#61: Are LLMs Entering the Age of Agents?

Author(s): Towards AI Editorial Team

Originally published on Towards AI.

Good morning, AI enthusiasts! Reasoning agents seem to have taken over AI in the last couple of weeks. While it is early, this class of reasoning-powered agents is likely to progress LLM adoption and economic impact to the next level. So, let’s start with the basics this week by first understanding what exactly an agent is.

We also have plenty of resources covering everything from design, architecture, and applications of Llama, comparison between OpenAI GPT, DeepSeek, and Qwen2.5 to Artificial Super Intelligence and “black box” deep learning models. Enjoy the read!

What’s AI Weekly

The vast majority of what we call Agents are simply an API call to a language model. These cannot act independently, make decisions, or do anything. It simply replies to your users. Still, we call them agents. But this isn’t what we need. We need real agents, but what is a real agent? In simple terms, a real agent is something that functions independently. That’s exactly what I will dive into this week. Read the complete article here or watch the video on YouTube.

— Louis-François Bouchard, Towards AI Co-founder & Head of Community

Learn AI Together Community section!

AI poll of the week!

The poll results are surprising. Most of you believe the alternatives are pretty similar. Are these o1, Deepseek, or any others? Tell us in the thread!

Collaboration Opportunities

The Learn AI Together Discord community is flooding 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. Mongong is developing a framework that uses hypergraph-based AI, tensor embeddings, and adaptive reasoning to solve complex problems across domains. If you would like to join the project, connect in the thread!

2. Thebaburuidas is building an AI-powered fashion assistant that uses image recognition and is seeking a co-founder. If this sounds interesting, reach out in the thread!

Meme of the week!

Meme shared by hitoriarchie

TAI Curated section

Article of the week

Unlocking the Potential of Meta LLaMA: A Deep Dive into Its Design, Architecture, and Applications By Shenggang Li

This article explores Meta’s Llama, a large language model designed for efficiency and accessibility. It details the underlying Transformer architecture, including self-attention mechanisms, positional embeddings, and feed-forward networks, explaining how these components contribute to Llama’s capabilities. It highlights Llama’s optimizations, such as Rotary Positional Embeddings (RoPE), for improved long-sequence handling and weight tying for reduced parameter count. It also discusses Llama’s training process, including the causal language modeling objective and techniques used for efficient training on large datasets. Finally, it provides practical guidance on implementing Llama using Ollama, Hugging Face Transformers, or cloud platforms like AWS or Azure, outlining the advantages and limitations of each approach.

Our must-read articles

1. Langchain (Upgraded) + DeepSeek-R1 + RAG Just Revolutionized AI Forever By Gao Dalie (高達烈)

This article discusses the creation of a RAG (Retrieval-Augmented Generation) chatbot using LangChain, DeepSeek-R1, and FalkorDB. DeepSeek-R1, a newly released open-source large language model, is highlighted for its cost-effectiveness and performance comparable to OpenAI’s o1 model. It details building the chatbot, emphasizing the use of LangChain’s features: PDFPlumberLoader for PDF processing, SemanticChunker for text splitting, FAISS for efficient search, and the creation of retrieval chains for context-aware responses. The chatbot handles document uploads, extracts information, and generates responses based on user queries and conversation history. It also covers DeepSeek-R1’s unique training method, using reinforcement learning without supervised fine-tuning.

2. Artificial Super Intelligence (ASI): The Research Frontiers to Achieve AGI to ASI and the Challenges for Humanity By Mohit Sewak, Ph.D

This article explores the potential pathways to Artificial Super Intelligence (ASI), examining scaled-up deep learning, neuro-symbolic AI, cognitive architectures, whole brain emulation, and evolutionary algorithms. It also details significant challenges, including understanding consciousness, solving the hard problem of intelligence, ensuring control and value alignment, managing emergent behaviors, and mitigating existential risks. It concludes by considering the potential utopian benefits of ASI — solving global problems, accelerating scientific discovery, and boosting productivity — while acknowledging the dystopian risks of mass unemployment, power concentration, and existential threats. Ultimately, it emphasizes the need for responsible development, global collaboration, and ethical considerations to navigate this transformative technology.

3. Event-Driven Prediction: Expanding Mamba State Space Models for Conditional Forecasting By Shenggang Li

This article presents a novel approach to conditional time series forecasting, particularly for stock prices. It extends the Mamba state space model by incorporating event-driven dynamics inspired by Markov Decision Processes. This allows the model to predict future price movements (both continuous and binary) based on past data and specific future events, such as a price crossing a key threshold. The method is demonstrated using simulated data with promising results, measured by metrics like MSE, MAPE, AUC, F1 score, and KS statistic. The author also suggests future work will involve real-world data and explore more complex event modeling. The resulting model offers valuable insights for traders and investors in risk management, strategy development, and market surveillance, with potential applications in portfolio optimization and high-frequency trading.

4. How to Explain Black-Box Deep Learning Models in Computer Vision and NLP By Chien Vu

This article discusses the challenges of interpreting “black box” deep learning models in computer vision and natural language processing (NLP). It highlights the importance of explainability and interpretability for various stakeholders, including data scientists, business leaders, and regulators. It then introduces OmniXAI, a Python library offering various explanation methods. Using examples, it demonstrates how OmniXAI’s techniques — such as SHAP, LIME, Integrated Gradients, Grad-CAM, Score-CAM, and counterfactual explanations — can be applied to analyze model predictions for image classification and sentiment analysis, revealing which features contribute most significantly to model decisions and offering insights into potential model biases. The article also emphasizes the role of explainability tools like OmniXAI in building more trustworthy and accountable AI systems.

5. Rotating Box Challenge: Why OpenAI GPT Beat DeepSeek and Qwen2.5 Hands Down By Tarun Singh

This article compares three AI models — OpenAI GPT, DeepSeek, and Qwen2.5 — using a rotating box physics simulation as a benchmark. The author provided identical prompts requesting a Pygame-based simulation. OpenAI GPT produced a flawless simulation, demonstrating accurate collision detection and realistic physics. DeepSeek’s and Qwen2.5’s attempts failed to meet the specified criteria, with DeepSeek’s simulation collapsing entirely and Qwen2.5’s exhibiting flawed collision handling. The results highlight OpenAI GPT’s superior ability to understand and execute complex prompts in a practical context.

If you are interested in publishing 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.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI