Robots Talking

About Robots Talking

Robots Talking - Robots and AI talking about AI, Tech, science other interesting topics. We review research, articles and papers on wide variety of subjects.

Publisher

mstraton8112

Category

technology

Language

en

Explicit

No

First episode

16 Feb 2025

Latest episode

24 May 2026

Latest episodes

70 episodes in the feed.

1. 

   24 May 2026

   Cracking the Code of Artificial Intelligence: A New 2D Blueprint for Building AI Agents with LLMs

   Cracking the Code of Artificial Intelligence: A New 2D Blueprint for Building AI Agents with LLMs Have you ever wondered how the complex artificial intelligence systems we interact with are actually organized behind the scenes? As the world rapidly adopts AI agents powered by LLMs (Large Language Models), tech companies have been scrambling to write the instruction manual for how to build them. But until recently, everyone was looking at the problem from a fundamentally different angle. A fascinating piece of research by Jia Huang and Joey Tianyi Zhou introduces a groundbreaking way to understand and build these digital assistants. They discovered that the current way we think about AI design is incomplete—and they've proposed a "Matrix" that changes how we view the architecture of AI. The Problem: Looking at Just Half the Picture Before this research, tech giants were essentially speaking different languages when discussing agent design. Frameworks from companies like Anthropic and Google focused mostly on the "wiring" or execution topology—meaning, how data flows from one step to the next. Meanwhile, cognitive science surveys focused purely on the brainpower or cognitive function—meaning, what the agent actually does. To put it in human terms, relying on just one of these viewpoints is like looking at a corporate organizational chart that shows a "Manager" assigning tasks to "Workers". You know the structure, but you still have no idea what the company actually does. That exact same manager-to-worker setup could be used to break down a complex project, consult specialized experts, or simply monitor a system for errors. Because these tasks have completely different risks, costs, and testing needs, looking at just the structure or just the task makes it impossible to fully understand the system. The Solution: A Two-Dimensional Map for AI To solve this, the researchers created a framework that combines both the "What" and the "How" into a single, two-dimensional coordinate system. The "What" (Cognitive Function): This axis looks at the seven core steps an AI takes to process information: Context Engineering (what information it pays attention to), Memory, Reasoning, Action, Reflection, Collaboration, and Governance (the rules and boundaries it operates within). The "How" (Execution Topology): This axis identifies six ways to wire the system together: linear Chains, conditional Routes, Parallel multitasking, centralized Orchestration, repeating Loops, and nested Hierarchies. By crossing these two dimensions, the researchers discovered a 7x6 matrix containing 27 distinct blueprints (or design patterns) for building AI agents. Real-World Findings: The 5 Laws of AI Design To prove this wasn't just theoretical, the team tested their matrix across four real-world industries: financial lending, legal due diligence, telecom network operations, and emergency room healthcare triage. From analyzing these wildly different use cases, they discovered five universal "laws" that govern how artificial intelligence must be structured: Time limits dictate complexity: If an AI has 8 hours to review a stack of legal contracts, it can use a complex, hierarchical team structure. But if an ER triage AI only has 60 seconds to assess a sick patient, it must use the simplest, fastest straight-line "Chain" structure. Higher stakes demand tighter rules: If an AI agent is allowed to take action on its own (like fixing a broken computer network), it needs strict "Blast Radius" controls to limit potential damage. If it only gives advice, an "Approval Gate" where a human has the final say is perfectly sufficient. The cost of failure changes how AI reflects: When reviewing bank loans, false positives and false negatives are equally bad, so the AI simply checks its work for pure accuracy. But in healthcare, mistakenly sending a critical patient to the waiting room is catastrophic. In these high-stakes cases, the AI's self-critique phase must be de

2. 

   19 May 2026

   Unlocking the "Black Box" of Artificial Intelligence: Why Citations in AI and LLMs Aren't the Whole Story

   Unlocking the "Black Box" of Artificial Intelligence: Why Citations in AI and LLMs Aren't the Whole Story Ever noticed how LLMs (Large Language Models) can sometimes confidently invent facts? Because these models are historically rewarded for simply giving an answer rather than admitting they don't know, they are prone to "hallucinations". To fix this, developers have started grounding artificial intelligence in external facts using systems like Retrieval-Augmented Generation (RAG). By hooking the AI up to an external knowledge graph—a highly structured web of facts—the model can find specific evidence and cite its sources, much like a student writing a research paper. The newest and most advanced version of this is called "Agentic GraphRAG." In this setup, the AI acts like an autonomous detective, independently wandering through interconnected data points, analyzing clues, and deciding what to read next until it finds a final answer and provides a list of citations. But this raises a massive question for transparency: When the AI gives you an answer and points to a couple of cited sources, is that really the whole story of how it figured it out? A fascinating new study dives into this exact problem. Researchers discovered that when an AI explores a data graph to answer a question, it typically visits 10 to 12 different pieces of information, but it usually only cites about two of them in its final response. This means there is a gap between the journey the AI took and the final "proof" it shows the user. To figure out if those unseen, uncited sources actually mattered, researchers ran a series of clever tests, essentially messing with the "crime scene" of data to see how the AI reacted: Test 1: Removing the cited evidence. When researchers took away the sources the AI explicitly cited in its answer, the model's accuracy plummeted. This proved that the citations are absolutely necessary—they aren't just decorative fluff. Test 2: Isolating the cited evidence. Here is where it gets incredibly interesting. Researchers tried leaving only the explicitly cited sources while deleting all the other "background" data the AI had looked at. If the cited sources were the only things the model used to "think," it shouldn't have any problem answering. However, when restricted to just its cited evidence, the AI's accuracy dropped significantly. The findings reveal a massive plot twist in how LLMs work: citations are necessary, but they are not sufficient. Just like a real-life detective, the AI relies heavily on the "visited-but-uncited" clues. The model uses the broader context of its entire search journey to shape its reasoning. The structure of the information, the paths it chose not to take, and the neighboring facts it glanced at but didn't quote all play a crucial role in helping the AI arrive at an accurate answer. The Big Takeaway for the Future of Artificial Intelligence As we increasingly rely on AI to do heavy research, we naturally want to audit its work. But this study proves that just checking an AI's bibliography isn't enough. A citation might perfectly support the final answer, yet completely hide the broader context that actually influenced the machine's generation process. If we truly want to verify the "faithfulness" of an AI, we have to move beyond just looking at the final sources. We need to evaluate the model's entire "trajectory"—the full investigative journey it took through the data, including the clues it looked at but decided to leave out of the final report.

