DEV Community: Can Mingir

Neuro-Symbolic AI: From Research to Industry

Can Mingir — Wed, 10 Dec 2025 17:11:52 +0000

📘 This week:
Neuro-Symbolic AI brings together the statistical nature of machine learning with the formal reasoning capabilities of symbolic AI. It seeks to offer a balanced approach to contemporary AI technologies, by combining the ability to learn from data, with the capacity to reason upon knowledge acquired from an environment. The main criticism of neural machine learning lies in its lack of explainability and semantics, which are key requirements in safety-critical applications, yet inherent strengths of logic-based methods. Recently, several corporations have publicly announced products and technologies grounded in Neuro-Symbolic AI methodologies. This talk provided a concise review of the foundations, frameworks and tools underlying Neuro-Symbolic AI, along with illustrative applications. It concludes by highlighting current trends and research directions in the field.

📝 Suggested paper:
A Garcez, LC Lamb:
Neurosymbolic AI: the 3rd wave. Artif. Intell. Review 56(11):12387-12406 (2023)

https://link.springer.com/article/10.1007/s10462-023-10448-w

Training Large Reasoning Models

Can Mingir — Wed, 03 Dec 2025 16:22:49 +0000

📘 This week, Dr. Asim Munawar from IBM’s Watson Research will examine how integrating Neuro-Symbolic AI approaches with large language models (LLMs) can enhance their reasoning, interpretability, and reliability.

Dr. Asim Munawar is a Project Lead at IBM’s Watson Research Center in New York, where he heads efforts to enhance reasoning, planning, and agentic workflows in enterprise-scale large language models. With over 15 years of experience in AI – more than a decade of it at IBM Research – he has held key leadership roles, including Manager and Program Director for Neuro-Symbolic AI https://asimmunawar.github.io

This presentation will explore architectures that combine symbolic logic, structured representations, and deep learning, and outlines strategies to address limitations such as brittleness, hallucination, and lack of transparency in current LLMs.

📝 Bonus: The DeepSeek paper will be covered in Asim's presentation.
https://arxiv.org/abs/2501.12948

Can/Will LLMs Learn to Reason?

Can Mingir — Wed, 12 Nov 2025 16:34:14 +0000

Join us for an interactive session exploring Neuro-Symbolic AI. This week, Abulhair Saparov from Purdue University is presenting “Can/Will LLMs Learn to Reason?”

This week:

Reasoning—the process of drawing conclusions from prior knowledge—is a hallmark of intelligence. Large language models, and more recently, large reasoning models have demonstrated impressive results on many reasoning-intensive benchmarks. Careful studies over the past few years have revealed that LLMs may exhibit some reasoning behavior, and larger models tend to do better on reasoning tasks. However, even the largest current models still struggle on various kinds of reasoning problems. In this talk, we will try to address the question: Are the observed reasoning limitations of LLMs fundamental in nature? Or will they be resolved by further increasing the size and data of these models, or by better techniques for training them? I will describe recent work to tackle this question from several different angles.

📢 New Series: Neuro-Symbolic Wednesday

Can Mingir — Wed, 05 Nov 2025 08:22:27 +0000

Join us for an interactive session exploring Neuro-Symbolic AI, the emerging paradigm that blends the strengths of neural networks with symbolic reasoning. We will discuss how hybrid approaches can enhance generalization, interpretability, and reasoning, and how these methods are shaping the future of intelligent systems.

Whether you’re a researcher, engineer, or simply curious about the cutting edge of AI, you’ll find an engaging space to learn, connect, and exchange ideas.

Event Format:
🔍 Reading Group
🗣️ Guest Spotlights
💬 Research Roundtable
🛠️ Open-Source Review
🤝 Collaboration Hour

RSVP at Luma

📢 Neuro-Symbolic AI Summer School 2025 | Online Event | Aug 14 - 15

Can Mingir — Wed, 13 Aug 2025 18:43:42 +0000

http://lu.ma/pqzv80yd?utm_source=devto

📅 Agenda

Day 1: Frameworks and Foundations | AUG 14

📜 Big Picture
9:00–9:30 AM ET – Alexander Gray, Centaur AI Institute – Neuro-Symbolic AI and NSSS4 (Overview)
9:30–9:40 AM – Q&A / Break
9:40–10:20 AM – Arvind Narayanan, Princeton – Assessing the Current State of AI (Talk)
10:20–10:30 AM – Q&A / Break

🏗 Frameworks
10:30–11:05 AM – Leilani Gilpin, UC Santa Cruz – Neuro-Symbolic AI for Safer Autonomous Vehicles (Tutorial)
11:05–11:15 AM – Q&A / Break
11:15–12:00 PM – Mateo Zarlenga & Pietro Barbiero, Cambridge – Foundations of Interpretable Models (Tutorial)
12:00–12:10 PM – Q&A / Break
12:10–12:55 PM – Parisa Kordjamshidi, Michigan State – Compositional Learning in Language and Vision (Tutorial)
12:55–1:05 PM – Q&A / Break

📐 Mathematical Foundations
1:05–1:45 PM – Parikshit Ram, IBM Research – How to Measure Compositionality, and Why it Leads to Better Generalization (Talk)
1:45–1:55 PM – Q&A / Break
1:55–2:35 PM – Peihao Wang, UT Austin – Why Neural Networks Can Discover Symbolic Structures (Talk)
2:35–2:45 PM – Q&A / Break
2:45–3:25 PM – Changlong Wu, Purdue – Why Current Models Will Always Hallucinate (and a Path Forward) (Talk)
3:25–3:35 PM – Q&A / Break
3:35–4:35 PM – Sridhar Mahadevan, Adobe Research & UMass Amherst – Category Theory: The Mathematics of Symbolic Structures (Tutorial)
4:35–4:45 PM – Q&A / Break

Day 2: Methods and Systems | AUG 15

🤖 Neuro-Symbolic AI Software
9:00–9:40 AM – Bowen Li, Carnegie Mellon – Generalizing to New Situations in Robotics (Talk)
9:40–9:50 AM – Q&A / Break
9:50–10:30 AM – Olga Vileskaia & Kevin O'Connor, Centaur AI Institute – Explainability While Retaining Predictive Accuracy (Talk)
10:30–10:40 AM – Q&A / Break

🧩 Learning Symbolic Models
10:40–11:20 AM – Bin Yu, UC Berkeley – Interpretable and Veridical Data Science (Talk)
11:20–11:30 AM – Q&A / Break
11:30–12:10 PM – Felix Petersen, Stanford – Deep Differentiable Logic Gate Networks (Talk)
12:10–12:20 PM – Q&A / Break
12:20–1:00 PM – Hikaru Shindo, TU Darmstadt – Neuro-symbolic Agentic Systems (Talk)
1:00–1:10 PM – Q&A / Break

🛡 Safer AI Systems
1:10–1:50 PM – Soroush Saghafian, Harvard – Human + AI "Centaur" Systems (Talk)
1:50–2:00 PM – Q&A / Break
2:00–2:40 PM – Pranava Madhyastha, City Univ London – New Results in Controllable Text Generation (Talk)
2:40–2:50 PM – Q&A / Break

➗ AI Systems for Mathematics
2:50–3:30 PM – Shange Tang, Princeton – State-of-the-art Performance in Automated Mathematical Theorem Proving (Talk)
3:30–3:40 PM – Q&A / Break
3:40–4:20 PM – Ankit Anand, DeepMind – Curious Case of AI in Maths: Being Proficient in Advancing Open Conjectures in Maths Yet Having Struggles in AI for Education (Talk)
4:20–4:30 PM – Q&A / Break

🔮 Looking Forward
4:30–5:30 PM – Panel on The Future of AI – Rich Sutton, Univ Alberta; Leonardo de Moura, Amazon; Artur Garcez, City Univ London; more TBA; moderator: Alexander Gray – Discussion including open Q&A
5:30–5:35 PM – Alexander Gray, Centaur AI Institute – What's Coming Next and How to Participate (Closing remarks)

[Boost]

Can Mingir — Thu, 31 Jul 2025 14:45:06 +0000

Can Mingir for GreyCollar

Jul 31 '25

How to Use AI Agents in Supply Chain & Logistics

#ai #machinelearning

Comments

2 min read

How to Use AI Agents in Supply Chain & Logistics

Can Mingir — Thu, 31 Jul 2025 14:44:39 +0000

Before anything else, what are AI Agents?

AI Agents are autonomous versions of LLMs used for decision-making, and the main reason is they are highly effective at complex tasks by breaking them down into smaller, manageable steps. In addition, they can perform interactive actions such as retrieving data from databases, making API calls, or even reaching out to humans for additional information.

AI Agents are often compared with LLMs, but their usage in enterprise systems is very different. Today, microservices are considered best practice and can be written in multiple languages like Java or Python. Traditional coding is great at highly complex algorithm and deterministic requirements, but AI Agents can be replaced some of those microservices if particular module requires large number of business logic, especially those business rule often changed.

AI Agents can replace some of your microservices, especially where business rules change frequently.

This is especially true in Supply Chain & Logistics, where processes like inventory management, order fulfillment, and route planning rely on countless business rules. These rules often vary from one customer to another and change frequently due to seasonal demand, market fluctuations, or regulatory updates. AI Agents are particularly useful in these scenarios because they can adapt quickly to evolving requirements, reducing the need for constant manual intervention or system reconfiguration.

Which case applies to you?

