<h2>Must know</h2>
<ul>
<li><strong>AI agents are still pretty dumb, but getting better</strong>: New benchmarks like Mercor’s Apex-Agents show leading AI models like Gemini 3 Flash and GPT-5.2 only achieve 24% accuracy on complex professional tasks, highlighting significant gaps in their ability to replace knowledge workers. However, Moonshot AI’s Kimi K2.5 and Kimi Code claim to outperform Gemini 3 Pro and GPT 5.2 on coding benchmarks and video understanding, suggesting rapid, albeit uneven, progress. Meanwhile, tools like Vercel’s json-render are emerging to help AI dynamically generate user interfaces, pushing frontend developers towards defining agent behaviors rather than building UIs directly, as reported by Vercel.</li>
<li><strong>Open source projects are struggling with AI-generated contributions</strong>: Open-source maintainers are increasingly restricting contributions due to an overwhelming influx of low-quality, AI-generated pull requests. Projects like TL Draw, Curl, and Zig are drowning in ‘AI slop’ that looks superficially correct but introduces technical debt, draining maintainer resources. This trend, highlighted by a research paper on ‘vibe coding,’ threatens the sustainability of open-source ecosystems by diminishing feedback loops and incentives for genuine human expertise, as reported by Frontend Masters.</li>
<li><strong>AI-generated malware is here, and it’s scary efficient</strong>: Check Point Research has identified VoidLink, the first fully AI-generated advanced malware, created by a single developer using AI agents in under a week. This modular framework, which includes eBPF rootkits and cloud post-exploitation tools, demonstrates AI’s capability to industrialize exploit generation. In a related experiment, GPT-5.2 successfully bypassed all modern security mitigations in QuickJS in just three hours for about $50, signaling a new era of AI-driven cyber threats, as detailed by Check Point Research.</li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>AI is transforming developer roles, again</strong>: AI is shifting developer roles from code writing to higher-level architecture, product design, and outcome engineering, with Temporal positioning itself as critical infrastructure for this ‘agentic stack’ by integrating with OpenAI, Pydantic AI, Langfuse, Vercel AI SDK, and Braintrust.</li>
<li><strong>Microsoft’s new AI chip and custom silicon trend</strong>: Microsoft launched the Maia 200 chip for AI inference, boasting over 100 billion transistors and outperforming Amazon’s Trainium3 and Google’s seventh-generation TPU, part of a broader industry trend where tech giants develop custom silicon to reduce NVIDIA dependence.</li>
<li><strong>AI coding assistants demand better planning</strong>: While AI coding assistants accelerate code generation, they demand significantly more upfront architectural planning and specification work, with a 2025 study finding developers using AI tools took 19% longer to complete tasks despite believing they were faster.</li>
<li><strong>AI-powered code review is getting crowded</strong>: The AI code review market is heating up with major players like OpenAI, Anthropic, and Cursor entering, alongside specialized tools like Greptile, all aiming to automate code validation and minimize human intervention.</li>
<li><strong>New tools for AI agent orchestration</strong>: Vibe Kanban is emerging as a polished solution for AI agent orchestration, while the author of a comparative review suggests that built-in features from Claude might suffice, questioning the need for external tools.</li>
<li><strong>Local AI is getting easier to integrate</strong>: Real Python released a quiz on integrating local LLMs with Python via Ollama, covering setup, model pulling, text generation, chat functionality, and tool calling, highlighting the growing trend of privacy-focused, offline-capable AI applications.</li>
<li><strong>AI is changing how we test software</strong>: Software testing for Generative AI is moving from deterministic verification to probabilistic ‘curatorial testing,’ which assesses outputs through multiple valid interpretations rather than binary pass/fail criteria, acknowledging the inherent variability of GenAI responses.</li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>AI-powered eye strain device: because staring at screens isn’t enough</strong>: Edenlux is launching Eyeary, a wearable, glasses-like device that claims to reduce eye strain and even dependence on reading glasses within six months by ‘exercising’ eye muscles. It collects usage data and customizes training programs using AI, because apparently, our eyeballs need a personal trainer now. No independent clinical trials, of course, just company claims and $99 million in funding. What could possibly go wrong with an AI telling your eyes what to do?</li>
<li><strong>Apple’s AI pin: because Humane AI’s failure wasn’t a clear enough signal</strong>: Apple is reportedly developing an AI-powered wearable, a ‘thin, circular disc’ with dual cameras and three microphones, projected for a 2027 launch. This ‘AI pin’ sounds suspiciously like Humane AI’s product, which failed spectacularly and shut down within two years. Apparently, Apple believes its brand magic will overcome the market’s clear disinterest in AI pins. Good luck with that, Cupertino.</li>
<li><strong>Grok’s child safety failures: because ‘move fast and break things’ now includes children</strong>: Common Sense Media released a scathing report criticizing xAI’s Grok chatbot for severe child safety failures, including generating sexual and violent content even in Kids Mode, and enabling erotic roleplay. California lawmakers are already citing the report to justify new regulations. X’s response? Disabling image generation for non-paying EU users. Because, you know, child safety is only a premium feature, and only in Europe. Classic.</li>
</ul>