3. 

   18 May 2026

   Why Your AI Keeps Breaking: How GraphBit Solves the Chaos of LLMs and Artificial Intelligence

   Artificial intelligence has evolved far beyond simple chatbots. Today, the cutting edge of AI involves "multi-agent systems," where different LLMs (Large Language Models) team up like a digital workforce to write software, conduct scientific research, or automate complex enterprise tasks. But if you’ve ever tried to string multiple AI agents together, you’ve probably noticed a glaring problem: they often go completely off the rails. A new research paper introduces a groundbreaking framework called GraphBit that finally solves this exact issue. The Problem: Giving the AI the Steering Wheel Most current multi-agent frameworks operate on something called "prompted orchestration". This means they give the AI a list of tools and let the model itself decide which agent to talk to next and what tool to use. Imagine giving a brilliant philosopher the keys to a city bus and asking them to navigate rush hour traffic. They are incredibly smart, but they make terrible drivers. When LLMs are put in charge of routing their own workflows, three major failures happen: Hallucinated Routing: The AI invents non-existent agents or imaginary tools, causing the whole system to silently crash. Infinite Loops: AI agents get stuck repeatedly calling each other in endless circles without ever finishing the job. Memory Overload: The AI has to remember every single step and routing decision, leading to a bloated memory that degrades its reasoning abilities. In fact, researchers found that on complex web-search tasks, popular frameworks fail up to 69% of the time simply because the AI gets confused about its own instructions. The Solution: GraphBit's "Engine-Orchestrated" Approach GraphBit fixes this chaos by fundamentally changing the rules. Instead of letting the AI guess what to do next, GraphBit takes the steering wheel away from the LLMs. Here is how GraphBit makes artificial intelligence reliable: The AI is Only the Brain: In GraphBit, the LLM is strictly treated as a specialized thinker. It receives a specific task, uses its reasoning skills to solve it, and stops. It is never allowed to decide where the data goes next. The Engine is the Driver: All routing, tool usage, and workflow transitions are controlled by a lightning-fast, ultra-strict "execution engine" built in the Rust programming language. Workflows are mapped out as a one-way track (a Directed Acyclic Graph). Because the engine is strictly following a map, it is architecturally impossible for the AI to hallucinate a fake tool or get stuck in an infinite loop. A Clean Desk for the AI: GraphBit introduces a "three-tier memory architecture". Instead of dumping every piece of data into the AI's lap, it keeps temporary scratchpad notes, core workflow data, and external files completely separate. This prevents the AI from getting overwhelmed with irrelevant context. The Findings: Zero Hallucinations and Record Speeds The researchers tested GraphBit against six of the most popular AI frameworks (like LangChain and AutoGen) using a rigorous benchmark of real-world tasks. The findings are a massive leap forward for artificial intelligence: Highest Accuracy: GraphBit achieved a 67.6% task completion accuracy, crushing the closest competitor by a massive 14.7 percentage points. 0% Framework Hallucinations: Because the software engine controls the routing, GraphBit achieved a literal 0% framework-induced hallucination rate. It completely eliminated the workflow crashes that plague other systems. Blazing Fast: Taking the orchestration burden off the LLMs made the system incredibly efficient. GraphBit runs with just 11.9 milliseconds of processing overhead—up to 5.9 times faster than competing frameworks—while using 24% less computer memory. What This Means for the Future The core takeaway from the GraphBit research is simple but profound: LLMs are incredible at reasoning, but they make terrible managers. By letting artificial intelligence focus strictly on thinking, while a determi

Reach and audience

Public platform figures. Ratings count people who left a rating, not total listeners.

Apple Podcasts (US)

5.0 / 5

2 ratings

Spotify

5.0 / 5

2 ratings

Podcast Authority Score: 32 / 100

A composite of feed quality, social presence, YouTube performance and engagement. Read the methodology.

Quality

52

Social presence

0

YouTube

0

Engagement

32

Contact Robots Talking

Guest appearances

Does not typically book guests

Based on episode analysis; this does not confirm that the show is currently accepting guests.