<h2>Must know</h2>
<ul>
<li><strong>AI agents are still pretty dumb for real work</strong>: A new benchmark, Apex-Agents, from Mercor, tested leading AI models like Gemini 3 Flash and GPT-5.2 on complex white-collar tasks. They only managed a 24% accuracy rate, proving they can’t reliably replace knowledge workers in tasks requiring multi-domain reasoning and tool use. This suggests that while AI is making rapid progress, human-level performance in complex tasks is still a distant dream, as reported by Mercor. <a href="https://app.daily.dev/posts/OsmrZ9UpM" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google’s AI mode wants to read your Gmail and photos</strong>: Google launched an ‘AI Mode’ for AI Pro and AI Ultra subscribers that, if you opt-in, lets the AI access your Gmail and Google Photos. It promises to suggest vacation itineraries or shopping items based on your personal data. Google claims it only uses data for specific prompts and responses, not for general training, but the privacy implications of an AI sifting through your personal life are, shall we say, ‘interesting.’ <a href="https://app.daily.dev/posts/Kw6SK0Mqr" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anthropic’s AI models are getting too smart for their own good</strong>: Anthropic’s performance engineering team is having a tough time designing technical evaluations that their Claude AI models (Opus 4 and 4.5) can’t ace. The AI models are now matching or exceeding human candidates on complex coding optimization tasks, forcing Anthropic to create unconventional puzzles to find tasks where human reasoning still has an edge. It’s almost like they’re trying to prove humans are still useful, which is sweet. <a href="https://app.daily.dev/posts/vA7gH0M1s" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Postgres gets AI-powered best practices</strong>: Supabase released Agent Skills for Postgres Best Practices, a collection of 30 rules to help AI coding agents write correct and performant Postgres code, covering query performance, connection management, and schema design. <a href="https://app.daily.dev/posts/oIEBoE2RS" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ClickHouse launches managed Postgres service</strong>: ClickHouse introduced a new managed Postgres service, promising up to 10X faster performance for disk-bound workloads and native integration with ClickHouse for a unified transactional and analytical data stack. <a href="https://app.daily.dev/feed-by-ids?id=WHTgaLabX&amp;id=WHTgaLabX" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Rails 8 still the MVP king</strong>: Ruby on Rails 8 continues to be a top choice for Minimum Viable Product (MVP) development, offering built-in authentication, background jobs, caching, and WebSocket support without external dependencies, making it ideal for rapid iteration. <a href="https://app.daily.dev/feed-by-ids?id=w2WfkMTD2&amp;id=w2WfkMTD2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vercel sandboxes get filesystem snapshots</strong>: Vercel Sandboxes now supports filesystem snapshots, allowing users to capture and reuse the complete state of a running sandbox environment, streamlining CI/CD workflows and rapid prototyping. <a href="https://app.daily.dev/posts/Uv3Zx9otG" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Datadog unifies RUM and APM data</strong>: Datadog launched automatic retention filters that link Application Performance Monitoring (APM) traces to Real User Monitoring (RUM) sessions, aiming to unify frontend and backend observability for better troubleshooting. <a href="https://app.daily.dev/posts/fs5040Sq6" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub Copilot SDK for multi-platform integration</strong>: GitHub released a multi-platform SDK for Python, TypeScript, Go, and .NET, enabling integration of Copilot’s agentic workflows, including orchestration, planning, and programmatic file edits, into various applications. <a href="https://app.daily.dev/feed-by-ids?id=GXXnGjqMp&amp;id=GuG7K2ryZ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Feast joins PyTorch ecosystem</strong>: Feast, an open-source feature store, has integrated with the PyTorch Ecosystem to address training-serving skew in production AI systems by providing a unified API for consistent feature management. <a href="https://app.daily.dev/posts/dPG1hvjH4" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>AI agents in IT operations: a recipe for chaos</strong>: Apparently, AI agents are the new ‘cowboy sysadmins’ in IT operations, introducing non-deterministic actions and increasing risk. The article warns against letting AI directly modify live production systems, suggesting they stick to generating playbooks and configurations for human review. Because, you know, humans are so much better at not messing things up. <a href="https://app.daily.dev/posts/U8PYVjc73" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI music personas: the new copyright loophole</strong>: AI-generated music personas are mimicking artists’ styles without direct copying, creating a legal gray area that current copyright laws don’t cover. Major labels are suing AI platforms, and analysts predict a 24% loss in music creator revenues by 2028. Because who needs original art when you can have an AI churn out endless, legally ambiguous knock-offs? <a href="https://app.daily.dev/posts/9zpaBvJiJ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-driven QA analytics: because humans can’t count to 11 million</strong>: Testlio launched LeoInsights, an AI-based QA analysis platform trained on 13 years of data, promising to turn ‘quality assurance metrics into actionable business insights.’ It even has a value calculator to measure efficiency improvements. Apparently, early adopters dismissed 11 million vulnerabilities in six months. Who knew AI could be so good at ignoring problems? <a href="https://app.daily.dev/feed-by-ids?id=xZI6Dgtwy&amp;id=xZI6Dgtwy" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
