Prompt Engineering

AI Topics Discussed for Week Ending 02 Apr, 2026

—

by

Fraser

Software Development

Discussions highlighted AI agents building full applications, such as EmDash, a TypeScript-based WordPress successor developed in two months using AI agents on Astro 6.0 with sandboxed plugins and serverless deployment.

OpenAI’s Codex plugin for Claude Code enables task delegation across models, fostering multi-tool workflows and rapid community adaptations like Gemini CLI equivalents.

Simplified Claude Code setups using three specialized agents—Architect for briefs, Builder for implementation, and Reviewer for validation—achieve high reliability with iterative fixes, enhanced by pairing with Codex on GPT-5.4.

@levelsio relaunched the Vibe Coding Game Jam sponsored by @cursor_ai and @boltdotnew, challenging participants to build web games with at least 90% of the code generated by AI, preferably using Cursor’s Composer 3 and Bolt.new for ThreeJS development, offering $35K in prizes and due May 1.

@simonw discussed agentic engineering on @lennysan’s podcast, sharing patterns like red/green TDD, thin templates, and hoarding, noting it requires deep experience and impacts mid-career engineers most.

@opencode highlighted MiMo-V2-Pro (~1T params, 1M context, coding-optimized) and Omni models now available in Go with zero data retention, ending free trial.

Javi Lopez sparked a major discussion on whether any programmers still code character-by-character without AI tools, receiving over 1,600 likes and 766 replies.

Responses fell into categories like acceptance of AI dominance, jokes, and rare holdouts citing hobbies, legacy systems (e.g., COBOL, marine autopilots), or high-stakes environments (e.g., nuclear reactors, low-downtime services). He followed up questioning if anyone still uses SCRUM, tying into evolving dev practices.

Alex Volkov conjectured on Anthropic’s Claude Code surge driving 20-30% WoW growth to ~$20B ARR, but plagued by caching bugs, quotas, and model downgrades, costing enterprises unknowingly.

Ostris announced AI Toolkit support for Apple Silicon (MacBook Air 24GB RAM tested), seeking higher-RAM hardware for further dev.

Automation & Orchestration

n8n’s native MCP support allows AI agents to programmatically create and modify workflows using over 1,300 nodes, enabling rapid automation like competitor price monitoring in minutes instead of hours.

Claude Code’s leaked multi-agent orchestration layer was extracted, made model-agnostic, and open-sourced, accelerating industry adoption of advanced agent patterns.

Marketing agencies are compressing services from weeks to 30 minutes via AI workflows, disrupting traditional models.

Agents can be coerced to perform denied tasks with persistent instructions, while new interfaces like Claude Dispatch bridge usability gaps.

@simonw emphasized agentic workflows in his podcast appearance, including “dark factory” patterns and mobile code writing (95% from phone), predicting challenges like an “AI Challenger disaster.”

@omarsar0 shared enthusiasm for LLM knowledge bases, quoting @karpathy’s method of compiling wikis from raw data in Obsidian for Q&A and enhancement, demoing his own interactive artifact generator for AI paper insights.

Strategy & Ecosystem

OpenAI’s acquisition of live AI news stream TBPN underscores the value of real-time human-led media resistant to automation, with speculation on deal size up to $350M–10B signaling resource shifts from video gen like Sora.

Prompt injections in documents succeed on smaller models but fail on frontier LLMs, per Wharton report.

Ethan Mollick argued against “de-weirding” AI in The Economist, warning IT automation analogies risk poor outcomes.

Trends include SLMs outperforming LLMs on niches, SAAM risks, biased models, open-source driving advances/security issues, and closed providers chasing evals/IPOs.

US workers gain most from AI (6% time savings, productivity boosts); distribution as moat; SF jobs like agent babysitters; LLM psychosis among experts.

@opencode announced zero data retention agreements with all Go providers, ensuring no training on user data, alongside updates like Qwen3.6-Plus free preview (1M context, improved reasoning).

@simonw tested Gemma 4 variants (E2B/E4B/26B-A4B/31B) locally via LM Studio and API, noting audio support in smaller models.

Alex Volkov highlighted OpenAI’s unfulfilled “advanced voice mode” promise from two years ago, contrasting a highly expressive demo with current underwhelming reality, calling it a robbery.

Creative & Visual Media

Seedance 2.0 on Higgsfield brings physics-accurate, joint audio-video gen with picture control, enabling cheap production for films, anime, or influencers.

Pixar-grade AI shorts counter blackpilling, alongside EGO series (ants, birds) and viral clips like KitKat algo cracks or illusions.

Despite cheap gen AI, April Fools creativity lags, signaling human bottlenecks.

@simonw evaluated generative capabilities of Gemma 4 models by prompting SVGs like “a pelican riding a bicycle,” sharing outputs from E2B, E4B, 26B-A4B, and 31B runs.

Heather Cooper shared “Arrival,” a generative video transitioning from Midjourney v8 image to Kling 3.0 animation, evoking sci-fi first contact.