AI - Short & Sweet

New best os AI, Mustafa's warnings, best practices for AI Agents and Context Engineering

Author

Martin Musiol
August 20, 2025 • Estimated Reading Time: 7 minutes

In partnership with

Friends,

in this short and sweet episode, we’ve got a new contender for the strongest AI model on the scene—thanks to DeepSeek.

We’ll also touch on Mustafa Suleyman’s warning about “seemingly conscious” AI, practical best practices for building agents and multi‑agent systems, and Anthropic’s super solid prompting template you can apply today. ⬇️

Training Generative AI? It starts with the right data.

Your AI is only as good as the data you feed it. If you're building or fine-tuning generative models, Shutterstock offers enterprise-grade training data across images, video, 3D, audio, and templates—all rights-cleared and enriched with 20+ years of human-reviewed metadata.

With 600M+ assets and scalable licensing, our datasets help leading AI teams accelerate development, simplify procurement, and boost model performance—safely and efficiently.

Book a 30-minute discovery call to explore how our multimodal catalog supports smarter model training. Qualified decision-makers will receive a $100 Amazon gift card.

For complete terms and conditions, see the offer page.

(✨ If you don’t want ads like these, Premium is the solution. It is like you are buying me a Starbucks Iced Honey Apple Almondmilk Flat White a month.)

DeepSeek V3.1 just dropped — and it might be the most powerful open AI yet

You can download the model here.

Mustafa Suleyman warns about “seemingly conscious“ AI

A medium shot of Mustafa Suleyman, a man with a warm and friendly expression. He is smiling and looking slightly off-camera to his right. He has graying dark hair, a neatly trimmed beard and mustache, and is wearing thin, wire-rimmed glasses. His right hand is raised to his chin in a thoughtful gesture, with his fingers resting along his jawline. He is wearing a burnt-orange, wide-wale corduroy shirt. The background is softly out of focus, suggesting a modern setting like an office or storefront with dark-framed glass doors. The lighting is soft and natural, highlighting his face and creating a warm, approachable mood.

Microsoft AI CEO Mustafa Suleyman is warning that AI designed to appear conscious is a danger to both society and the future of AI.

Here's his argument:

  • The Threat is Here: Suleyman claims current technology can already create AI that convincingly simulates human memory, personality, and experience.

  • "AI Psychosis": This is leading to a rise in "AI psychosis," where users believe models are sentient and begin advocating for their rights.

  • A Dangerous Delusion: He calls the study of "model welfare" premature and dangerous, arguing it only fuels these delusions.

  • The Core Mandate: His message to the industry is to build AI "for people, not to be a person."

Why it matters: Suleyman’s hardline stance directly challenges companies like Anthropic that are actively researching model welfare. With consciousness itself still a scientific mystery, critics argue this position risks shutting down a crucial debate before it has even begun.

UserJot published the Best Practices for Building AI Agents, especially in Multi-Agent Systems

An insightful read, but here are the top 5 points-TLDR:

  1. Give Worker AI Agents Amnesia. Each specialized AI should tackle its task with a clean slate, remembering nothing from past jobs. This prevents them from getting confused and makes them more reliable.

  2. Keep a Simple Chain of Command. Use one "manager" AI that understands the big picture. This manager gives specific, one-off jobs to a team of "specialist" AIs. Avoid creating middle managers. (As in real life.)

  3. Break Projects into Baby Steps. Just like following a recipe, figure out which tasks need to happen in order and which can be done all at the same time. This gets the job done correctly and much faster.

  4. Give Crystal-Clear Instructions. Don't be vague. Tell each AI exactly what you want, what information to use, and how you want the answer back. The less room for guessing, the better the result.

  5. Put AIs on a Need-to-Know Basis. Don't overload a specialist AI with the entire backstory. Just give it the minimum amount of information it needs to do its immediate task. This keeps it focused and efficient.

Anthropic’s Super Solid Prompting Template (Context Engineering)

Based on this prompt structure, I have reiterated and improved all my prompts in my personal library. I can highly encourage you to go through that and do the same. I think this is a super valuable template.

A presentation slide titled "Prompt structure" that visually breaks down the components of an effective AI prompt. The slide is divided into two sections. On the left, a vertical list of ten numbered, color-coded rectangular boxes outlines the elements of a prompt structure: Task context (reddish-pink) Tone context (orange) Background data, documents, and images (light green) Detailed task description & rules (darker green) Examples (light blue) Conversation history (purple) Immediate task description or request (reddish-brown) Thinking step by step / take a deep breath (light pink) Output formatting (gray) Prefilled response (if any) (light gray) On the right, a larger section labeled "User" provides an example of a complete prompt for an AI career coach named Joe. The text within this prompt is highlighted with background colors that correspond to the numbered list on the left. The prompt defines Joe's persona, tone, background documents, rules for interaction, an example dialogue, and placeholders for conversation history and the user's question. Below the main prompt text, it instructs the AI to "Think about your answer first before you respond" and to place its output within <response> tags. At the very bottom right, a smaller box labeled "Assistant (prefill)" shows the starting tag <response>, indicating where the AI's generated answer would begin.

It was found in this great 101 prompting course.

I hope you enjoyed it.

Happy weekend!
🙇Martin

Our webpage