Prompt Engineering

AI Topics Discussed on 22 Jan, 2026

—

by

Fraser

Qwen3-TTS Model Release

Simon Willison detailed his experiments with Alibaba’s new Apache 2.0 licensed Qwen3-TTS model, highlighting its strong voice cloning capabilities and effectiveness.

He adapted MLX-Audio code into a CLI tool for generating speech with custom voices, such as a pirate impression.

Advances in Coding Agents and Workflows

@levelsio revived his plane game (fly.pieter.com) using current AI models to overhaul the flight physics for realism, including stalling and takeoff fixes in just two prompts—capabilities last year’s models couldn’t handle.

He integrated Claude Code with Chrome to enable the agent to control keys, host a local server, refresh pages, and autonomously fly the plane for iterative improvements.

Simon Willison shared a TIL on using Claude Code on iPhone with GitHub Pages for real-time previews of changes during agent execution.

Vaibhav Srivastav (@reach_vb) described his Codex workflow: finding verifiable loops, minimal context, iteration, and agent MD updates for drift.

He announced Codex integration into JetBrains IDEs (IntelliJ, PyCharm, etc.) powered by GPT-5.2, available via ChatGPT plans, with app server embedding support.

Dan Shipper (@danshipper) hosted Vibe Code Camp, a live marathon featuring AI builders demoing agent-native apps, Claude Code, compound engineering, and more.

@goose_oss announced multichat support (up to 10 chats in sidebar) and built-in data safety checks against LLM manipulation via red-teaming.

Generative Media Finetunes and Contests

A.I.Warper (@AIWarper) showcased a Klein 9B finetune on 80s anime, producing stylistic effects on existing anime images.

@fal partnered with Alibaba Cloud for a Winter Olympics fan video contest using their Wan model: create 5-15s landscape videos inspired by sports like figure skating, for prizes including tickets and museum features.

Agentic Reasoning Frameworks

Elvis (@omarsar0) highlighted a comprehensive 135+ page survey on agentic reasoning in LLMs, covering foundational (planning/tools), self-evolving (reflection/memory), and multi-agent paradigms, distinguishing in-context vs. post-training optimization; applications include math, robotics, and web research.

Economics in the AGI Era

Emad Mostaque (@EMostaque) argued conventional economics fails post-AGI, promoting his single-axiom framework and policy recommendations via “The Last Economy” book; forthcoming papers, tools, and studies.

Claude Code and Agentic Development

Discussions highlighted Claude Code’s capabilities for rapid app building, with Riley Brown demonstrating a ManusAI clone created in 30 minutes using the Claude Agent SDK, integrating web search, Slack, Notion, Nano Banana, and Remotion skills.

Machina shared a GitHub repo with 30 premium Claude Code skills and advised using Claude projects for context in long-term building goals, like committing to one project for 6 months.

CJ Zafir used Claude Code (Opus 4.5) to analyze X data against the open-sourced X algorithm, creating a “X verifier suite” for post optimization after a 6-hour run costing $6.78.

Ethan Mollick noted that views on AI capabilities are skewed by CLI tools like Claude Code/Codex, predicting better agentic harnesses will transform frontier model performance.

Model Usage and Performance Optimization

Ethan Mollick emphasized that few users switch from default models, significantly boosting AI value with a simple change, and praised GPT-5.2 Pro for checking academic paper reproducibility and robustness (though not full replications).

OpenRouterAI launched regional LLM performance tracking by provider and API stats parsing with Claude Code, plus AI-friendly docs updates.

Riley Brown contrasted average Claude vs. ChatGPT users, predicting OpenAI’s direction will widen the gap.

AI in Academic Research and Language Influence

Ethan Mollick highlighted GPT-5.2 Pro’s role in flagging reproducibility issues in papers across fields, recommending authors use it for self-review.

He shared analysis of 280k academic video transcripts showing increasing use of ChatGPT-favored words, dubbing it “model collapse, except for humans.”

Mollick also discussed accumulating AI skillsets from experience, countering claims that catching up takes only a day.

AI Influencers and Monetization

Justine Moore detailed Fanvue’s $100M ARR growth via AI influencers, using photorealistic models and motion transfer (e.g., NB Pro, Kling), with IG Reels driving subscriptions; one creator gained 700k followers in 6 weeks.

Machina amplified an AI influencer blueprint and noted quick mastery (7-8 weeks) suffices for monetization, dismissing “years to master AI” narratives.

Prompting and Workflow Best Practices

Machina warned against over-relying on Claude projects/SaaS for prompts (leads to quality loss), advocating manual AI-assisted refinement across models; also suggested interactive AI summaries for YouTube transcripts.

The Boring Marketer released a free “vibe marketing playbook” with Claude Code examples for agencies/SaaS/ecom, including prompts, skills, and outputs.

Ethan Mollick proposed terms like “Marcus Chen-ing” for AI-directed human actions.