Prompt Engineering

AI Topics Discussed on 07 Mar, 2026

—

by

Fraser

Creative & Visual Media

Heather Cooper (@HBCoop_) demonstrated advancements in generative video models by transforming a Midjourney image into a Veo 3.1 video clip. The prompt specified a woman in low-key lighting exhaling and tucking hair behind her ear, with precise control over motion and a still camera.

Omar Sar noted rapid improvements in AI long-form video generation after gaining early access to Utopai Studios’ new PAI model, praising its capabilities and editing tools.

Justine Moore (@venturetwins) highlighted major advances in Kling’s new Motion Control feature for generative video, particularly its impressive character swap capabilities. She shared a video demo using “known” faces (hard mode for consistency), noting it handled tricky shots with multiple angles well, improving facial consistency across cuts and reducing warping—especially with face-locking.

Software Development

Alex Volkov (@altryne) highlighted AI-assisted coding workflows using Codex on Windows, integrated with Raycast and GPT 5.4 native sandboxes. He described building three projects simultaneously on a 115″ projector screen, calling it “ridiculously overpowered.”

Jonathan Fischoff suggested that Cursor should focus on aggregating intelligence from multiple models rather than developing its own foundational model like Opus or GPT, leveraging its position in AI-assisted coding.

GPT 5.4 emerged as a standout model, praised by Justine Moore (@venturetwins) for its coding prowess and general utility, feeling like conversing with a “smart friend” for explaining concepts or troubleshooting.

Ethan Mollick (@emollick) noted its role alongside tools like Claude Code and Codex in boosting research assistants’ workflows.

Automation & Orchestration

Omar Sar highlighted research on automatic harness synthesis for LLM agents, emphasizing how automating the scaffolding for tools, code execution, and APIs could lower barriers to agent development.

He also shared a survey on agentic reinforcement learning for LLMs, proposing a taxonomy for capabilities like planning, memory, and tool use in open-ended environments.

Machina (@EXM7777) advocated building personal knowledge bases via note-taking (Notion, Obsidian) as persistent context for agents, turning SOPs and thinking into superior AI outputs.

He also predicted “model consensus” as the future: multiple provider agents reasoning in parallel, merged by an orchestrator to highlight agreements/divergences, akin to Perplexity’s model council.

Strategy & Ecosystem

Javi Lopez (@javilopen) critiqued New York’s regulation mandating disclosure of AI models in advertisements starting June 9, 2026, with fines up to $5,000 per violation. He argued it unfairly targets AI compared to 3D VFX or mixed media, while malicious deepfakes are already covered by existing laws.

Omar Sar discussed Yann LeCun’s research on Transformer phenomena like massive activations and attention sinks, attributing them to pre-norm architecture and their implications for quantization, pruning, and efficient inference.

Simon Willison pointed out Qwen 3.5’s 4B model outperforming GPT-4o on benchmarks while running on iPhones, expressing suspicion of potential test-specific training.

A.I. Warper sought recommendations for the best real-time text-to-speech solutions.

Ethan Mollick (@emollick) emphasized AI’s transformative impact on science, enabling ambitious questions and RA upskilling with tools like Claude Code—shifting from replacement fears to expanded reach.

He critiqued persistent misinformation on outdated papers (e.g., 2025 hallucination study), noting rapid progress on benchmarks like SimpleBench.

Pre-AI errors (e.g., Excel mangling gene names in 1/3 of top genetics papers) underscore how AI can already outperform imperfect human baselines.