The AI community has long dreamed of the day that AGI (Artificial General Intelligence) becomes a reality. It was debated whether AGI was even possible, but the last 2-3 years has shown a startling increase in just how fast AI can evolve. The rise of LLMs like ChatGPT have demonstrated that AI can be much more than a single-problem algorithm. Giant models now exist, and can be used by anyone, that can cover nearly any topic and provide deep insights in almost no time. Looking at where we are today seems like a natural progression, but even three years ago this seemed almost like science fiction.
Taking that same lens forward, the “impossible” seems more and more likely. And one of the biggest milestones on this journey is the development of true AGI.
A recent paper by Dr. Ben Goertzel suggests an architecture that doesn’t take us all the way to AGI, but it takes a very large step in that direction. The paper provides case studies that support the architecture’s intent, and seem to indicate that it can accomplish different levels of insight and intelligence. Moreover, this architecture may just move us past some of the performance plateaus that have famously restricted LLMs from improving beyond a certain point.
The biggest point of concern for AGI is that of ownership. Whoever controls AGI first may just have a frightening advantage over the rest of the market, along with the economy, society, and even government. AGI may evolve independently in a number of different places simultaneously, but given the power wielded by just a handful of AI giants today, the thought of a few megacorporations controlling AGI is deeply disturbing. Dr. Gorertzel’s paper has some interesting ideas on that as well, and it might offer some hope for us after all. Let’s take a look at the ideas suggested, consider their feasibility, and contemplate what this might mean for our global society in the very near future.
On the shoulders of giants: OpenClaw and MeTTaClaw
The paper discusses an orchestration layer that builds from a number of existing AI tools. The foundation of this is OpenClaw, which is an open-source AI agent that can run autonomously and was designed to be a highly effective workflow manager. The tool can connect to messaging apps, calendars, emails, or other types of workflows. It gained popularity because it was effective, because it was free (and open source), and because it could run locally–solving many privacy issues that encumber other AI agents. Building on OpenClaw is MeTTaClaw, a compact agentic system that was coded in MeTTa (SingularityNET’s AGI programming language). The creator, Hyperon team member Patrick Hammer, was able to develop a very compact agent, with a core of around 200 lines. This makes it lightweight, but also very easy to audit, verify, and adapt.
The goal of MeTTaClaW was to build from OpenClaw, focusing on three shortcomings: building a space for long-term memory; adding the ability to use a wider set of tools such as web searches, file access, and even shell commands; and creating the ability to acquire new skills. Key to this suite of improvements is the MeTTa programming language, and its ability to self-rewrite code. This statement immediately raises the pulses of traditional programmers, and greatly increases those opportunities for exponential gains–or substantial losses.
HyperClaw, the cognitive orchestration layer
The subject of Dr. Goertzel’s paper, HyperClaw, was built from the foundations of MeTTaClaw. The underlying problem seen in other AI agentic platforms is the need for the human operator not just in creative decisions, but in more basic mechanical processes. An example used was a quantitative finance research workflow that built up approximately 100 iterations of experiments working to build up an efficient and effective result. Using traditional AI tools, this entire effort still uses around 30 hours of the human user’s time, but much of this is spent performing essential (but non-value added) tasks such as fixing bugs, parameter sweeps, and reworking cycles. The real value-adds for the human are decisions such as pivoting on the larger strategies, or catching LLM hallucinations with cross-checking (or using their own expertise for sanity checks).
In summary, HyperClaw extends the reach of MeTTaClaw’s capabilities by adding three key architectural elements. First, the layer adds Context Frames. These are structures that capture the state of the multi-module task. This includes the stated goals, current hypotheses, the experiment logs/results, and other evaluation information. A critical element to a Context Frames is that it becomes the authoritative source of truth instead of relying on an LLM. This allows the larger model to break through current LLM performance limitations.
Second, the Attention-Metaprotocol is designed to manage the MeTTa rewrite function mentioned above. However, instead of focusing on discrete tasks, the structure operates at two timescales: a faster loop examines those operations happening in the seconds-to-minutes range, rewriting accordingly in order to improve performance. The other loop runs on a longer timescale, and is capable to examining and modifying entire strategic postures that the overall model is running.
Third, the Module Spaces wrap up the various external systems, allowing the HyperClaw orchestrator to interact with them in a consistent manner, leveraging the Hyperon interface. This creates a universal interface with the system, but will be critical as these external systems continue to evolve: the Spaces can adapt to new technology, but the HyperClaw fundamentals can remain the same.
The results
It’s important to note that HyperClaw is a preliminary design, and the goal of the paper is to spark discussion, debate, and generate improvements. That said, the initial case studying involving the financial workflow is projected to take the original 30 hours of human effort and reduce it down to 5 hours; most importantly, nearly all the time remaining is time specifically meant for the human operator, handling flags for complex issues, deciding on strategic options, and thinking outside the box for unconventional approaches.
That said, one of the best features of HyperClaw isn’t what it can do, but where it resides. Built as a decentralized platform for SingularityNET’s forthcoming decentralized AI infrastructure ASI:Chain, HyperClaw won’t be controlled by an AI monolith corporation. It will be shared by a global community, and distributed in a way that can’t be censored, threatened by competition, or hijacked by a government. This is an area where we see the dark side of AI today, and control of AI means power, and power means control over people. The decentralized AI movement as a whole was born to stand against this abuse of power, and HyperClaw will be another critical tool in that fight.
Concerns, hopes, and what to expect next
The ability to self-rewrite itself to modify and improve an agent’s architecture is uncharted territory at this scale. This immediately gives an uncomfortable chill to anyone who has seen a dystopian science fiction movie. However, this evolution is not only likely, it is necessary to improve our ability to use AI as a whole. Moreover, this is the same chill we collectively had as AI began to evolve and perform without explicit programming instructions. We’ve since discovered that AI can perform disastrously if we don’t design it correctly, and we are constantly learning from our mistakes.
With tools like HyperClaw, the stakes are higher as it has even more autonomy, along with the ability to rewrite itself. That said, tackling this step in the road to AGI should be done as soon as possible so we can fail small, fail often, improve significantly before AGI is embedded into our everyday lives. Even an initial look at the tool sparks curiosity and interest, as it begins to pave a clearer road to AGI. We don’t yet have clarity on how long it will take to get to AGI, but this is one of the first architectures that has been designed in a way to handle more and more AGI-driven tools as part of the workflow, all without having to tear out the code and start over.
Most of all though, there is hope for AGI not as a tool controlled by a few, but a revolutionary technology shared by all. Decentralized AI is a cornerstone of this effort, and architectures like HyperClaw give a clear idea what is possible in the near future. It’s up to us decide who will leverage that power.
