<h2>Must know</h2>
<ul>
<li><strong>AI agents are coding, but are they ready for prime time?</strong>: AI coding agents are rapidly advancing, with tools like Claude Code and Agent Flywheel enabling autonomous code generation and environment setup. Claude Code, powered by Opus 4.5, is now capable of handling most development work, including strategic planning and task breakdown, as highlighted by a YouTube demonstration. However, a new benchmark from Mercor reveals that even leading models like Gemini 3 Flash and GPT-5.2 achieve only 24% accuracy on complex white-collar tasks, suggesting they can’t reliably replace knowledge workers yet. This raises concerns about security risks when non-developers use AI to create applications without understanding potential vulnerabilities, a point emphasized by ThePrimeTime. <a href="https://app.daily.dev/feed-by-ids?id=oecQfT5G1&amp;id=xBP8EMs83&amp;id=OsmrZ9UpM&amp;id=RlSMmScC7" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New AI hardware promises massive efficiency gains</strong>: Neurophos, a Duke University spinout, secured $110M in Series A funding to develop optical processing units (OPUs) that claim to achieve 235 POPS at 675 watts, significantly outperforming Nvidia’s B200 GPU (9 POPS at 1,000 watts). Microsoft’s venture arm, M12, participated in the round, signaling strong interest. These metamaterial-based optical chips, expected by mid-2028, perform AI inferencing in the optical domain, promising substantial efficiency and speed advantages over electronic GPUs. <a href="https://app.daily.dev/posts/7OO0Cb60y" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google and Microsoft push AI deeper into personal and enterprise workflows</strong>: Google launched AI Mode with Personal Intelligence for its AI Pro and AI Ultra subscribers, allowing AI to access users’ Gmail and Google Photos for tailored responses and suggestions. This opt-in feature, while raising privacy questions, aims to enhance personalization. Meanwhile, Microsoft released VibeVoice-ASR, a unified speech-to-text model capable of processing long-form audio with automatic speech recognition, speaker diarization, and timestamping, targeting meeting and conversational scenarios. Both moves indicate a strong push to integrate AI more deeply into daily digital interactions and enterprise tools. <a href="https://app.daily.dev/feed-by-ids?id=Kw6SK0Mqr&amp;id=zq6H3l6U4" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>GitHub Copilot expands with new SDK</strong>: GitHub released a multi-platform SDK for integrating Copilot’s agentic workflows into applications, supporting Python, TypeScript, Go, and .NET, enabling orchestration, planning, and programmatic file edits. <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>Anthropic introduces agent skills for specialized AI work</strong>: Anthropic announced Agent Skills, a new paradigm to package domain expertise as organized collections of files, documentation, and scripts, designed to extend AI agent capabilities and operate with MCP servers. <a href="https://app.daily.dev/posts/lvPOiATWV" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Feast integrates with PyTorch for better feature management</strong>: Feast, an open-source feature store, joined the PyTorch Ecosystem to provide a unified API for consistent feature management across AI model training and deployment, addressing training-serving skew. <a href="https://app.daily.dev/posts/dPG1hvjH4" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Inferact raises $150M to commercialize vLLM</strong>: Inferact secured $150M in seed funding at an $800M valuation to commercialize vLLM, an open-source inference optimization tool, following a trend of successful open-source AI projects transitioning to funded startups. <a href="https://app.daily.dev/posts/qqbAFhp2s" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-driven security agents assess vulnerabilities contextually</strong>: ZEST Security introduced AI Sweeper Agents that identify truly exploitable cybersecurity vulnerabilities by analyzing exploitation requirements against specific IT environments, aiming to reduce false positives and prioritize real risks. <a href="https://app.daily.dev/posts/LDTl5G8Jm" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vercel sandboxes get filesystem snapshot support</strong>: Vercel Sandboxes now support 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 with automatic retention filters</strong>: Datadog launched cross-product retention filters that automatically retain Application Performance Monitoring (APM) traces linked to Real User Monitoring (RUM) sessions, unifying frontend and backend observability for improved troubleshooting. <a href="https://app.daily.dev/posts/fs5040Sq6" 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, U8PYVjc73, suggests AI should only propose changes to version-controlled artifacts, not directly modify live production systems. Because, you know, we definitely need more layers of abstraction and human oversight to slow down those lightning-fast AI agents. Who needs efficiency when you can have ‘deterministic automation’? <a href="https://app.daily.dev/posts/U8PYVjc73" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Prompt engineering: the new snake oil?</strong>: Richard Socher, the alleged inventor of prompt engineering, founded <a href="http://you.com" target="_blank" rel="noopener nofollow">you.com</a> to combat LLM hallucinations and built it into a $1.5 billion company. Now, <a href="http://you.com" target="_blank" rel="noopener nofollow">you.com</a> is shifting to enterprise API infrastructure, used by OpenAI and Amazon, to ensure ‘accurate, citation-supported answers.’ So, the solution to LLM’s inherent flaws is… more LLMs, but with a fancy new API and a $1.5 billion valuation. Sounds like a solid plan, definitely not just repackaging the same old problems. <a href="https://app.daily.dev/posts/CvKOkn4RS" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-driven development: the next big buzzword bingo?</strong>: A new ‘AI-DLC’ methodology claims AI can be the main orchestrator of software development, initiating workflows and generating code, with humans merely validating decisions. Iteration cycles are supposedly shortened to hours or days, not weeks. This sounds suspiciously like every other ‘revolutionary’ methodology that promises to fix all our problems by just adding AI to the mix, without mentioning any actual companies, products, or real-world case studies. Color me skeptical. <a href="https://app.daily.dev/posts/SNNDpvqm6" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
