Prompt Engineering

AI Topics Discussed on 28 Jan, 2026

—

by

Fraser

Creative & Visual Media

Ostris (@ostrisai) tested training a LoRA via student-teacher distillation using Z-Image (base) as student and Z-Image-Turbo as teacher, producing shocking results at 5k steps, batch 1, with 8-step inference samples showing clear before/after improvements.

Gokay (@gokayfem) noted fal.ai’s new “Chasing 6+ TB/s: an MXFP8 quantizer on Blackwell” kernel drop, hinting at upcoming releases leveraging it for high-performance inference.

Cristóbal Valenzuela (@c_valenzuelab) promoted the 4th Runway AI Festival, now expanded to celebrate AI works in film, design, new media, fashion, advertising, and gaming, with submissions open.

Heather Cooper (@hbcoop_) reviewed Genspark AI Workspace, which hit $100M ARR in 9 months as an all-in-one alternative to juggling ChatGPT, Midjourney, Canva, and multiple tabs.

fal.ai announced several new generative image models available on their platform, including HunyuanImage 3.0-Instruct with chain-of-thought reasoning for better blending and intent alignment,

Qwen Image Max for realistic faces, text rendering, and editing,

and Z-Image base trainer for high-quality LoRA training.

AIWarper tested Luma Lab’s Ray 3.14 video-to-video update using a middle frame insertion

and launched a fully automated Instagram “slop channel” pipeline for rage-bait content involving destroying collectibles.

Machina (@EXM7777) highlighted generating mascot assets using Nano Banana Pro as part of an AI workflow for content production.

Software Development

Simon Willison highlighted Dan Shapiro’s model of five levels of AI-assisted programming, from autocomplete to a “dark software factory,” adding notes on the final level.

Omar Sar noted NVIDIA’s VibeTensor, a full deep learning stack (PyTorch-style with CUDA, autograd, etc.) autonomously built by LLM coding agents, spanning 63k+ lines of code, with benchmarks showing speedups in kernels.

Machina (@EXM7777) praised Kimi K2.5’s UI/UX for building websites, noting its real-time reasoning, 3D visuals, todo lists, and component generation, outperforming Gemini in following instructions and web research for React landing pages.

Ethan Mollick (@emollick) shared observations from coders indicating a shift toward AI-assisted engineering, with leaders at OpenAI, Cursor, and others reporting less hands-on coding and more oversight.

Automation & Orchestration

Alex Volkov (@altryne) praised Clawd/Moltbot’s multi-threaded capabilities for handling separate conversations across Telegram/Slack/Discord threads with independent memory and sequences, avoiding context overload.

He shared mind-blowing demos of Clawdbot autonomously handling unsupported voice memos by inspecting file headers, converting with FFmpeg, detecting missing Whisper, and curling OpenAI API—plus tips on enabling memory flush and session search for better persistence.

Fofr (@fofrAI) demonstrated Gemini 3’s Agentic Vision in AI Studio, where it executes code to add bounding boxes, labels, and arrows to images for tasks like cleaning instructions.

Omar Sar discussed Clawdbot’s success with agentic loops and tools, urging developers to build custom agentic harnesses for reliability, personalization, and control using Claude Code.

Machina (@EXM7777) detailed a one-shot landing page workflow using Kimi K2.5 agent: deep research for ICP/product profiles, injected copywriting skills, mascot generation, and full React page build; also launched Real-Time AI Ops Community for weekly workflows and agentic systems.

Ethan Mollick (@emollick) discussed a multimodal Gemini 2.5 agent matching or exceeding 14,000 medical students in a physician training sim for case completion, time, and diagnostic accuracy.

Strategy & Ecosystem

Ostris (@ostrisai) inquired about the current state-of-the-art in training step-distilled diffusion models.

@levelsio observed Claude targeting high-IQ creative users (e.g., artsy non-coders via sponsored content), positioning against ChatGPT’s “normie” image—the first clear market differentiation among major AI apps.

Ethan Mollick (@emollick) covered trends including a paper analyzing global job ads showing AI’s impact on labor markets, particularly devaluing expertise; funding for alternative architectures beyond LLMs; and Riley Brown (@rileybrown) emphasized Meta’s acquisition of Manus as a key event.

Machina (@EXM7777) promoted AI upskilling via a community focused on meta-skills, real-world workflows, and leading the AI space.

OpenRouter (@OpenRouterAI) hosted a live session with Arcee AI’s CTO.