Prompt Engineering

AI Topics Discussed on 21 Feb, 2026

—

by

Fraser

AI Topics Discussed on 21 Feb, 2026

Creative & Visual Media

@levelsio shared updates to Photo AI, introducing a model editor that allows users to set default prompts (e.g., “wearing thin frame glasses and blue hair”) applied to every photo generated with that model, and manually select key photos from the training set for higher resemblance and realism.

Peter Omallet noted that the percentage of people able to perceive improvements in generative image and music models is declining rapidly, predicting it will be below 1% across major modalities by year-end.

Software Development

The Boring Marketer shared pro tips for OpenClaw, recommending Sonnet 4.6 for coding tasks and using Claude Code to SSH into VPS for debugging and optimization.

EXM7777 emphasized Claude Code’s ability to ship full DApps in days and audit smart contracts, criticizing crypto’s underuse of such AI-assisted engineering tools.

Automation & Orchestration

@simonw described “Claw” as an emerging term for OpenClaw-like agent systems that run on personal hardware, handle orchestration, scheduling, context, tool calls, and persistence—positioning them as a new layer atop LLM agents. He highlighted variants like NanoClaw (with ~4K lines of code, containerized, and configurable via AI skills that fork the repo) alongside others such as nanobot, zeroclaw, ironclaw, and picoclaw, noting preferences for local setups to connect with home devices.

EXM7777 advocated heavily for “deep research” features in models like ChatGPT Pro to enhance workflows in content, outreach, strategy, and product research, urging 5-10 daily runs backed by data, SOPs, and frameworks.

They also promoted building ICP subagents in OpenClaw trained on customer data sources for content and ad feedback, with ongoing market monitoring.

Ethan Mollick discussed AI’s impact on phone OS, favoring conversational agents for most tasks beyond legacy apps.

OpenClaw tips included spawning subagents with various models like Gemini and Opus.

Strategy & Ecosystem

@omarsar0 covered a new Google paper proposing “deep-thinking tokens” to measure genuine LLM reasoning effort by detecting prediction instability across transformer layers, outperforming token count on benchmarks like AIME, HMMT, and GPQA. The paper also introduces Think@n for efficient test-time compute by prioritizing high deep-thinking samples.

Ethan Mollick called out accounts reposting old 2025 AI “failure” papers (e.g., Apple paper, model collapse) as new with AI commentary, noting their outsized buzz despite irrelevance amid model progress.

The Boring Marketer warned of “false productivity” as a key risk with current AI tools.

Peter Omallet highlighted open source AI’s potential against concentrated capital in entertainment.

EXM7777 shared an article on AI strategies, garnering high engagement.

Mollick expressed skepticism toward an unspecified AI development or output.