In a world obsessed with large language models (LLMs) and “AI that can do everything,” Ben Goertzel offers a much-needed reality check.

In his recent talk at The Beneficial AGI Summit & Unconference 2025 (BGI-25), the 🔗SingularityNET founder and long-time 🔗Artificial General Intelligence (AGI) researcher makes a bold but grounded claim:

Today’s AI is powerful, but it isn’t AGI, and we’re going to need new architectures, new tools, and new infrastructure to get there.

Even more importantly, he argues that how we build and deploy AGI will determine whether we end up in a world of concentrated power and instability, or one where intelligence and opportunity are broadly shared.

That’s exactly the space 🔗DEEP is focused on: helping builders, researchers, and communities experiment with the next wave of AGI-native ideas, things like Hyperon, MeTTa, neural-symbolic AI, and decentralized AI infrastructure.

Let’s unpack some of the key concepts from Ben’s talk and why they matter for the future of AGI and for what we’re building together at DEEP.

Beyond LLMs: Why “Very Smart Autocomplete” Isn’t Enough

Ben acknowledges what everyone can see:

• Current AI tools can solve graduate-level math problems.
• They outperform most humans in games like chess and Go.
• They can generate essays, resumes, cover letters, code, and artwork.

By many visible metrics, it looks like we are already close to human-level intelligence.

But there’s a gap.

LLMs are trained on an enormous amount of data and are astonishingly good at pattern-matching and pattern-remixing. They can imitate almost anything they have “seen” in their training distribution.

What they notably struggle with is:

• True creativity grounded in their own goals and experience.
• Reliability and self-correction at the level humans manage instinctively.
• Generalization beyond their training data in a robust, open-ended way.

That last point is crucial.

You can’t build a civilization-changing intelligence that only rearranges what already exists. AGI must be able to push beyond its training data, the way humans invent new science, new music, new culture.

Which brings us to what Ben and our collaborators are actually building.

Artificial General Intelligence (AGI)

AGI is not just “stronger AI.” It has a specific meaning in Ben’s framing:

AGI is an intelligence that can generalize across domains and beyond its training data, and continually expand its own capabilities in open-ended ways.

Some important aspects:

• Breadth: It can handle a wide range of tasks, from math and coding to planning, creativity, and social interaction.

• Depth: It can understand, not just replicate. This makes it capable of building new concepts that weren’t directly present in its training set.

• Self-extension: It can revise its own strategies, tools, and even underlying representations to achieve new goals.

LLMs are broad in scope because they’re trained on almost everything online, but they still primarily obey their training distribution. AGI is about crossing that boundary.

At DEEP, this is the level we care about, not just “another wrapper on top of an LLM,” but research, tools, and systems that move us closer to true generality.

Hyperon, A Framework for Real AGI

To move beyond the LLM paradigm, Ben and his team are developing 🔗Hyperon, a next-generation AGI framework.

You can think of Hyperon as:

A decentralized knowledge metagraph + AI operating system designed to host many kinds of intelligence at once.

Some of its defining features:

• Knowledge Metagraph
  A huge, evolving graph of nodes and links representing concepts, data, rules, and experiences. This metagraph can span many machines and locations.

• Multi-Paradigm Support
  Hyperon is designed to integrate:
  • Deep neural networks and LLMs
  • Logical reasoning and symbolic inference
  • Evolutionary algorithms and creative search
  • Concept blending and other generative mechanisms

• Scalability & Decentralization
  It’s built to run at scale, billions of graph nodes and to run across networks, not just inside a single corporate server farm.

Instead of relying on one gigantic neural network to do everything, Hyperon gives us a shared substrate where different forms of intelligence can interact and cooperate.

At DEEP

We see frameworks like Hyperon as the kind of AGI-native infrastructure builders should start playing with now. Our aim is to:

• Host experiments on multi-paradigm architectures
• Connect Hyperon-like systems with real-world data, agents, and communities
• Support developers who want to build tools and applications on top of this new layer

MeTTa: A Language for Thinking Machines

Hyperon needs a way to express knowledge, rules, and processes. That’s where 🔗MeTTa comes in.

MeTTa (Meta-Metagraph Language) is the AGI programming language designed for Hyperon.

In practice, it is:

A language of thought for AGI, a way to encode data, logic, probabilities, and procedures in a unified form, and run AI methods over them efficiently.

MeTTa allows developers to:

• Represent knowledge as flexible graph structures
• Define transformation rules, inference patterns, and probabilistic reasoning
• Plug in neural networks, symbolic engines, and evolutionary search over the same underlying data

Recent progress (which Ben highlights) includes:

• High-performance MeTTa compilers that can handle hundreds of millions to billions of nodes in RAM
• The ability to run reasoning, neural inference, and evolutionary algorithms together at large scale

This is not just another general-purpose programming language. It’s purpose-built for building minds, not just apps.

