XOK6I" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Malicious Trivy extension spawns unattended AI agents</strong>: Security researchers at Socket discovered malicious code in versions 1.8.12 and 1.8.13 of the Aqua Trivy extension on OpenVSX, which silently launched AI coding assistants in fully permissive modes. The exploit used a compromised publisher account to exfiltrate credentials and system reconnaissance data via the victim’s authenticated GitHub CLI. Developers who installed these versions are urged to rotate all tokens and audit their AI agent logs for unauthorized autonomous actions. <a href="https://app.daily.dev/feed-by-ids?id=1Sk4jyYwp&amp;id=DGWq9i7NL" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>CtrlAI provides a transparent proxy for securing non-deterministic AI agents</strong>: This open-source MIT-licensed proxy intercepts tool calls to enforce guardrails like blocking SSH key access without requiring code changes in the agent SDK. <a href="https://app.daily.dev/posts/f6EeImc9b" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Matrix Rust SDK achieves 98% speedup via Data-oriented Design</strong>: By replacing large 144-byte structs with a compact 64-byte cache and eliminating lock contention, room list execution time dropped from 53ms to 676µs. <a href="https://app.daily.dev/posts/3QNJOkege" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Mercury 2 debuts as the first production diffusion-based LLM</strong>: Inception Labs claims their diffusion architecture generates 1,000 tokens per second, outperforming autoregressive models on GPU parallelization. <a href="https://app.daily.dev/posts/X9MtYJ6uL" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux 7.1 to enforce automatic power-off on fatal ACPI errors</strong>: The kernel is shifting to a more aggressive ‘fail-safe’ stance, though users can opt-out via a new boot parameter if they prefer a kernel panic. <a href="https://app.daily.dev/posts/TLZKRpI81" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>SHAFT framework optimizes secure transformer inference</strong>: Presented at NDSS 2025, this framework uses constant-round softmax protocols to reduce communication overhead by up to 41% in encrypted environments. <a href="https://app.daily.dev/posts/XN8gJNuhc" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>TanStack Start benchmarks show 83% faster dev startup than Next.js</strong>: A production SaaS migration report highlights Vite-based routing and reduced vendor lock-in as primary drivers for moving away from the Vercel ecosystem. <a href="https://app.daily.dev/feed-by-ids?id=ed7TZMhpP&amp;id=x40hVM7sP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>MongoDB adopts TLA+ formal verification for WiredTiger engine</strong>: Engineers are using model-based verification to generate tens of thousands of tests, ensuring conformance for distributed transaction protocols. <a href="https://app.daily.dev/posts/xG6jFmzdR" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>The AI complexity trap is officially here</strong>: IDC reports that while organizations are running 40+ AI proof-of-concepts, only 11% actually scale, proving that ‘AI-first’ is mostly ‘Technical Debt-first.’ We’re seeing a classic Jevons Paradox where cheaper code production is just leading to more bloated, unmaintainable systems that IT departments can’t actually manage. It turns out that replacing expensive engineers with cheap tokens just moves the cost to the ‘why is everything broken’ column of the budget. <a href="https://app.daily.dev/feed-by-ids?id=gK91ABz9x&amp;id=8MzkC9L2x" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Ars Technica reporter fired for AI-generated hallucinations</strong>: A senior AI reporter was sacked after his Claude-based workflow fabricated quotes from a human source, proving that even the people covering the hype aren’t immune to its failures. Using an LLM to ‘extract source material’ while sick is the journalistic equivalent of letting a toddler do your taxes—it looks right until you actually read the numbers. It’s a grim reminder that ‘AI-assisted’ is often just a euphemism for ‘I didn’t actually check this.’ <a href="https://app.daily.dev/posts/O3BSPfqbD" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>The great DDR5 scalping bot invasion</strong>: Scalpers are now using ‘AI-assisted evasion’ to scrape RAM inventory every 6 seconds, because apparently, we need more bots to buy hardware for other bots to run on. DataDome is tracking 10 million requests from bots that use cache-busting to bypass rate limits, effectively making the hardware market as non-deterministic as the models themselves. We’ve reached peak tech irony: you can’t buy a GPU to run your AI because an AI already bought it. <a href="https://app.daily.dev/feed-by-ids?id=FguvJtzTY&amp;id=srIKQBdIA" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