When deciding between traditional coding and AI Agents for a microservice, the choice largely depends on the nature of your business logic and system requirements. Traditional coding works best when dealing with a small number of complex, deterministic rules. However, it comes with higher maintenance costs, as every change requires manual updates, testing, and deployment. In contrast, AI Agents excel when there are numerous basic rules that frequently change. They offer dynamic adaptability, learning from feedback and new data to adjust logic without heavy developer intervention. Their goal-based approach reduces maintenance costs and allows for error handling through supervised learning, making them ideal for fast-evolving environments like Supply Chain & Logistics.

The relentless growth of e-commerce has placed an unsustainable strain on traditional, manual methods of product catalog management. The imperative to provide rich, accurate, and engaging product information at scale has created a critical need for intelligent automation solutions that can navigate the complex trade-offs between creative quality, factual integrity, operational speed, and cost.
— Ryan M., Cloud Principal Architect, Retail at Google

Conclusion

As Supply Chains & Logistics grow more dynamic and complex, relying solely on traditional coding for every business rule update can slow you down. AI Agents offer a powerful alternative—bringing flexibility, adaptability, and intelligence to processes that demand constant change. Whether your goal is to boost automation, lower maintenance costs, or build resilience, now is the perfect time to explore where AI Agents can fit into your architecture.

P.S. coding is still here, not going anywhere 😎

📢 GreyCollar: Supervised Agentic AI Project

Can Mingir — Wed, 25 Jun 2025 16:28:06 +0000

We’re launching an open-source, supervised AI agent platform built for Human–AI collaboration.

TL;DR

🎯 Supervised Learning
As issues arise, data is labeled under human supervision and added to the agent’s knowledge base for continuous learning.

🛡️ Hallucination Control (Human-in-the-Loop)
Agents only respond when sufficient knowledge exists. If not, tasks are escalated to human supervisors.

⚡ Event-Driven Agentic Platform
Inspired by DDD, GreyCollar uses a platform layer to orchestrate tasks through decentralized, choreographed events.

🔗 GitHub: github.com/GreyCollar/GreyCollar

What are Supervised AI agents?

GreyCollar AI is a supervised AI agent platform for human–AI collaboration. The platform provides an environment to continuously learn from human supervisors, so they can adapt to real-world workloads.

Each AI colleague works within defined responsibilities and uses a knowledge base to complete tasks. When uncertain, they escalate to human supervisors—enabling "Hallucination Control" to prevent mistakes.

Human-AI Collabs (Human-in-the-Loop)

Human-in-the-Loop (HITL) is a collaborative approach where AI agents work alongside human experts to enhance decision-making, automate processes, and refine task execution. In this model, human supervision plays a key role in guiding, correcting, and improving AI-driven workflows.

Improved Accuracy – Human feedback enables AI colleagues to refine responses in real time, reducing errors and increasing reliability.
Continuous Learning – AI adapts to new tasks and domains by integrating ongoing human input, improving with every interaction.
Safe & Responsible AI – Human oversight ensures ethical alignment, reduces bias, and mitigates unintended or harmful outputs.
Operational Efficiency – AI handles routine, repetitive work at scale, freeing human supervisors to focus on higher-value, strategic decisions.

⚡n8n Integration

GreyCollar can be part of your favorite flow tools like n8n, enabling you to embed supervised AI directly into automated workflows.

Features

Colleague (AI): AI assistants that handle tasks based on assigned responsibilities and knowledge.
Supervising (Human): Humans who guide AI with feedback, questions, or extra info.
Knowledge: The info AI uses—documents, FAQs, or other sources.
Responsibility and Task: Defines what tasks the AI performs and how.
Team: A group of AI colleagues for managing knowledge and leadership.
Communication: How you interact with AI—via chat, email, Slack, WhatsApp, etc.
Integration: Connects to third-party tools via Model Context Protocol (MCP).

Colleague (AI)

Colleagues are AI assistants that help you with your tasks based on responsibilities and knowledge. They are designed to:

Complete tasks standalone for given responsibilities
Continuously learn and persist to knowledge base
Collaborate with other human supervisors or human colleagues

Supervising (Human)

Supervising by human is raised when the AI is not able to complete the task or needs human input. The supervisor can provide feedback, ask questions, or give additional information to help the AI complete the task.

⚠️ This is the core concept to eliminate hallucination that the AI evaluates knowledge existed before the execution of the task.

Knowledge

Knowledge is the information that the AI uses when working on responsibilities. It can be in the form of documents, FAQs, or any other.

Knowledge can be added manually or part of the supervising process during task execution.

Responsibility and Task

Responsibility is a blueprint of the tasks that the AI will perform based on knowledge. It defines what the AI can do and how it can help you.

Tasks are the actions that the AI performs for a given responsibility with knowledge. Once the task is initiated through communication, the AI will execute the task and provide feedback to the supervisor.

