<h2>Must know</h2>
<ul>
<li><strong>AI agent platforms are exploding, but so are their security flaws</strong>: LangChain, now valued at $1.25B, continues its rapid expansion with LangGraph 1.0 and new agent builders, while Anthropic and OpenAI are integrating advanced coding agents directly into browsers and IDEs. However, security experts like Bruce Schneier warn that fundamental architectural flaws in LLMs make AI agents inherently vulnerable to prompt injection and poisoned data, creating a trilemma where systems can’t be fast, smart, and secure simultaneously. This means all these shiny new tools are built on a shaky foundation, requiring architectural changes, not just patches.</li>
<li><strong>Big tech pours billions into AI infrastructure, ignoring efficiency</strong>: Google Cloud’s G4 VMs are pushing multi-GPU performance, and NVIDIA is shifting to a platform provider model with CUDA and NIM microservices. Meanwhile, Meta is building massive AI clusters with disaggregated fabrics. Yet, Big Tech, including Meta, Microsoft, and Amazon, is investing billions in nuclear power and gas turbines for AI data centers, while academic and startup efforts show drastically cheaper ways to train models. It seems the focus is on throwing money at hardware rather than optimizing the models themselves, which is a classic move.</li>
<li><strong>Structured reasoning and control for reliable AI agents</strong>: Researchers are tackling LLM hallucinations with Attentive Reasoning Queries (ARQs), using JSON schemas to enforce structured reasoning and achieving higher task success rates than open-ended methods. Companies like Samesurf are building control plane architectures for agent navigation, while Salesforce’s Agent Graph uses hybrid reasoning with external graphs and finite state machines to maintain agent intent. These efforts highlight a critical shift towards making AI agents more predictable and reliable for enterprise use, especially in regulated industries.</li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>AI startups rake in cash for critical industries</strong>: 1001 AI secured $9M to build an AI-native operating system for aviation, logistics, and oil/gas in MENA, while Cercli raised $12M for an AI-powered HR and payroll system, both targeting the region’s underserved markets.</li>
<li><strong>AMD’s ROCm 7.9 preview targets next-gen AI hardware</strong>: AMD released ROCm Core SDK 7.9, a rapid preview with a new build infrastructure (TheRock) and early support for Ryzen AI Max 300 Series APUs and MI350/MI300 GPUs, aiming to accelerate Linux-centric AI workloads.</li>
<li><strong>AI engineering: math, MLOps, and ethics over frameworks</strong>: Success in AI engineering by 2026 will hinge on core math, MLOps, and production infrastructure skills, with ethical operationalization and system-level thinking being more crucial than chasing the latest frameworks.</li>
<li><strong>Enterprise AI agents: customization and workflow automation</strong>: Adobe’s AI Foundry offers custom generative AI models for enterprises, Block enabled 12,000 employees with AI agents for fraud detection, and Microsoft’s 365 Copilot now features Agent Mode for multi-step workflows.</li>
<li><strong>New tools for cross-platform Windows apps</strong>: InstallAware released an Application Porting Toolkit using Wine to bring Windows apps to Linux/macOS, while open-source WinBoat runs full Windows systems in containers on Linux, offering different approaches to compatibility.</li>
<li><strong>OpenTelemetry adds declarative config and AI gateway tracing</strong>: OpenTelemetry introduced declarative YAML configuration to filter health check traces and is integrating distributed tracing into AI Gateways, with Java having the most advanced support for these experimental features.</li>
<li><strong>New Java releases: Spring, JUnit, and enterprise updates</strong>: Spring Cloud Stream Applications 2025.0.1 and Spring Security updates are out, JUnit 6.0.0 unifies versioning and requires Java 17, and Jakarta EE 12 and WildFly 38 continue to advance enterprise Java.</li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Humanoid robots: still mostly hype, less human</strong>: Over $65 billion has been poured into humanoid robotics by the likes of Tesla and Amazon, with claims of AI integration pushing human-like movement. Yet, these prototypes still struggle with basic reliability, cost, and locomotion. Apparently, the ‘imminent’ replacement of human workers is still largely a marketing fantasy, despite researchers making digital characters do parkour.</li>
<li><strong>AI job replacement: the numbers are always wrong</strong>: A Norwegian government report, citing GPT-4, claims AI could replace 155,000 public sector jobs. Critics, however, are quick to point out the methodology is flawed, unreliable, and highly sensitive to phrasing. It’s almost as if relying on an LLM to predict job displacement is a recipe for speculative policymaking, but hey, it makes for a great headline!</li>
<li><strong>AWS US-EAST-1 outage: the gift that keeps on giving</strong>: Another day, another AWS US-EAST-1 outage, this time blamed on DNS resolution failures with DynamoDB. Major services like Alexa, Fortnite, and Zoom went down, proving once again that putting all your eggs in one cloud basket is a brilliant strategy. The Register even suggested ‘brain drain’ at Amazon, linking prolonged recovery to massive layoffs and loss of senior engineers. Who needs tribal knowledge when you have ‘retry failed requests’?</li>
</ul>
