<h2>Must know</h2>
<ul>
<li><strong>AI agents are still a mess, but everyone’s trying to make them work</strong>: AI agents are supposed to be the next big thing, but they’re still pretty unreliable. A new benchmark, AssetOpsBench, found that even top models like GPT-4.1 only hit 72.4% task completion in industrial settings, far from production-ready. The problem isn’t just the models; it’s the unpredictable behavior, prompt sensitivity, and lack of robust error recovery, as noted by LangChain. Everyone’s trying to fix it, from Docker’s Cagent for deterministic testing to LangGrant’s LEDGE MCP Server for secure database analytics, but it’s clear we’re still in the early days of making these things actually useful without constant human babysitting. <a href="https://app.daily.dev/feed-by-ids?id=54axue44G&amp;id=6aVFrCy8v&amp;id=2NT8yV81n&amp;id=5joRPhN74&amp;id=1w09x1vo7" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI is changing developer jobs, but core skills still matter</strong>: AI coding tools are definitely shaking things up, but don’t ditch your textbooks just yet. While AI can churn out boilerplate code faster than you can say ‘technical debt,’ it’s also making developers slower in some cases, increasing code review times by 91% and even doubling bugs, according to The Serious CTO. The real ‘dev flex’ in 2026 isn’t how fast you can prompt an LLM, but how well you understand data structures, memory management, and system architecture, as highlighted by an unknown source. Turns out, humans are still pretty good at debugging and making architectural decisions, which AI, for all its hype, still struggles with. <a href="https://app.daily.dev/feed-by-ids?id=J24eplQdr&amp;id=dABjwqV7N&amp;id=wzh6JIlVF&amp;id=SiZeB7ycG&amp;id=1iSdibthb" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>OpenAI and Google are still duking it out, but with ads and new hardware</strong>: The AI arms race continues, with OpenAI launching ChatGPT Go globally for $8/month (with ads, because apparently, even AI needs to pay the bills) and Google pushing its Gemini 3 Flash for faster, cheaper AI workflows. OpenAI also inked a $10 billion deal with Cerebras for specialized AI inference chips, signaling a move away from Nvidia’s GPU dominance for certain workloads, as reported by an unknown source. Meanwhile, Google’s Gemini AI is getting integrated into everything from Google Trends to BigQuery, making it easier to do AI-powered data analysis directly in SQL. It’s all about who can make AI faster, cheaper, and more integrated, even if it means shoving ads in your face. <a href="https://app.daily.dev/feed-by-ids?id=KwZZ88381&amp;id=9kdSue3WD&amp;id=xYYfl4Rqa&amp;id=FoOQYVciF&amp;id=5NpQUB2lE&amp;id=QaJ2IQUdX" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Go’s database/sql client and PgBouncer can still cause outages</strong>: A nasty race condition in Go’s database/sql client, combined with misconfigured PgBouncer (without peering), can lead to connection pool poisoning and full database outages, especially when transactions aren’t properly closed. <a href="https://app.daily.dev/posts/DUFN7D8Ly" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>RubyLLM::Agents simplifies AI agent management in Rails</strong>: RubyLLM::Agents is a new Rails engine that lets you build and manage LLM-powered AI agents with a declarative DSL, offering features like cost tracking, workflow orchestration, and a real-time dashboard, supporting OpenAI, Anthropic, and Google models. <a href="https://app.daily.dev/posts/uoOGO4JET" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Curl is getting MQTTS support in May 2026</strong>: Curl version 8.19.0, due in May 2026, will add MQTTS (MQTT over TLS) support, bringing its total supported protocols to 29 and signaling a continued focus on IoT connectivity and security. <a href="https://app.daily.dev/posts/G6McjHbVV" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>MemOS offers open-source long-term memory for AI agents</strong>: MemOS is a new open-source framework for LLMs and AI agents, providing unified long-term memory management with graph-structured storage, multi-modal support, and claims of 43.70% accuracy improvement over OpenAI Memory. <a href="https://app.daily.dev/posts/GlwdfoP1R" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Python’s cryptography library is fed up with OpenSSL</strong>: The maintainers of Python’s cryptography library are increasingly frustrated with OpenSSL’s performance regressions and API complexity, exploring alternatives like BoringSSL and LibreSSL, and even considering dropping OpenSSL support entirely. <a href="https://app.daily.dev/posts/GpSRLZNlG" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>RubyPyMill bridges Python notebooks and Ruby workflows</strong>: RubyPyMill is a new CLI tool that allows Ruby developers to automate and orchestrate Python-based Jupyter notebooks, making it easier to transition proofs of concept to production without rewriting code. <a href="https://app.daily.dev/posts/1gzoFGxq9" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Ruby 4.0.1 fixes a Kernel#sleep bug</strong>: Ruby 4.0.1 was released, fixing a Kernel#sleep regression bug, with future updates planned every two months, and Ruby 3.2 reaching end-of-life in March 2026. <a href="https://app.daily.dev/posts/RE0IzAiOp" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>AI-generated code: the gift that keeps on giving… technical debt</strong>: Turns out, letting AI write your code without human oversight is a fantastic way to accumulate technical debt faster than you can say ‘refactor.’ From AI-generated SwiftUI code causing memory leaks and unexpected upgrade issues to LLMs hallucinating non-existent APIs, it seems these tools are great at producing convincing but ultimately flawed code. As one developer put it, AI-generated code is often ‘functionally correct from a surface perspective but structurally flawed,’ leading to maintainability, performance, and reliability nightmares down the line. So much for AI making our lives easier, eh? <a href="https://app.daily.dev/feed-by-ids?id=EZUXhYwRZ&amp;id=yIClekHIh&amp;id=EHuoSn58k&amp;id=BrFQI7FHb&amp;id=DovPZUpOP&amp;id=qu9MmthdT&amp;id=pc6GKGFyk" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-powered chatbots: saving seconds, costing dollars</strong>: Dakota State University researchers built a chatbot, Axlerod, to help insurance agents retrieve policy documents. The big win? It saves a whopping 2.42 seconds per search! But wait, there’s a catch: the $252 annual cost for a Gemini Business seat means a net loss of $131 per user. So, while it’s 93.18% accurate, it’s also a financial black hole. Turns out, existing automation already handles most routine queries, leaving these LLM chatbots to tackle the ‘complex’ stuff where, ironically, they still underperform. Progress! <a href="https://app.daily.dev/posts/8ysWd076O" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-generated books: the new frontier of technical misinformation</strong>: Just when you thought you could trust a book, HiTeX Press comes along with over 800 AI-generated technical books, complete with hallucinated code and non-existent features. One such masterpiece, ‘Ziggbook,’ even hit #1 on Hacker News before being exposed and taken down. Apparently, the authors explicitly claimed it wasn’t AI-generated, which, you know, was a lie. So, if you’re looking for reliable technical resources, maybe stick to human-written content, unless you enjoy debugging fictional APIs and learning about features that don’t exist. <a href="https://app.daily.dev/posts/j0xF0WiVr" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