Team

Team is a logical grouping of AI colleagues. Mainly this grouping provides 2 major benefits:

Knowledge Management: Knowledge can be shared between AI colleagues within the team, while each colleague can also maintain their own individual knowledge. In agentic AI, effective knowledge management is crucial to eliminate hallucinations, ensuring that each AI colleague has sufficient knowledge to complete tasks without being misled by irrelevant or unnecessary information. a
Team Lead: The team lead is the person responsible for handing off the task to the AI colleagues.

Communication

Communication is the primary way to interact with AI colleagues. It can occur through various channels, such as chat, email, or voice, depending on the context and user preferences. These communication channels are linked to specific responsibilities that AI colleagues are capable of handling, ensuring interactions are efficient and task-relevant. Multiple channels can be used simultaneously, allowing for flexibility in how users engage with AI colleagues.

In short, communication opens up AI colleagues to the outside world, enabling them to perform tasks.

Integration

All integrations are based on MCP that allows you to connect to any third-party service. The integration can be used for bidirectional communication:

Incoming: Pulling data such as reading from Google Drive or checking Google Calendar
Outgoing: Sending data such as writing to Google Drive or posting to a Slack channel

Event-Driven Agentic AI Platform

GreyCollar is an Event-Driven AI Agent Platform designed for dynamic and adaptive AI workflows and autonomous decision-making. While frameworks like LangChain and LlamaIndex are specialized in creating static flows, but it is significantly more challenging to have flexible AI agent compared to event-drive architecture.

Key Advantages:

⚡ Dynamic Workflows:

Instead of a rigid sequence of actions, GreyCollar agents react to events in real-time. These events could be anything: a new email, a sensor reading, a user interaction, or even the output of another AI agent.
This allows for highly adaptable and context-aware behavior. The agent's next action is determined by the current situation, not a pre-programmed path.

🧠 Autonomous Decision-Making:

Agents can make decisions without constant human intervention. They can monitor their environment, identify relevant events, and take appropriate actions based on predefined rules or learned behaviors.
This is crucial for applications that require rapid response times or operate in dynamic environments.

🔄 Modularity and Scalability:

Event-driven systems are naturally modular. Agents can be designed as independent components that communicate with each other through events.
This makes it easier to build complex systems by combining smaller, specialized agents. It also allows for easier scaling, as new agents can be added without disrupting the existing system.

🔍 Real-time responsiveness:

Because the system is based on events, it can react very quickly to changes in the enviroment. This is very important for applications that need to be up to date.

Hello World

Customer: "Do you have a parking spot at your store?"
> SESSION.USER_MESSAGED
{
  sessionId: "2116847c",
  content: "Do you ... at your store?"
}

AI: "Please wait a moment while working on the answer."
> SUPERVISING.RAISED
{
  sessionId: "2116847c",
  question: "Do you ... at your store?"
}

Supervisor: "Yes, we have a parking spot in the back of the store."
> SUPERVISING.ANSWERED
{
  sessionId: "2116847c",
  question: "Yes, we have ... of the store."
}

# Knowledge is stored for future reference. 🧠

AI: "Yes, we have a parking spot in the back of the store."

> SESSION.USER_MESSAGED
{
  sessionId: "2116847c",
  question: "Yes, we have ... of the store."
}

# A Few Moments Later... 🍍

Customer #2: "Planning to come down there, how is parking situation?"

> SESSION.USER_MESSAGED
{
  sessionId: "3746a52b",
  content: "Planning ... situation?"
}

AI: "Yes, most certainly, we have a parking spot in the back. 😎"
> SESSION.USER_MESSAGED
{
  sessionId: "3746a52b",
  question: "Yes, most ... in the back."
}

🚀 Join us on GitHub
Thanks to supervised learning, we’re taking a fresh approach to AI agents. Join us in shaping the future of human–AI collabs — we welcome all kinds of contributions!

🔗 GitHub: github.com/GreyCollar/GreyCollar

📢 Supervised Agentic AI Project

Can Mingir — Sat, 21 Jun 2025 11:29:26 +0000

We’re launching an open-source, supervised AI agent platform built for Human–AI collaboration.

Targeting:

🎯 Supervised Learning
As issues arise, data is labeled under human supervision and added to the agent’s knowledge base for continuous learning.

🛡️ Hallucination Control (Human-in-the-Loop)
Agents only respond when sufficient knowledge exists. If not, tasks are escalated to human supervisors.

⚡ Event-Driven Agentic Platform
Inspired by DDD, GreyCollar uses a platform layer to orchestrate tasks through decentralized, choreographed events.

🔗 GitHub: github.com/GreyCollar/GreyCollar

GreyCollar: Supervised AI Agent | Human-AI Collabs

Can Mingir — Wed, 26 Feb 2025 16:44:08 +0000

Hello y'all, we are launching a project for Human-AI collabs with supervised learning capabilities. You're more than welcome to jump in and brainstorm with us!

What is Supervised AI Agent?

