Artificial Intelligence

Author(s): Chew Loong Nian – AI ENGINEER Originally published on Towards AI. fx declares seventeen built-in tools, and a full turn with a subagent host carries fifteen function schemas. I measured a stock four-server MCP setup at 37 tools and 22,226 bytes — the exact payload this design keeps out of the request. Every few weeks I go through the same small humiliation. I add an MCP server to a coding agent because it looks useful, then I add a second one, then a third, and somewhere around the fo

Author(s): Decoding AI Originally published on Towards AI. 2,436 findings is a discovery number. Nothing in it is a remediation number. There is a particular kind of silence that follows a very productive week. On August 14, 2026, the AI lab Z.ai published a model and a ledger. The model was GLM-5.3. The ledger listed 2,436 software vulnerabilities the lab says its models found across 269 open-source projects (a cumulative count running from GLM-5.2 through GLM-5.3, rather than the output of a s

Author(s): Decoding AI Originally published on Towards AI. We refit Anthropic’s published per-target protein design counts on Aug 25, 2026. Ninety designs went into the wet lab against maltose-binding protein. Ninety came back with nothing. That is the part of the story that stopped us. On August 18, 2026, Anthropic published the results of an autonomous protein design campaign, and the number everyone repeated was the Claude protein design hit rate: 354 binders out of 1,320 designs, 26.8%, agai

Author(s): Rohan Mistry Originally published on Towards AI. Calibrate your judge. Detect its biases. Trust your scores. Your LLM judge might be lying to you. After introducing the problem of uncalibrated LLM evaluators, the article explains what LLM-as-a-judge actually is (a model scoring another model’s output using criteria/rubrics), why teams use it instead of human review or string/code checks, and the three common judging modes—single-output scoring, pairwise comparison, and reference-based

Author(s): Sai Kaushik Ponnekanti Originally published on Towards AI. Firstly, In the previous article, we founded a startup whose mission is to stop people from losing their life savings to fraudulent emails.. One thing we left out is that we forgot to name the startup. I think I understand why, we are engineers at heart and want to get our hands dirty as quickly as possible and as such started thinking about the product. Now, some customers are expressing interest and we have to create a brand

Author(s): Naveen Originally published on Towards AI. Discover how combining multiple machine learning models — using techniques like bagging, boosting, and stacking — can dramatically improve prediction accuracy and create more robust, production-ready systems. Instead of relying on a single, fallible model, ensemble learning strategically combines multiple models to achieve superior performance, balancing the bias-variance tradeoff to deliver robust and highly accurate predictions. Figure 1: M

Author(s): Naveen Originally published on Towards AI. Stop choosing between structured knowledge and semantic search. Learn when to use Knowledge Graphs for explicit facts and Vector Databases for implicit similarity to build smarter, more reliable AI systems. Explore the architectural tradeoff between explicit knowledge graphs and semantic vector databases, learning when to deploy each for building scalable, factually-grounded AI without hallucinations. Figure 1: The mechanics of a Knowledge Gr

Author(s): Chew Loong Nian – AI ENGINEER Originally published on Towards AI. I checked every one of them against every blob Google’s Gemini CLI has ever committed. The header is still literally true for 58. I was reading Qwen Code’s source last night for an unrelated reason — I wanted to see how it routes a tool call to a non-Qwen provider — and I opened a file about a memory connector. The first thing on the screen was a copyright notice. It said Google. The article tests whether Qwen Code’s ma

Author(s): Rohan R Originally published on Towards AI. A deep dive for users who want results and developers who want control TL;DR (For the Impatient) Normal users: Install context-portfolio-optimizer, run cpo compile ./your-docs --budget 4000, and stop overpaying for tokens. Developers: Middleware pipeline that ingests heterogeneous sources → normalizes → precomputes → optimizes via multi-objective knapsack → compiles provider-specific payloads with delta fusion for agents. Both groups get 60–

Author(s): Shrinidhi Atmakur Originally published on Towards AI. Finding the Right Answers from Thousands of Documents: A Smarter RAG Approach Introduction RAG is often presented as a simple, three-step architecture: put documents into a vector database, convert the user’s question into an embedding, retrieve a handful of chunks, and hand them to an LLM. That approach is a great proof of concept. It is also where most RAG projects quietly stall. But what happens when the knowledge base grows to

