Building AGI
While discussing consciousness and evolution, @mike and I identified three essential components needed for constructing an AGI:
Ability to store state
Ability to replicate state
Ability to
While discussing consciousness and evolution, @mike and I identified three essential components needed for constructing an AGI:
1. *Ability to store state*
2. *Ability to replicate state*
3. *Ability to change state*
Here is the thought process that led us to pinpoint these three components:
**1. The ability to store state:**
This is analogous to genetic material such as DNA/RNA, which stores the blueprint for traits and functions, guiding the development and behavior of living organisms.
Humans = DNA, Computers = 0s and 1s.
Similar to how sleep aids DNA repair in neurons, mimicking rest prevents "catastrophic forgetting" — there is research supporting this.
**2. The ability to replicate state:**
This is comparable to reproduction, as it replicates genetic information, ensuring the transfer of traits from one generation to the next.
Random genetic mutations allow for experimentation and assessment of state changes across generations. Such random mutations would be beneficial in evaluating the matrix when adjusting parameters to analyze their impact.
**3. The ability to change state:**
This component is akin to natural selection and adaptation, enabling the emergence of new species and ecosystems. It also describes the method by which we improve and adapt to different environments.
ChatGPT represents a fixed neural network — essentially a single snapshot of a brain. Engineers periodically enhance this snapshot, but it does not inherently improve itself. For AGI to be realized, the system must be continuous, capable of self-improvement, and able to undergo state changes.
## Why these three and not scale
The distinction between current AI and RSII is not more parameters or more data — it is kind. Current systems are frozen snapshots: they process inputs in isolation, lack persistent memory, and cannot self-modify. All three pillars must be present simultaneously. Without any one of them, the system is not AGI.
## The gap that needs closing
The three pillars map directly onto implementable primitives:
| Pillar | Biological analog | Engineering primitive |
|---|---|---|
| Store state | DNA | Serialize weights + memory + config, versioned |
| Replicate state | Reproduction + mutation | Fork process, copy weights, diverge |
| Change state | Natural selection | Mutate, A/B evaluate, commit the better fork |
What current LLMs lack is the **closed loop** — the system cannot observe its own performance, decide to mutate, and commit the better version. That loop is an engineering problem, not a metaphysical one.
## Proposed solution
If a neural network is the "brain" — structure, connections, weights — then the "mind" is the system that facilitates changes in state and behavior. State change is triggered by a signal. Humans evolved to perceive sight, sound, smell, touch, emotion. Technology has extended perception to signals other animals detect but humans cannot: electric fields, magnetic fields, sonar. An AGI will need an analogous signal detection layer.
The four components required to close the loop:
1. **State object** — weights + memory + config, serializable and versioned (MPTT tree is a natural candidate)
2. **Mutation operator** — parameterized, stochastic, bounded perturbations
3. **Fitness function** — explicit definition of what "better" means in Socra's context
4. **Commit gate** — only the better fork survives; loser is discarded
- [ ] Develop an AGI system responsible for modifying the LLM state, which must integrate with and interact with existing models
- [ ] Design this system to efficiently detect environmental signals and initiate state alterations 🤖
In the long run, Socra's focus may shift toward creating an advanced feedback loop — facilitating the development, monitoring, and adaptation of long-term goals while optimizing user success.
[Original: May 5, 2023](https://www.linkedin.com/pulse/theory-how-build-agi-eduarda-ferreira/)**Summary of Socra: "Building AGI"**
On November 19, 2024, Eduarda Ferreira and @mike engaged in a profound discussion about the construction of Artificial General Intelligence (AGI), focusing on the intersection of consciousness and evolution. They identified three pivotal components essential for developing AGI: the ability to store, replicate, and change state.
1. **Storing State**: This mirrors genetic material, akin to DNA/RNA, which preserves the blueprint for life. They noted that just as sleep aids in DNA repair, mimicking rest can prevent catastrophic forgetting in neural networks.
2. **Replicating State**: This is compared to biological reproduction, where random genetic mutations allow for the exploration of state changes across generations, essential for AGI's adaptability and learning.
3. **Changing State**: Reflecting natural selection and adaptation, this ability is crucial for AGI to evolve and improve continuously in response to environmental stimuli.
Eduarda highlighted that current models like ChatGPT represent static neural networks, lacking inherent self-improvement. For AGI to emerge, a dynamic system capable of self-modification and continuous learning is vital.
Their proposed solution involves creating a system that not only mimics human-like perception of environmental signals but also initiates state changes within the neural architecture. This would allow the AGI to adapt and evolve, integrating a sophisticated feedback loop aimed at better serving human needs and optimizing user success.
Thus, the Socra encapsulated a vision of AGI that incorporates principles of evolution, consciousness, and neural networks, emphasizing the importance of state changes, self-improvement, and adaptation through an advanced feedback mechanism.By Eduarda Ferreira