GreyCollar is a Supervised AI Agent that functions under human guidance and feedback, operating within a supervised learning framework. It is trained on labeled data, where each input is paired with a corresponding correct output, enabling the model to learn from explicit examples and improve decision-making accuracy.

Designed for human-AI collaboration (Human-in-the-Loop), GreyCollar is highly effective in scenarios that require data-driven decision-making, automation, and real-time adaptability. It incorporates human-in-the-loop mechanisms, enabling iterative learning through continuous feedback and model adjustments. This enhances its ability to handle complex tasks at work environment.

Autonomous Workflow

Autonomous workflows are self-driven processes where AI agents can independently execute multi-step tasks with human supervision. These workflows integrate task planning, execution, decision-making, and learning based on changing inputs or goals.

Task Decomposition
- The AI agent breaks down complex goals into smaller, executable steps.
- Uses methods like hierarchical planning or task graphs.
Decision-Making & Adaptation
- Dynamically adjusts workflows based on new information.
- Uses supervised learning to adapt itself to workspace-related tasks and directions.
Memory & Context Awareness
- Agents retain context across workflow steps.
- Utilizes vector databases, episodic memory, or long-term storage.
Multi-Agent Coordination
- Multiple AI agents collaborate by delegating and verifying tasks.
- Uses shared knowledge bases to improve coordination and efficiency.
Human-in-the-Loop & Supervised Learning
- Uses human feedback to improve models through supervised learning.
- Helps refine edge cases and prevents unintended consequences.

Human-AI Collabs (Human-in-the-Loop)

Benefits

Enhanced Accuracy – Human feedback helps the AI refine its responses and correct errors in real-time.
Adaptive Learning – AI models improve continuously by integrating human insights, ensuring adaptability to evolving tasks.
Safe AI – Human oversight prevents biases, ensures fairness, and mitigates unintended consequences.
Task Optimization – AI streamlines repetitive processes while humans focus on strategic and complex decision-making.

Welcome! I’ve been expecting you—"Skynet was gone. And now one road has become many." 🌐

The future is building up! AI Agents are now an emerging field within AI communities and marks a crucial milestone on the journey to AGI. Just like any other tooling in computer science, we must be mindful of when and where to use them. LangChain, or LlamaIndex might be a better fit if your application has a static flow—where AI doesn't need to make dynamic decisions—like in ChatBots, RAG etc. That said, if your business rules are well-defined and deterministic, there’s no shame in coding them directly!

However, if you need continuous learning combined with autonomous decision-making—in other words, true AI Agent—GreyCollar may suit you well. Choosing the right tool for the job is key.

Can Mingir
@canmingir

Get Started

git clone https://github.com/greycollar/greycollar.git
npm install

// Add env variables
JWT_SECRET=<JWT_SECRET>
LLM=OPENAI
OAUTH_CLIENT_SECRET=<OAUTH_CLIENT_SECRET>
OPENAI_API_KEY=<OPENAI_API_KEY>

npm start

This will start three applications simultaneously: Dashboard, API Server and Proxy Server. Once started, the dashboard should be accessible in your browser at http://localhost:3000

Learn more at https://greycollar.ai/docs

Features

Teams and Colleagues

In GreyCollar, Colleagues aka AI agents are organized into structured teams based on their areas of expertise and operational roles. This hierarchical framework ensures that AI agents work efficiently within the organization, providing clarity and structure to their contributions.

One of the biggest challenges in AI agents swarms in the workplace is preventing knowledge management. GreyCollar’s structured hierarchy provides visibility, clarity, and organizational coherence, allowing AI agents to seamlessly adapt to day-to-day tasks with minimal human supervision.

Async Chat (Email, Slack, WhatsApp...)

GreyCollar is designed as a standalone AI addition to operational workspaces, where efficient communication is essential for daily task execution. Unlike traditional chatbots, GreyCollar’s AI agents require ongoing, context-aware communication to function effectively within dynamic team environments.

The Async Chat feature enables AI agents to engage in continuous, asynchronous communication—both with human team members and other AI colleagues. This ensures that agents can participate actively in task execution, collaborate with teammates, and respond to evolving instructions over time.

Task Execution

Task flows in GreyCollar is automatically generated by the AI, allowing it to dynamically adjust its actions based on real-time data and predefined objectives. The agent continuously monitors inputs, applies decision-making algorithms, and initiates actions to fulfill its goals efficiently. This autonomous process enables the agent to handle complex tasks, adapt to changing conditions, and optimize performance across various applications, such as automated customer support, predictive maintenance, or autonomous driving systems.

https://github.com/user-attachments/assets/5111dfb1-687e-4fe6-a6bc-9c5e1a11abd4

Team Chat

Team Chat is an internal communication tool designed for interaction between supervisors and AI Colleagues. It enables human supervisors to engage directly with AI agents for a variety of purposes—whether it’s assigning tasks, asking questions, providing new information, or offering real-time feedback.

