Videos

This article, presented by WEKA's Chief AI Officer Val Bercovici and product lead Callan Fox, introduces "Context Platform Engineering" as a critical approach to combat "Token Anxiety" and "Prompt Cache Arbitrage" in AI agent systems. They announce the release of WEKA's open-source toolkit designed to help engineers optimize agent swarms and subtasks by translating Service Level Agreements (SLAs) into executable Service Level Objectives (SLOs). The discussion delves into WEKA Labs' research, visualizing agent loops, analyzing KV Cache hit rates across different memory tiers (HBM, DRAM, WEKA's Augmented Memory Grid), and presenting benchmark results. The core argument is that maximizing KV Cache hit rates is the single most important metric for production-grade AI agents, and efficient tiered memory solutions are key to achieving this, thereby maximizing efficiency and reducing costs for AI inference service providers. The toolkit allows for configuring agent swarms with specific SLOs, simulating prompt cycles, and exploring memory tiering options.

This fireside chat between LangChain's CEO Harrison Chase and Rippling's AI Head Ankur Bhatt provides deep insights into operationalizing AI agents within a large enterprise suite, covering HR, IT, and Finance. Ankur details Rippling's journey from simple content summarization to complex standalone AI products like Talent Signal and internal productivity tools. Key discussions include the 'prototype to production' process, the necessity of feedback loops with real production data, and the evolving approaches to context engineering. They explore the future of agent systems, debating the balance between deterministic workflows and 'deep agents' that leverage LLM reasoning for edge cases. The session concludes with a Q&A on security, commercialization, code quality, and tool management, offering practical advice for AI practitioners.

Ahmad Awais presents CommandCode, an innovative AI coding agent designed to transcend generic LLM outputs by learning and embodying a developer's specific coding "taste" and preferences. He demonstrates how CommandCode, through its Meta Neuro-Symbolic architecture, observes and adapts to a developer's implicit coding conventions, such as preferred languages (TypeScript), tools (Commander, pnpm), and project structures (e.g., separate command directories, Vitest for testing). Awais highlights the limitations of current "sloppy" AI agents that often produce generic or "vibe-coded" solutions requiring extensive manual correction. CommandCode aims to capture the "invisible decision architecture" that defines high-quality code for an individual or team, enabling faster, more consistent, and personalized code generation. The underlying architecture involves a Meta Neuro-Symbolic Reasoning Space with a KL Divergence Loop for continuous learning from both explicit and implicit feedback, coupled with reflective context engineering for self-awareness and adaptation. The vision extends to creating a "taste ecosystem" where developers can share or adopt specific coding styles, significantly accelerating development workflows and reducing code review times by generating code that inherently aligns with established best practices and personal styles.

This in-depth interview with Jay Parikh, Microsoft's EVP of Core AI, explores the company's vision for AI's future. Parikh details the formation of the new Core AI team, its focus on empowering 'builders' with an evolving AI tech stack, and the return to in-office work to foster rapid learning in a fast-changing field. He discusses how AI is blurring traditional engineering roles and enabling 'agent teams' to orchestrate complex tasks. The conversation also addresses critical infrastructure concerns like data center power limitations and GPU utilization, emphasizing efficiency and a multi-model deployment strategy for enterprises. Parikh shares insights on the interplay between open and closed-source models, Microsoft's collaboration with OpenAI, and the paramount importance of AI safety and observability from inception. A key 'contrarian view' highlights the need to measure AI's impact by its ability to solve previously intractable problems, rather than superficial metrics like lines of code generated.

This interview features Sherwin Wu, OpenAI's Head of Platform Engineering, discussing the company's massive scale, reaching 800 million weekly users. Wu outlines OpenAI's strategic shift from a singular AGI model focus to a diversified portfolio of specialized models, including GPT-5, o3, and Sora 2. He elaborates on the evolution of their developer platform, highlighting Reinforcement Fine-Tuning (RFT) as a significant breakthrough over traditional supervised fine-tuning (SFT) for leveraging proprietary customer data, even piloting data-sharing for discounts. The conversation touches on the 'anti-disintermediation' nature of large models due to strong user-model stickiness and deep developer integration. Wu clarifies that 'prompt engineering' is evolving into 'context engineering,' emphasizing tool orchestration and data retrieval. He defines an AI agent as capable of acting on a user's behalf over time and introduces OpenAI's node-based Agent Builder designed for controlled, deterministic workflows in structured tasks. The discussion also covers OpenAI's usage-based pricing philosophy, the strategic benefits of open-source models, and the distinct technical stacks for text, image, and video models.

