<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, revealing a maximum accuracy of only 24%. This suggests current AI agents can’t reliably replace knowledge workers in multi-tool, multi-domain tasks, despite year-over-year improvements. Meanwhile, Anthropic’s performance engineering team is struggling to design technical evaluations that advanced Claude models (Opus 4 and 4.5) can’t ace, forcing them to resort to Zachtronics-inspired puzzles to find tasks where human reasoning still has an edge. <a href="https://app.daily.dev/feed-by-ids?id=OsmrZ9UpM&amp;id=vA7gH0M1s" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>OpenAI Codex now integrated into JetBrains IDEs</strong>: OpenAI Codex, powered by GPT-5.2-Codex, is now natively integrated into JetBrains IDEs (IntelliJ, PyCharm, WebStorm, Rider) starting with version 2025.3. This allows developers to use the AI agent for refactoring, code migrations, and security vulnerability detection with adjustable autonomy levels and ‘reasoning budgets.’ Authentication is flexible, supporting JetBrains AI subscriptions, ChatGPT accounts, or OpenAI API keys, as reported by JetBrains. <a href="https://app.daily.dev/posts/uAvdGTErT" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitHub releases multi-platform SDK for Copilot agent integration</strong>: GitHub has launched a multi-platform SDK for integrating Copilot’s agentic workflows into applications, supporting Python, TypeScript, Go, and .NET. This SDK enables embedding the same engine used by the Copilot CLI, offering functionalities like orchestration, planning, tool invocation, and programmatic file edits. It’s currently in technical preview and requires a GitHub Copilot subscription. <a href="https://app.daily.dev/feed-by-ids?id=GXXnGjqMp&amp;id=GuG7K2ryZ" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Inferact raises $150M to commercialize vLLM</strong>: Inferact, a new startup, secured $150 million in seed funding at an $800 million valuation to commercialize vLLM, an open-source inference optimization tool, with Amazon already using vLLM in its cloud service and shopping app. <a href="https://app.daily.dev/posts/qqbAFhp2s" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Feast joins PyTorch ecosystem to enhance feature management</strong>: Feast, an open-source feature store, has joined the PyTorch Ecosystem to provide a unified API for consistent feature management across model training and deployment, aiming to eliminate training-serving skew in production AI systems. <a href="https://app.daily.dev/posts/dPG1hvjH4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>BoxLang AI v2 brings autonomous agents to JVM</strong>: BoxLang AI v2, a major update to the JVM AI framework, introduces autonomous AI agents with memory and tool usage, supporting 12+ AI providers via a unified API and enabling serverless deployment on AWS Lambda or local operation via Ollama. <a href="https://app.daily.dev/posts/rm0dNUEhj" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Datadog launches automatic retention filters for RUM and APM</strong>: Datadog released cross-product retention filters that automatically retain Application Performance Monitoring (APM) traces linked to Real User Monitoring (RUM) sessions, unifying frontend and backend observability data for improved troubleshooting. <a href="https://app.daily.dev/posts/fs5040Sq6" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vercel Sandboxes introduces filesystem snapshot feature</strong>: Vercel Sandboxes now supports filesystem snapshots, allowing users to capture and launch new sandbox environments from saved states, streamlining workflows for parallel testing and experimentation by eliminating repetitive setup tasks. <a href="https://app.daily.dev/posts/Uv3Zx9otG" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Pulumi adds replacementTrigger for resource recreation</strong>: Pulumi v3.215.0 introduces the <code>replacementTrigger</code> option, enabling automatic recreation of infrastructure-as-code resources when specified metadata changes, useful for cycling secrets or updating resources based on external version numbers. <a href="https://app.daily.dev/posts/ay0AhT4GQ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Rust 1.93.0 released with musl update and API stabilizations</strong>: Rust 1.93.0 is out, updating the bundled musl libc to 1.2.5 for improved DNS resolution in static Linux binaries, stabilizing 23 APIs, and enhancing inline assembly handling, with a breaking change for legacy musl compatibility symbols that had a fix available since June 2023. <a href="https://app.daily.dev/feed-by-ids?id=wF6DW7Hp4&amp;id=O9g10PE7Q" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Google’s AI Mode wants to read your Gmail and Photos</strong>: Google’s new AI Mode for AI Pro and AI Ultra subscribers lets the AI access your Gmail and Google Photos to suggest vacation itineraries or shopping items. Google assures us it only uses data for specific prompts, not your entire inbox, but hey, what could possibly go wrong with an AI sifting through your personal life? It’s opt-in, so you get to decide if you trust Google’s AI with your digital soul. <a href="https://app.daily.dev/posts/Kw6SK0Mqr" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI agents in IT operations: the new ‘cowboy’ sysadmins?</strong>: Apparently, AI agents are introducing ‘non-deterministic actions’ into IT operations, which is just a fancy way of saying they’re making things unpredictable. The article warns against letting AI directly modify live production systems, comparing it to ‘cowboy’ sysadmins making unreproducible changes. So, AI is great for generating playbooks, but don’t let it touch the actual buttons, because apparently, it’s still too much of a loose cannon. <a href="https://app.daily.dev/posts/U8PYVjc73" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Yann LeCun says LLMs are dumb, launches his own ‘world model’ lab</strong>: Yann LeCun, a Turing Award winner, has launched AMI Labs in Paris, claiming large language models are limited because they lack understanding of the physical world. His solution? ‘World models’ trained on video, audio, and sensor data to enable AI that can ‘reason, plan, and interact with the physical world.’ So, while everyone else is busy with LLMs, LeCun is off building Skynet’s sensory system, because apparently, understanding the physical world is the missing piece for true AI. <a href="https://app.daily.dev/feed-by-ids?id=bsXOFS9xq&amp;id=wOdmxGV2q" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
