PIANO

Link : https://arxiv.org/pdf/2411.00114

---

(Parallel Information Aggregation via Neural Orchestration) architecture

Notes

• Using PIANO, author shows’ that agents form their own professional identities, obey collective rules, transmit cultural information and exert religious influence, and use sophisticated infrastructures, such as legal systems.
• Just as a pianist coordinates multiple notes to create a harmony, the PIANO architecture selectively and concurrently executes various modules in parallel to enable agents to interact with the environment in real-time.
• 

Why is it hard to build AI civilizations

• Single agents don’t make progress: LLM-powered agents often struggle to maintain a grounded sense of reality in their actions and reasoning. Even a small rate of hallucinations can poison downstream agent behavior when agents continuously interact with the environment via LM calls
• Groups of agent’s don’t make progress: Agents that miscommunicate their thoughts and intents can mislead other agents, causing them to propagate further hallucinations and loop.
• A lack of benchmarks for civilizational progress: There is a lack of large-scale benchmarks that can attributed to how technically difficult it is to perform simulations of hundreds or thousands of agents in a single world.

PIANO Architecture

• Author propose two brain-inspired design principles for the composite architecture of human-like AI agents.
• There are two main problems with agent architecture
  • Concurrency
  • Coherence

• Concurrency

  • P: Agent should be able to think and act concurrently. Agent be able to interact in real time with low-latency, but also have the capacity to slowly deliberate and plan.
  • CS: Vast majority of LLM-based agents today primarily use single-threaded, sequential functions. None are designed for concurrent programming.
  • S: The brain solves this problem by running different modules concurrently and at different time scales.
    • Author designed modules, such as
      • cognition,
      • planning,
      • motor execution,
      • and speech
    • to run concurrently in agent brain.
    • Each module can be seen as a stateless function that reads and writes to a shared Agent State.

• Coherence

  • P: An immediate challenge with concurrent modules is that they can produce independent outputs, making the agent incoherent. For instance, agents say one thing but actually do something else.
    • Incoherence also scales exponentially as the number of independent output modules increases
  • CS: The incoherence problem is usually not obvious for sequential architectures.
    • But a significant problem when multiple output modules can interface with the environment.
  • S: To ensure that the multiple outputs produced by agent are coherent, author introduced a Cognitive Controller (CC) module that is solely responsible for making high-level deliberate decisions.

---

Cognitive Controller and Core modules

• Cognitive Controller is solely responsible for making high-level deliberate decisions.
  • The decisions are then translated downstream to produce appropriate output in each motor module
• The Cognitive Controller synthesizes information across the Agent State through a bottleneck.
• This design of a bottlenecked decision-maker that broadcasts its outputs has been suggested as a core ingredient for human consciousness and is used in some neural network architectures.

• Core Modules

  • Memory: Stores and retrieves conversations, actions, and observations across various timescales.
  • Action Awareness: Allows agents to assess their own state and performance, enabling for moment-by-moment adjustments.
  • Goal Generation: Facilitates the creation of new objectives based on the agent’s experiences and environmental interactions.
  • Social Awareness: Enables agents to interpret and respond to social cues from other agents, supporting cooperation and communication.
  • Talking: Interprets and generates speech.
  • Skill Execution: Performs specific skills or actions within the environment.

---

Experiments

• There is bunch of really fun and expected outcomes, i suggest reading through the paper in detail to enjoy it completely. Link above.