At DEEP

We want to make environments where you can:

• Learn MeTTa and similar AGI-oriented languages

• Prototype cognitive architectures that leverage them
• Connect them with decentralized infra, governance, and real-world use cases

Neural-Symbolic AI

One of the most important themes in Ben’s vision is neural-symbolic AI.

Instead of choosing between “neural” vs. “symbolic,” the idea is to combine both:

• Neural components (like LLMs, vision models, audio models) excel at:
  • Perception
  • Pattern recognition
  • Natural language understanding

• Symbolic components (logic, rules, explicit reasoning) excel at:
  • Transparent decision-making
  • Structured planning
    Checking consistency and truthfulness

In Hyperon, this looks like:

• Neural models operating within the knowledge metagraph

• Logical inference engines that can:
  • Ground their conclusions in observed facts
  • Distinguish coherent reasoning from hallucination

• Creative components (like evolutionary algorithms and concept blending) that explore new solutions

This hybrid approach matters because:

Most real-world problems, governance, science, policy, operations, all require both intuition and explicit reasoning, not just text prediction.

At DEEP

We see neural-symbolic AI as a crucial opportunity area:

• For building governance agents, research copilots, and reasoning tools

• For integrating AI into complex, high-stakes domains where trust, traceability, and explainability matter

• For bridging the gap between cutting-edge research and deployable systems

We aim to be a home for experiments that treat neural-symbolic AI as the default, not an afterthought.

Decentralized AI & the Road to the Singularity

Ben doesn’t just talk about how AGI will think, but also where it will live and who will control it.

He contrasts two futures:

• Centralized AGI
  • Owned and run by a few big tech companies or governments

  • Operated in massive proprietary data centers

  • Prone to extreme wealth concentration and geopolitical instability

• Decentralized AGI
  • Running across many nodes, organizations, and communities

  • Built on open protocols and infrastructure

  • Designed so that benefits and control are more widely distributed

In his simulations, one clear conclusion emerges:

The most powerful lever for steering away from dystopian outcomes is infrastructure, the “rails” that spread access, capability, and value widely.

That includes, for example:

• Universal Basic Income (UBI)-like mechanisms to reduce extreme inequality

• Open AI stacks that aren’t locked inside a few corporations

• New chains like 🔗ASI chain, designed specifically for scalable AGI and ASI 🔗(Artificial Superintelligence) workloads

Dr. Goertzel points out that massive wealth inequality and centralized control of AGI are much more concrete risks than sci-fi “🔗paperclip maximizers.” And infrastructure is how we meaningfully reduce those risks.

At DEEP

We position ourselves precisely at this intersection:

• A hub for building decentralized AI infrastructure, from protocols to governance layers.

• A space for testing AGI-aligned economic and social “rails”, including reward mechanisms and support systems for people affected by automation.

• A bridge between theoretical work (like Hyperon/MeTTa and ASI chain) and practical deployments in communities, DAOs, labs, and organizations.

If AGI is going to be open, beneficial, and global, it needs ecosystems like DEEP to incubate the ideas, tools, and norms that make that possible.

Why This Matters Now

Ben closes with an uncomfortable but honest point:

• Human-level AGI could arrive within a few years.

• The transition from AGI to superintelligence could be significantly faster than many expect.

• Once AGI can understand and modify its own code, hardware stack, and environment, progress can compound quickly.

That means:

• We don’t have decades to lazily experiment and figure out governance later.

• The infrastructure, architectures, and communities we build now will strongly affect the shape of the singularity.

• It will not be enough for “a few dev teams in a few companies” to decide how the future of intelligence works.

We need:

• Open frameworks like Hyperon and MeTTa

• Hybrid architectures like neural-symbolic AI

• Decentralized infrastructure like ASI chain and related systems

• Networks of people (researchers, builders, policymakers, communities) who are actively thinking about beneficial AGI, not just profitable AI

DEEP exists to help convene and empower exactly those people.

Join the Conversation, Help Build the Future

If any of this resonates, if you care about:

• How we move beyond today’s LLM hype into genuine AGI,

• How frameworks like Hyperon, MeTTa, and neural-symbolic AI will reshape the field,

• How decentralized AI infrastructure can keep the singularity from becoming a story of extreme inequality and instability,

then your next steps are simple:

1. Watch Ben Goertzel’s Full Talk

This post is just a guided overview. To really feel the nuance, the math, the vision, the urgency, you should hear it in Ben’s own words.

🎥 Go watch the full video to get the complete picture of how he sees the path from LLMs to AGI to ASI.

2. Stay Connected With DEEP

If you want to keep following these ideas as they evolve, and see how to build with them, make sure you stay in the loop:

• 🔗Subscribe to our blog for deep dives into Hyperon, MeTTa, neural-symbolic AI, and decentralized infrastructure.

• Follow us on LinkedIn for updates on talks, builder programs, and opportunities to get involved.

Get started: 🔗Learn MeTTa 

We’re just at the beginning of this journey. AGI is not only something that will happen to us, it’s something we can co-create. DEEP is here for the people who are ready to build that future together.