<h2>Must know</h2>
<ul>
<li><strong>AI coding agents are a security nightmare for IAM</strong>: AI agents are fundamentally breaking traditional Identity and Access Management (IAM) systems. Because they get permissions from multiple users and operate autonomously, they create massive authorization risks and pathways for bypasses. Organizations are now facing urgent security risks, especially with ‘organizational agents’ shared across teams, lacking clear ownership and audit trails, as highlighted by Security Boulevard. This means we need entirely new security protocols, not just tweaks, to manage these high-risk entities. <a href="https://app.daily.dev/posts/3wv1aKzIr" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google Cloud opens public datasets via Apache Iceberg REST Catalog</strong>: Google Cloud is making public datasets, like the NYC Taxi dataset, accessible through the Apache Iceberg REST Catalog on BigLake. This means data engineers and analysts can query these read-only datasets without any infrastructure setup, using any compute engine that supports the Iceberg protocol, including Spark, Trino, Flink, and BigQuery. This move, announced by Google Cloud, aims to simplify access to large-scale analytical data and emphasizes interoperability. <a href="https://app.daily.dev/posts/tn371FITl" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>NVIDIA B200 GPU driver bug causes prolonged system hang</strong>: NVIDIA’s open GPU kernel driver version 570.133.20 is causing indefinite <code>nvidia-smi</code> hangs after about 66 days of uptime on B200 GPUs running Linux kernel 6.6.0. The issue, traced to NVLink communication failures, impacts cross-GPU operations and requires system reboots to recover functionality. NVIDIA has acknowledged the bug and is working on a fix, according to public issue tracking. <a href="https://app.daily.dev/posts/xhTAzdzCR" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Turbopuffer ANN v3 achieves 200ms p99 latency over 100 billion vectors</strong>: Turbopuffer announced its ANN v3 system, delivering 200ms p99 query latency for 100 billion vectors at 1,000 queries per second, utilizing hierarchical clustering and 16-32x compression with a 92% recall rate. <a href="https://app.daily.dev/posts/EHAeemnne" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>LLM inference bottleneck is memory, not compute</strong>: The main bottleneck for large language model inference is memory and interconnect, particularly during the autoregressive Decode phase of Transformer models, leading to research opportunities in high-bandwidth flash, processing-near-memory, 3D memory-logic stacking, and low-latency interconnects. <a href="https://app.daily.dev/posts/t0LNVAkom" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux kernel formalizes succession process for mainline repository</strong>: The Linux 6.19 kernel now includes a formal document outlining a succession plan if Linus Torvalds’ mainline Git repository becomes inaccessible, ensuring continuity of development within 72 hours. <a href="https://app.daily.dev/posts/EogQGzxtg" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>SVT-AV1 4.0 delivers significant encoding performance gains</strong>: SVT-AV1 4.0 has been released, bringing 5-8x speedup for still image encoding, 5-15% for RTC modes, and 10-25% for Random Access mode, alongside new visual quality enhancements like AC Bias and Arm Neon/SVE2 optimizations. <a href="https://app.daily.dev/posts/HOUGYgKhJ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>PostgreSQL adopted as dead letter queue in Wayfair’s event-driven system</strong>: Wayfair is using PostgreSQL as a Dead Letter Queue (DLQ) for its event-driven systems, redirecting failed Kafka events to a PostgreSQL table for improved visibility, queryability, and concurrency-safe retries using <code>FOR UPDATE SKIP LOCKED</code>. <a href="https://app.daily.dev/posts/eKm9n1imv" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>PHPStan 2.1.34 delivers major performance upgrades</strong>: PHPStan 2.1.34, a static analysis tool for PHP, achieves 25-40% faster analysis times on real-world projects and up to 50% reduction in runtime, significantly shortening CI pipelines, though memory usage is currently higher. <a href="https://app.daily.dev/posts/hEVb27hAV" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>MIT team demonstrates low-error superconducting nanowire memory array</strong>: MIT researchers have developed a 4x4 superconducting nanowire memory array with a bit error rate of 10⁻⁵, operating at 1.3 Kelvin and achieving a functional density of 2.6 megabits per square centimeter, addressing scalability and error rate challenges in superconducting memory. <a href="https://app.daily.dev/posts/UFLhnkgMo" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>AI is making software engineering easier, so now we just need SRE, right?</strong>: Two identical articles are confidently proclaiming that AI is making code-writing ‘much easier,’ thus shifting the core value of software engineering to operational excellence and reliability. Apparently, building a working demo is ‘straightforward’ now, and SRE is the new holy grail. This sounds suspiciously like a convenient narrative to explain why we still need engineers after AI writes all the code, without actually providing any concrete examples or data on how AI specifically impacts these operational challenges. So, AI writes the code, and we just… babysit it? Got it. <a href="https://app.daily.dev/posts/otuC5YedQ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI agents will automate your database migrations in minutes, apparently</strong>: Neo4j is touting specialized AI agent skills that can automate database migration tasks, claiming to reduce weeks of work to mere minutes. They even provided an example of migrating Python driver code and Cypher queries, which sounds fantastic until you remember the last time an ‘automated’ migration actually went smoothly. Sure, it can analyze codebases and generate plans, but I’ll believe it when I see it handle real-world legacy spaghetti without turning it into a bigger mess.</li>
<li><strong>AI coding agents are limited by your creativity, not their tech</strong>: An opinion piece suggests that the real bottleneck for AI coding agents isn’t their technical capability, but rather programmers’ lack of imagination. Apparently, we’re all just too dull to come up with innovative uses for these tools, reducing their impact. It’s a bold claim to shift the blame from the tech to the user, especially when these ‘AI coding agents’ are still struggling with basic context and consistency. Maybe the tools aren’t quite as brilliant as their marketing departments claim, and it’s not just our feeble human brains holding them back. <a href="https://app.daily.dev/feed-by-ids?id=3vC5OgaHL&amp;id=gov03HPhm" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