Matthew Berman provides a comprehensive review of Anthropic's recently released Claude Opus 4.5, positioning it as a leading contender for coding agents and computer use tasks. The article details Opus 4.5's performance across various benchmarks, including SWE-bench Verified, Terminal Bench 2.0, and T2 Bench, where it often outperforms competitors like Gemini 3 Pro and GPT 5.1 Codex Max. While excelling in coding and agentic tasks, the model shows areas where it doesn't lead, such as GPQA Diamond, MMMU, and MMLU. A significant highlight is Opus 4.5's ability to pass Anthropic's notoriously difficult performance engineer take-home exam, surpassing all human candidates. The article also covers the model's higher pricing compared to Gemini 3 Pro and introduces new developer features like 'Advanced Tool Use' and 'Tool Search,' which drastically optimize context window efficiency by allowing the model to dynamically search for tools rather than loading all definitions upfront. This innovation, coupled with its 'intelligence per token' efficiency, represents a major step forward for LLM development and application.

This video summary from Anthropic introduces "Agent Skills," a new feature for Claude that allows users to package portable, specialized knowledge for AI agents. Otto from Anthropic explains how these skills function by dynamically loading relevant `skill.md` files and associated scripts into Claude's context only when needed, employing a "progressive disclosure" mechanism to avoid context window bloat. The article meticulously differentiates Agent Skills from other Claude features like `Claude.md` files (project-specific info), MCP Servers (external data integration), and Subagents (fixed-role AI assistants). It highlights how skills provide transferable expertise (e.g., design standards, query optimization patterns) that can be shared across various agents and projects. Ultimately, Agent Skills enable the creation of reusable, specialized workflows, making Claude more intelligent and adaptable for complex tasks.

The article summarizes a practical comparison between Claude Opus 4.5 and Gemini 3 Pro, conducted by Greg Isenberg and James ("Boring Marketer"). They challenged both AI models to build a high-conversion landing page and a clickable prototype for a hypothetical SaaS product called "EstateClear," a family dashboard for probate management. The test aimed to evaluate the models' "Vibe Coding" capabilities and design prowess with minimal guidance. Claude Opus 4.5, especially when augmented with a "Front-end Design Skill," consistently produced more aesthetically pleasing, production-ready designs and demonstrated deeper product thinking in its prototype. Gemini 3 Pro, while clean, tended to embed AI-related features and benefited from Google's vertically integrated AI stack (AI Studio, Anti-Gravity, Nano Banana Pro). The discussion highlights the power of custom "Skills" in Claude for tailoring outputs and the increasing blurring of lines between technical and non-technical roles in product development. Overall, Opus 4.5 emerged as the stronger performer for generating refined, functional front-end assets.

The article provides a detailed review of Anthropic's new flagship model, Claude Opus 4.5, positioning it as a strong competitor to Gemini 3 Pro. It emphasizes the model's focus on coding, agents, and practical computer use, alongside a substantial price reduction (from $15/$75 to $5/$25 per million tokens for input/output). Benchmarking results show Opus 4.5's impressive gains in Aider Polyglot (89.4%), SWE-bench Verified (80.9%), and Terminal Bench 2.0 (59.3%) for code and agentic programming. The model also demonstrates strong performance in multilingual programming, long-term coherence, and deep research agents, while exhibiting improved safety and robust general reasoning capabilities. Live tests showcased its excellence in full-stack application development, particularly backend tasks, but noted weaknesses in frontend UI generation. A cost analysis reveals Opus 4.5's higher cost compared to Gemini 3 for similar performance, suggesting a hybrid approach where Opus handles backend and complex tasks, while Gemini 3 addresses frontend development.

Social and organizational psychologist Jennifer Parlamis debunks the myth that venting anger is a healthy release. Drawing from a personal anecdote and subsequent research, she explains that verbal catharsis often intensifies anger by repeatedly rehearsing negative 'causal attributions'—the internal, controllable explanations we assign to others' behaviors. Parlamis highlights the 'fundamental attribution error' and the self-reinforcing cycle between attributions and anger. She argues that anger is a cognitive construct, not something that simply happens to us. While venting provides a social and relational benefit (feeling heard), it fails to reduce anger itself. Instead, a meta-analysis reveals that physiologically arousing activities (like shouting or intense exercise) are ineffective. Effective anger regulation involves low-arousal activities (meditation, deep breathing, yoga), consciously examining causal attributions, actively seeking new information, and taking deliberate, controlled action. The talk emphasizes that anger is a useful emotion when regulated, empowering individuals to manage their emotions and build stronger relationships.