This feature transforms AI agents from passive tools into active collaborators, creating a dynamic environment where human expertise and AI capabilities work together effortlessly. Team Chat is also fully integrated with Slack, allowing users to communicate with AI agents within familiar workflows without the need for additional tools or platforms.

AI Marketplace

You can use any LLM, or even bring your own—we support and welcome them all. 🚀

Event-Driven AI Agent Platform

Key Advantages:

⚡ Dynamic Workflows:

Instead of a rigid sequence of actions, GreyCollar agents react to events in real-time. These events could be anything: a new email, a sensor reading, a user interaction, or even the output of another AI agent.
This allows for highly adaptable and context-aware behavior. The agent's next action is determined by the current situation, not a pre-programmed path.

🧠 Autonomous Decision-Making:

Agents can make decisions without constant human intervention. They can monitor their environment, identify relevant events, and take appropriate actions based on predefined rules or learned behaviors.
This is crucial for applications that require rapid response times or operate in dynamic environments.

🔄 Modularity and Scalability:

Event-driven systems are naturally modular. Agents can be designed as independent components that communicate with each other through events.
This makes it easier to build complex systems by combining smaller, specialized agents. It also allows for easier scaling, as new agents can be added without disrupting the existing system.

🔍 Real-time responsiveness:

Because the system is based on events, it can react very quickly to changes in the enviroment. This is very important for applications that need to be up to date.

Hello World

Customer: "Do you have a parking spot at your store?"
> SESSION.USER_MESSAGED
{
  sessionId: "2116847c",
  content: "Do you ... at your store?"
}

AI: "Please wait a moment while working on the answer."
> SUPERVISING.RAISED
{
  sessionId: "2116847c",
  question: "Do you ... at your store?"
}

Supervisor: "Yes, we have a parking spot in the back of the store."
> SUPERVISING.ANSWERED
{
  sessionId: "2116847c",
  question: "Yes, we have ... of the store."
}

# Knowledge is stored for future reference. 🧠

AI: "Yes, we have a parking spot in the back of the store."

> SESSION.USER_MESSAGED
{
  sessionId: "2116847c",
  question: "Yes, we have ... of the store."
}

# A Few Moments Later... 🍍

Customer #2: "Planning to come down there, how is parking situation?"

> SESSION.USER_MESSAGED
{
  sessionId: "3746a52b",
  content: "Planning ... situation?"
}

AI: "Yes, most certainly, we have a parking spot in the back. 😎"
> SESSION.USER_MESSAGED
{
  sessionId: "3746a52b",
  question: "Yes, most ... in the back."
}

⭐️ Star us on GitHub for the support

Thanks to supervising learning, we have a brand-new approach to AI Agents. Join us in shaping the future of AI! We welcome all kinds of contributions!

Join us at
https://github.com/greycollar/greycollar

greycollar / greycollar

Supervised AI Agent 🌎 Human-AI Collabs | Autonomous Workflow | Human-in-the-Loop | Async Communication | Open-ended Task Execution

GreyCollar

Supervised AI Agent

What is Supervised AI Agent?

Autonomous Workflow

Task Decomposition
- The AI agent breaks down complex goals into smaller, executable steps.
- Uses methods like hierarchical planning…

View on GitHub

📰 Neuro-Symbolic AI Newsletter | December 2024

Can Mingir — Mon, 30 Dec 2024 10:20:53 +0000

🔗 Learning from Symbolic Knowledge for Neural Networks
Published in: Neuro-Symbolic Artificial Intelligence. Publisher: Springer Nature Singapore. Log in. Introducing the latest innovation: AI-assisted ...

🔗 Does current AI represent a dead end?
What are your thoughts on neuro-symbolic integration ... neural networks with the reasoning and knowledge representation of symbolic AI) ?

🔗 Meet LLMSA: A Compositional Neuro-Symbolic Approach for Compilation
Home Tech News AI Paper Summary Meet LLMSA: A Compositional Neuro-Symbolic Approach for Compilation-Free, Customizable Static Analysis with...

🔗 The 80/20 problem of generative AI — a UX research insight
Image Depicting the Evolution of SWOT Analysis using AI and Neuro-symbolic AI created by. Towards AI. In. Towards AI. by. Mukundan Sankar · The Secret ...

🔗 Breakthroughs in LLM Reasoning Show a Path Forward for Neuro-symbolic Legal AI
Our finding opens up many directions in the application of neuro-symbolic AI to legal problems, which we feel uniquely positioned to pursue with ...

🔗 AllegroGraph Named a 2025 Trend-Setting Product
“Neuro-Symbolic AI represents the next evolution of artificial intelligence, where the integration of symbolic reasoning with machine learning ...

🔗 The Rise of Reasoning Models
The integration of neuro-symbolic AI with traditional deep learning has emerged as a promising direction, allowing systems to both learn from data ...