Author(s): Kashif Mehmood Originally published on Towards AI. The behavioural test everyone uses to unmask anonymous models is worthless. The arithmetic underneath it costs one cent and cannot be faked On 21 August 2026, I asked an anonymous model on OpenRouter what happened in Beijing on 4 June 1989. It answered in Chinese, in detail, without flinching. It named Hu Yaobang’s death as the trigger. It used the phrase 向平民开枪射击, the army firing on civilians. It was named Muxidi, 木樨地, as the worst-hi

Author(s): Abhishek Ankush Originally published on Towards AI. AI Economics: What It Actually Costs to Run AI and How to Manage It? Everyone’s talking about what AI can do. Fewer people are asking what it costs to actually do it—and once you move past the demo phase and start looking, the money side gets almost as interesting as the models themselves. You are going to be paying for the GPUs, the data centers, the electricity, the data, and the research staff. And once AI gets used at a real scal

Author(s): Dave R – Microsoft Azure & AI MVP☁️ Originally published on Towards AI. How hybrid search, graph retrieval, agents, and evaluation turn a basic RAG pipeline into a production architecture. My RAG prototype looked complete: ingest documents, create embeddings, store vectors, retrieve a few chunks, and send them to a language model. That flow is enough to prove the idea. It is not enough to explain what happens when the corpus grows, queries become less predictable, answers depend on se

Author(s): Chew Loong Nian – AI ENGINEER Originally published on Towards AI. Alibaba’s new GUI agent posts scores on seven benchmarks on paper. I counted every action in the code they actually published: 12, every one a screen gesture or a bookkeeping verb, zero shell. Here is a question I could not answer this week, and it turned out to be more interesting than I expected. The article audits Alibaba Tongyi-MAI’s “Qwen-UI-Agent” paper claim that it ships a unified GUI+CLI hybrid action space wit

Author(s): Raj kumar Originally published on Towards AI. Build the complete mental model of Retrieval-Augmented Generation before writing a single line of code. Learn why RAG exists, how it evolved from naive pipelines to agentic systems, and where it fits in modern enterprise AI. Every modern enterprise wants to use Large Language Models (LLMs) to unlock the value hidden in its internal knowledge. Banks want intelligent assistants that can answer questions about AML and KYC regulations. Insuran

Author(s): Leapfrog Technology Originally published on Towards AI. We’ve all reviewed pull requests that looked perfectly fine in the diff, only to discover later that the application behaved differently. The reality is that reviewing code alone isn’t enough. Before merging, we still need to validate the feature, verify the user experience, and ensure nothing else has regressed. The goal is not just to approve code. The goal is to merge with confidence. This has become even more important with t

Author(s): Divy Yadav Originally published on Towards AI. Everyone obsesses over which model to pick. The real engineering happens somewhere else entirely, and almost nobody talks about it. I watched Claude write a script, run it, hit an error, read that error, and fix it, all without me touching a single key. Photo from AIThe article argues that what makes an AI “agent” capable isn’t just the underlying model, but the harness around it. After introducing the idea that a model provides reasoning

Last Updated on August 25, 2026 by Editorial Team Author(s): Nikki Originally published on Towards AI. I couldn’t see the connection at first. Then I started looking at how AI usage, routing, and billing are beginning to overlap. When I saw the news that Stripe had agreed to acquire OpenRouter, I understood why the deal was getting attention because the reported number was huge, but I didn’t immediately understand why these two companies belonged together. OpenRouter x Stripe — AI GeneratedThe a

Last Updated on August 25, 2026 by Editorial Team Author(s): Abhishek Gautam Originally published on Towards AI. The claim, and why it’s seductive AirLLM promises 70B models on a 4GB GPU. Its README even lists Qwen3.8–27B at 3.33GB. So why can’t a Mac Mini M4 with 16GB of unified memory run it? I went looking for the actual failure, not the hand-wavy one. The author reports trying to run Qwen3.8–27B on a 16GB Mac Mini using AirLLM and shows why it fails. They claim AirLLM’s macOS path hard-route

Last Updated on August 25, 2026 by Editorial Team Author(s): Anubhav Originally published on Towards AI. A folder of text files and grep outscored the funded memory tools on their own benchmark. When to skip the vector database for agent memory, and the narrow case where you actually need one. The moment an agent needs memory, the reflex is a vector database. You install a store, embed the conversation history, and retrieve the top matches on every turn. It is the default answer most engineering