<h2>Must know</h2>
<ul>
<li><strong>OpenAI releases GPT-5.4 with native computer use</strong>: OpenAI launched GPT-5.4 on March 5, featuring a 1M-token context window and a new ‘computer-use’ capability that reportedly outperforms humans on the OSWorld benchmark. The model introduces a Tool Search system to cut token overhead by 47% and is rolling out across ChatGPT Plus, API, and GitHub Copilot. High-end ‘Thinking’ and ‘Pro’ variants offer deeper reasoning but come with a steep $30/1M input token price tag, according to OpenAI and Microsoft. <a href="https://app.daily.dev/posts/RpPUWi7AZ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Salesforce cuts EC2 costs by 13 percent with custom K8s scheduling</strong>: The Data 360 team at Salesforce replaced the default Kubernetes ‘LeastAllocated’ strategy with a custom ‘MostAllocated’ bin-packing plugin to handle 2 million daily Spark jobs. This shift proactively stacks executors onto utilized nodes, allowing Karpenter to terminate empty instances more effectively and reducing node disruption by 50%. The optimization raised overall CPU and memory utilization by 15% while slashing infrastructure spend, as detailed by Salesforce Engineering. <a href="https://app.daily.dev/posts/dLNiW6hAy" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GKE launches native custom metrics for HPA</strong>: Google Cloud has removed the need for external adapters or complex IAM bindings by integrating custom metrics directly into GKE. Using a new AutoscalingMetric CRD, developers can now scale workloads based on application-level signals like KV cache usage or request counts without managing a separate observability stack. This agentless architecture aims to reduce scaling latency and ingestion costs while improving reliability, per Google Cloud. <a href="https://app.daily.dev/posts/Ga2ibrgtl" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Rust 1.94.0 stabilizes array windows and TOML 1.1</strong>: The release stabilizes the array_windows method for slice iteration and upgrades Cargo to support modular configuration via the include key. <a href="https://app.daily.dev/posts/ClZ7LT6Ey" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Databricks unveils KARL for enterprise agentic reasoning</strong>: KARL is a custom RL-trained agent designed to match frontier model performance on grounded reasoning with significantly lower latency and cost. <a href="https://app.daily.dev/posts/3letdm1bp" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google report shows record high zero-day attacks on enterprise tech</strong>: In 2025, 48% of exploited zero-days targeted enterprise networking and security gear, with commercial surveillance vendors surpassing state actors in attribution. <a href="https://app.daily.dev/posts/hYcntrpfQ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Datadog updates Bits AI SRE with MCP support</strong>: The autonomous agent now uses the Model Context Protocol to double investigation speeds and adds human-in-the-loop triage actions via Slack and Jira. <a href="https://app.daily.dev/posts/K7E28Dkg8" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Apache Polaris and Gluten graduate to top level projects</strong>: Polaris provides an open-source catalog for Iceberg interoperability, while Gluten accelerates Spark SQL by offloading compute to native engines. <a href="https://app.daily.dev/posts/0x6bMcZKJ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>PyTorch FlexAttention gets massive Blackwell speedup</strong>: A new FlashAttention-4 backend delivers up to 3.2x speedups on Hopper and Blackwell GPUs by leveraging low-level CuTeDSL kernel generation. <a href="https://app.daily.dev/posts/2CanCrYFg" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AMD Ryzen AI 400 series brings 50 TOPS to desktop</strong>: The new AM5 chips integrate Zen 5 cores and an NPU capable of meeting Microsoft’s Copilot+ PC requirements for business environments. <a href="https://app.daily.dev/posts/Iel0wUSWS" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>The great AI ratepayer protection pinky swear</strong>: Seven tech giants signed a White House pledge promising to pay for their own datacenter energy costs so grandma’s electric bill doesn’t go up. Critics at Gartner and elsewhere point out this ‘pledge’ has zero enforcement mechanisms and ignores the fact that utilities, not tech companies, set the rates. It’s a classic PR move to quiet community opposition while AI power demand is projected to double by 2030. <a href="https://app.daily.dev/posts/SI8TMzfHX" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anthropic sues the DoD over safety theater</strong>: Dario Amodei is taking the Pentagon to court because they labeled Anthropic a ‘supply chain risk’ after he refused to give the military unrestricted access to Claude. While he claims he’s standing up against autonomous weapons, he also had to apologize for a leaked memo calling OpenAI’s military deal ‘safety theater.’ It’s a messy legal battle where ‘principled AI’ meets the reality of losing massive government contracts to competitors. <a href="https://app.daily.dev/posts/M7ZS3XGwd" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cloudflare claims a one week AI rewrite is production ready</strong>: Cloudflare is bragging that a single engineer used $1,100 in AI tokens to rewrite Next.js into ‘vinext’ in just seven days. Vercel’s CEO was quick to point out that this ‘production-ready’ miracle is currently running exactly one low-traffic government beta site and is riddled with security holes. It’s the ultimate ‘AI replaced my dev team’ bait that ignores the decade of edge cases and security hardening that actually makes software usable. <a href="https://app.daily.dev/posts/13r2qcnvY" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
