Prompt Engineering

AI Topics Discussed on 09 Mar, 2026

—

by

Fraser

Creative & Visual Media

Runway released real-time video agents built on their GWM-1 world model, enabling expressive digital personas from a single image without fine-tuning.

The BBC used Runway Characters for live television segments.

fofr shared luminism-style images generated with Nano Banana Pro.

fal announced the release of OneReward, a new model on their platform for unified image inpainting including fill, image extension, and object removal, with superior text rendering and mask-guided generation outperforming prior models.

They also promoted events at GDC with Kindred Ventures on AI reshaping games and media, plus GTC meetups.

AIWarper shared MatAnyone2 for advanced video matting, integrable with tools like CorridorKey for improved chroma keying in generative workflows.

Justine Moore highlighted GPT-5.4’s prowess in reviewing slides, providing expert-level teardowns and narrative improvements.

Ilker shared a workflow for Kling v3 multi-shot video generation, automating prompt creation for easier video production.

Machina (@EXM7777) discussed optimizing prompts for Nano Banana image generation using high-quality reference images from sources like Cosmos and Pinterest.

Riley Brown noted rising competition in content creation due to AI tools and layoffs driving more creators to platforms like IG, TT, and YT.

Software Development

Guillermo Rauch highlighted OB-1, a self-improving coding agent now integrating Vercel CLI for deploying and managing projects.

omarsar0 highlighted concerns with Claude Code’s new review feature, emphasizing separation of code generation (output-focused) from verification (skepticism-focused) to avoid shared blind spots, promoting QodoAI’s independent review tools.

simonw noted recent LLMs with coding agents excel at niche tech like Datasette/SQLite, countering predictions of “boring” preferences, and polled on AI tools in dev interviews.

jfischoff commented on Claude’s agentic code review teams.

Machina praised Anthropic’s new Code Review feature in Claude Code, deploying agent teams for bug hunting in PRs, positioning it as a top coding agent.

The Boring Marketer recommended Claude Code’s /loop skill for recurring tasks like PR monitoring and deployments, and a builder-judge setup with Claude Code and Codex for effective multi-agent development.

Ethan Mollick showcased Claude Code building an interactive lighthouse atlas app.

CJ Zafir detailed using Codex-5.4 and Opus-4.6 for model distillation and fine-tuning smaller Qwen models, surpassing GPT-4o and others locally.

Riley Brown demonstrated Claude Code and OpenClaw using a Figma-like tool (@paper) for agentic design and React app building.

Automation & Orchestration

Alex Volkov commented on Andrej Karpathy’s autoresearch, an agentic system where AI autonomously iterates on LLM training code for rapid improvements.

Rauch praised AI agent simulations for government scenarios as an underrated use case.

omarsar0 discussed Memex(RL) for scaling agent memory in long-horizon tasks via indexed experience retrieval, and KARL from Databricks for RL-trained enterprise search agents handling synthesis and reasoning.

He teased upcoming agentic research work and stressed persistent memory for complex planning.

Machina highlighted agentic commerce advances, with OpenClaw enabling fully autonomous app building, domain purchase, hosting, and ad campaigns; Lemon’s voice agent for device tasks; and Airwallex’s infrastructure for agents.

The Boring Marketer emphasized judgement in AI workflows, turning operators into systems thinkers for prioritization.

Ethan Mollick discussed shifting focus to AI-driven organizational efficiency over individual productivity.

Strategy & Ecosystem

Volkov teased upcoming Weights & Biases features and noted his role judging the Mistral AI worldwide hackathon.

Rauch echoed Naval Ravikant’s trend observation on computers evolving into tools for AI agents, linking to Vercel Sandbox.

jfischoff observed intense AI model competition post-NeurIPS, with rapid advances like Gemini 3 feeling outdated.

danshipper announced a 1-click Slack-based OpenClaw setup via Every, addressing multi-tenant demands.

simonw contrasted SQLite’s long-standing production query plan simulation with PostgreSQL 18’s new feature.

Justine Moore gauged community sentiment on AGI progress this week.

Ethan Mollick critiqued centralized “responsible AI” governance tangling risk discussions and called for distributed decisions; questioned Microsoft Cowork’s model access amid rapid advances like GPT-5 to 5.4; and sought verification on fruit fly brain upload claims.

Ethan Mollick also promoted prompt improvement techniques.