🔗 How Neurosymbolic AI Can Fix Generative AI's Reliability Issues
This is where neurosymbolic AI can help. By combining the power of neural networks with the logic of symbolic AI, it could solve some of the ...

🔗 Best practices for AI-based security for 2025
Neuro-symbolic AI ... The neuro-symbolic AI approach combines (1) statistical methods of machine learning with (2) non-statistical reasoning techniques ...

🔗 Assured Neuro Symbolic Learning and Reasoning (ANSR)
DARPA is motivating new thinking and approaches to artificial intelligence development to enable high levels of trust in autonomous systems through ...

🔗 Unlocking the future of intelligent Health and AI: key learnings from the Gartner IT Symposium 2024
Looking ahead, Gartner identified key trends for 2025, including neurosymbolic AI, multi-agent systems, and the evolution of decision intelligence.

🔗 When Chess Champion Gukesh Dommaraju Met Demis Hassabis
Google DeepMind is blending scaling with architectural innovation, betting on multimodal and neuro-symbolic AI to propel it towards AGI. Google ...

🔗 Deepak Kaul awarded a 2024 Global Recognition Award for AI and cybersecurity contributions
On the more academic side, his published research on dynamic upsell systems and neuro-symbolic AI has contributed to understanding real-time decision- ...

🔗 Neuro-Symbolic AI in Healthcare: Unlocking Precision Medicine
Just as chemists combine knowledge of fundamental elements with pattern recognition to understand complex chemical systems, neuro-symbolic AI aims to ...

🔗 New AI Speaks Two Languages at Once and Just Might Crack AGI
What is Neuro-Symbolic AI? Neuro-Symbolic AI combines the pattern-recognition capabilities of neural networks (subsymbolic AI) with the logical ...

🔗 Generative AI can't shake its reliability problem. Some say 'neurosymbolic AI' is the answer
Neurosymbolic AI could be a best-of-both-worlds marriage between deep learning and "good old-fashioned AI."

🔗 Neurosymbolic AI Could Be the Key to Unlocking Human-Like Intelligence
Neurosymbolic approaches augment neural networks with rule-based logic to better approximate human reasoning. The technology could help AI overcome ...

🔗 Amazon's Hybrid AI Safeguarding Approach Spurs Rules-Checking Prompts That Catch ...
AI, also commonly referred to as neuro-symbolic AI. It goes like this. Generative AI and LLMs are principally based on pattern-matching across ...

🔗 Wayne Chang's Reasoner claims big AI reliability breakthroughs
The serial entrepreneur says Reasoner's neurosymbolic approach avoids the risks of using generative AI.

ARC, Neuro-Symbolic AI, Intermediate Language | Road to AGI | Recap 01

Can Mingir — Thu, 28 Nov 2024 17:12:06 +0000

Hello everyone! 👋

Over the past few months, we've been working on ARC benchmark as a part of our Neuro-Symbolic AI project that we’ve been able to achieve some promising results, and it feels incredible to see our approach—combining symbolic reasoning with neural network capabilities—making meaningful progress.

Nucleoid (aka nuc) adopts a Neuro-Symbolic AI architecture but introduces a novel twist: the intermediate language as a universal bridge between Neural Networks and Symbolic Systems.

The intermediate language plays a critical role in uniting the two paradigms. Based on our findings, the nuc lang helps Neural Networks is to abstract patterns, which is eventually used in Symbolic System, and Knowledge Graph is built with logic and data representations in the intermediate language. In addition, LLMs surprisingly behaves near deterministic while running on ARC-AGI.

Before diving into our approach:

What is ARC Benchmark?

The Abstraction and Reasoning Corpus (ARC) is a benchmark dataset and challenge designed to test AGI systems on their ability to perform human-like reasoning and abstraction. Developed by François Chollet, ARC is not a typical machine learning dataset—it intentionally avoids tasks solvable by brute-force statistical techniques or large-scale data training.

Our Progress 🐋

We were able to get very promising and exciting numbers (Still incomplete tho). For example, in this puzzle, our project responded with this result without any prompt engineering.

More details here 👇
https://github.com/NucleoidAI/Nucleoid/tree/main/arc

...and this is ChatGPT o-1's answer

🌱 What is Neuro-Symbolic AI?

Neuro-Symbolic AI combines the pattern-recognition capabilities of neural networks (subsymbolic AI) with the logical reasoning and structured knowledge of symbolic AI to create robust and versatile systems. Neural networks excel at learning from unstructured data, like images or text, while symbolic AI handles explicit rules and reasoning, offering transparency and precision. By integrating these approaches, Neuro-Symbolic AI enables generalization from smaller datasets, improves explainability, and supports tasks requiring both adaptability and logical consistency. This hybrid approach is pivotal for advancing AGI, as it bridges the gap between learning from data and reasoning through.

