LAI #84: Prompting as a Skill, DINOv2 Embeddings, and Claude vs. OLMo 2
No-code ML labs, AutoGen at scale, GPU internals, and a growing case for simpler reinforcement learning.
Good morning, AI enthusiasts,
This week’s issue starts at the foundation: prompting. As more teams adopt LLMs, the ability to shape outputs through structured prompting is becoming a core skill, more spreadsheet than science. In What’s AI, we walk through the techniques we teach in our AI for Business course.
You’ll also find a no-code local ML pipeline from the community, a deep dive into DINOv2 embeddings for visual classification, and a practical comparison between Claude 3.5 Sonnet and OLMo 2. We’ve also got:
A step-by-step guide to deploying multi-agent systems with AutoGen on Azure
A breakdown of CUDA vs. cuDNN for anyone trying to optimize under the hood
And a reminder that sometimes simpler RL methods outperform clustered ones in real-world applications
Plus: community collabs, memes, and this week’s poll on what would flip your opinion of Grok 4.
Let’s get into it.
What’s AI Weekly
This week in What’s AI, I am diving into the most foundational aspect of working with LLMs: Prompting. Prompt engineering has now become a standalone job title in some cases , but more broadly, it’s more of a new, valuable skill, much like knowing how to Google or use a spreadsheet. So, if you are unhappy with LLM outputs, before trying anything else, start by fixing your prompts, which is exactly what I share in this article. I provide an overview of the most popular and valuable prompting techniques that we also share in our course, AI for Business. Read it here or watch the video on YouTube.
— Louis-François Bouchard, Towards AI Co-founder & Head of Community
Learn AI Together Community Section!
Featured Community post from the Discord
Nickname2905 has created Angler ML, a local no-code ML lab designed to help you draft, tweak, and export machine learning pipelines completely offline. It can load CSVs, preprocess data, train models (Linear Regression, KNN, Ridge), export your model & even generate the Python code. Check it out and support a fellow community member. If you have any feature suggestions or feedback, share them in the thread!
AI poll of the week!
Grok 4 just topped ARC-AGI-2 and beat Claude Opus on HLE. Yet, 63% of this community still called it hype. That gap says something: the tech is real, but trust, pricing, and UX haven’t caught up. Yes, Grok 4 Heavy is a multi-agent system with reasoning breakthroughs. But it also costs $300/month and just had a public meltdown over a broken system prompt.
What would actually flip your vote to “amazing”? A new killer use case? A price drop? Better guardrails? Tell me in the thread!
Collaboration Opportunities
The Learn AI Together Discord community is flooded with collaboration opportunities. If you are excited to dive into applied AI, want a study partner, or even want to find a partner for your passion project, join the collaboration channel! Keep an eye on this section, too — we share cool opportunities every week!
1. Dhanush__45 wants to dive deeper into the cloud domain and is looking for someone who can help and study together. If this sounds like your thing, connect in the thread!
2. Jinj4 is looking for a collaborator who wants to dive deep into prompts and video generation. If you have been curious about this, reach out in the thread!
Meme of the week!
Meme shared by kolawole0952
TAI Curated section
Article of the week
Exploring Clustered Optimal Policies via Off-Policy Reinforcement Learning for Business Use Cases By Shenggang Li
This article compares four reinforcement learning methods for personalizing promotions using offline data: single-head DQN, single-head PPO, Fixed-K DQN, and adaptive-clustered PPO. The models were evaluated using inverse propensity scoring (IPS) to estimate potential profit. Across multiple trials, the single-head PPO model consistently delivered the highest returns. Its clipped surrogate objective provided more stable learning, whereas the more complex clustered PPO was hindered by sparse reward signals that prevented its gating mechanism from learning effectively. The findings suggest simpler, robust models can outperform intricate architectures in practical applications.
Our must-read articles
1. Harness DINOv2 Embeddings for Accurate Image Classification By Lihi Gur Arie, PhD
Leveraging pre-trained DINOv2 embeddings offers an effective method for image classification, especially when working with small datasets. The author demonstrates this by first using a zero-shot k-Nearest Neighbors (kNN) classifier on extracted features, which achieves 83.9% accuracy on a specialized microorganism dataset. To improve performance, a simple linear classification head is trained on these same embeddings, boosting accuracy to 95.8%. This process highlights the quality of DINOv2’s feature extraction for detailed visual tasks without requiring extensive labeled data.
2. OLMo 2 vs Claude 3.5 Sonnet: A Head-to-Head AI Showdown By Adi Insights and Innovations
An analysis of AllenAI’s OLMo 2 and Anthropic’s Claude 3.5 Sonnet highlights the distinct advantages of open-source versus proprietary AI models. The text details OLMo 2’s transparent, self-hosted architecture and contrasts it with Claude 3.5 Sonnet’s enterprise-focused, API-based approach. It includes practical coding comparisons where Claude 3.5 Sonnet often produces more comprehensive solutions. It also covers technical specifications, pricing models, and strategic use cases to help in selecting the appropriate model for different project needs.
3. Multi-Agent Systems with AutoGen on Azure By Naveen Krishnan
Transitioning a multi-agent system from a local proof-of-concept to a production environment requires a robust infrastructure. The author details an architecture for deploying Microsoft’s AutoGen framework on Azure, using services like AKS for scalability and Azure OpenAI for secure LLM operations. It provides a practical code example of collaborative agents handling research and file management tasks, along with essential practices for security, monitoring, and performance. It concludes with specific deployment configurations for containerization and Kubernetes.
4. CUDA vs cuDNN: The Dynamic Duo That Powers Your AI Dreams By Ojasva Goyal
While often mentioned together, CUDA and cuDNN play different but complementary roles in GPU acceleration. This article clarifies that CUDA is NVIDIA’s foundational platform for general-purpose parallel computing on GPUs. In contrast, cuDNN is a specialized library built upon it, providing highly optimized functions specifically for deep learning operations. This combination is essential for the performance behind many AI applications, from medical imaging to autonomous vehicles.
If you want to publish with Towards AI, check our guidelines and sign up. We will publish your work to our network if it meets our editorial policies and standards.