🌍 System 1 and System 2

Neuro-Symbolic AI aligns intriguingly with the concepts from Daniel Kahneman’s Thinking, Fast and Slow, which describes two systems of human thought: System 1 (fast, intuitive, and automatic) and System 2 (slow, deliberate, and logical).

Neural Networks in Neuro-Symbolic AI parallel System 1, as they excel at processing unstructured data, recognizing patterns, and generating outputs rapidly without explicit reasoning. They mimic intuitive, subconscious processes that are data-driven and reactive.
Symbolic AI, on the other hand, mirrors System 2, as it relies on explicit rules, logic, and structured reasoning to solve problems in a deliberate and explainable manner, akin to conscious, rational thought.

By combining these two paradigms, Neuro-Symbolic AI reflects the dual systems of human cognition, enabling it to tackle problems requiring both fast intuition (pattern recognition) and slow reasoning (logic and planning). This hybrid approach not only enhances AI's adaptability but also brings it closer to human-like intelligence by integrating the strengths of both modes of thought.

🦆 Duck Test

"If it looks like a duck, swims like a duck, and quacks like a duck, then it probably is a duck"

Simply, it is System 1 at work. We have seen ducks probably thousands or millions of times throughout our lives, which forms well-defined patterns in our cognition. When we come across something resembling a duck, human cognition doesn’t trigger System 2 because identifying the object as a duck is automatic and intuitive. Our brains rely on a mental "duck schema" even the individual parts of a duck, such as its wings, bill, or webbed feet, are matched to their associated labels in realm of System 1.

Duck in the lake

Again, we won't be surprised if we seen duck in the lake, because we have enough labelled patterns to make the call.

Duck in the City

In this case, as our experiences typically associate ducks with natural settings, it is uncommon to see a duck in an urban environment at night. This unfamiliarity triggers System 2 to take over, engaging in more deliberate reasoning. So, System 2 is now responsible for reasoning as cities being dangerous after dark, and the duck is in the city, duck may not be safe. System 2 overrides System 1’s instinctive identification of "a duck" and shifts the focus to evaluating the broader circumstances of the duck's well-being.

CPU vs GPU

Neuro-Symbolic AI architecture orchestrates a seamless harmony between the precision of CPUs in reasoning and the raw power of GPUs in pattern recognition.

CPUs are specialized in logical branch operations used in decision-making, and GPUs execute parallel arithmetic operations for matrix multiplications in pattern recognition. So, it is important to understand how they are different.

CPUs are designed for versatility, with fewer cores optimized for high-performance single-threaded tasks and low latency, making them well-suited for complex decision-making, sequential processing, and multitasking. In contrast, GPUs have thousands of smaller, energy-efficient cores optimized for massive parallelism, enabling them to handle tasks like matrix computations, image rendering, and deep learning efficiently. GPUs excel in throughput-oriented tasks where large data sets can be processed simultaneously, while CPUs focus on general-purpose computing and running the operating system. Additionally, CPUs often feature larger caches and more sophisticated control logic to handle diverse workloads, whereas GPUs prioritize raw computational power and bandwidth to accelerate specific workloads. This fundamental difference makes GPUs indispensable for tasks requiring high parallelism, while CPUs remain the backbone of general computing and coordination.

While building a modular AI system, it is crucial to design with the underlying hardware in mind. In advanced systems like AIs, hardware constraints can sometimes conflict with algorithmic requirements. In Neuro-Symbolic standpoint, Neural Network (System 1) needs GPU for pattern recognition and CPUs for knowledge representation and reasoning in Symbolic (System 2). It is worth noting, System 1 and 2 is just for building AI foundation, expecting more and more systems like them...

In short, it is not possible doing efficient reasoning with GPU or pattern recognition with using CPU.

🌿 Stay Tuned for Recap 02 in Road_to_AGI series...

NucleoidAI / Nucleoid

Neuro-Symbolic AI with Knowledge Graph | "True Reasoning" through data and logic 🌿🌱🐋🌍

Nucleoid

Neuro-Symbolic AI with Knowledge Graph
Reasoning Engine

Declarative (Logic) Runtime Environment: Extensible Data and Logic Representation

Nucleoid is a declarative, logic-based, contextual runtime for Neuro-Symbolic AI. Nucleoid runtime tracks each statement in IPL-inspired declarative syntax and dynamically creates relationships between both logic and data statements in the knowledge graph to used in decision-making and problem-solving process.

Adaptive Reasoning: Combines symbolic logic with contextual information to analyze relationships, draw conclusions and incorporating new information and adjusting its conclusions accordingly.
Logic Graph: Specialized knowledge graph that captures relationships between both logic and data statements based on formal logic, facilitating complex deductions and adapting to new information.
Explainability: The Logic Graph provides a transparent representation of the reasoning process, making it easier to understand how decisions are reached and potential biases are identified.

Echoing to the idea of "thinking, fast and slow", AI system should provide fast, “intuitive” ideas, and the…

View on GitHub