Useful General Intelligence — Danielle Perszyk, Amazon AGI

Danielle, a cognitive scientist at the Amazon AGI SF Lab, proposes a shift in the goal of artificial intelligence from creating thinking machines to augmenting human intelligence. This vision, termed “Useful General Intelligence” (UGI), prioritizes making humans smarter and giving them more agency, rather than simply making AI more autonomous.

Perception as Controlled Hallucination

The talk begins by drawing a parallel between AI “hallucinations” and the nature of human perception. Our brains operate as “prediction machines,” constantly reconciling internal models with sensory input. In this sense, all perception is a form of controlled hallucination. This is a feature, not a bug, of a flexible intelligence. The challenge with AI agents is not to eliminate hallucinations, but to control them, just as our brains do.

Two Paths for AI: Augmentation vs. Autonomy

Historically, AI has been driven by two competing visions:

1. Thinking Machines (Standard AGI): The goal, originating in the 1950s, is to build AI that is as smart as or smarter than humans. This path focuses on increasing the AI’s agency.
2. Augmenting Humans (Douglas Engelbart’s Vision): The goal is to use computers to make humans smarter. This path, championed by the inventor of the computer mouse, focuses on increasing human agency.

The speaker argues for the second path. The co-evolution of humans and technology shows that while automation can be a powerful tool for augmentation (freeing up our conscious attention), it can also reduce our agency, leading to issues like social media addiction and echo chambers. The UGI vision is to “unhobble humans” by giving us precise control over AI tools.

Nova Act: An Agent for the Visual Web

To realize the UGI vision, we must meet developers and models where they are today. The primary obstacle is that most of the digital world consists of visual user interfaces, not structured APIs. Nova Act is Amazon’s solution to this problem.

• Core Technology: Nova Act is an agentic model (a specialized version of Amazon’s foundation model, Nova) combined with an SDK. It is trained specifically to be highly reliable at interacting with UIs.
• Functionality: It translates natural language instructions into actions on the screen, such as clicking, typing, and scrolling. The primary developer interface is a simple act() call.
• Developer Experience: The SDK allows for powerful integrations with Python. The demo showed how to:
  • Use a natural language prompt to find an apartment on a website.
  • Define a pydantic class to perform structured data extraction from the webpage.
  • Use a Python helper function with a ThreadPool to spin up multiple browser instances and perform parallel searches (e.g., calculating biking distances via Google Maps for each result).
  • Load the final, enriched data into a pandas DataFrame for analysis.

The Evolutionary Path to Reliable Agents

Current agents are often LLM wrappers that lack a true understanding of their environment. Computer use agents like Nova Act are different because they are grounded in a shared environment—the UI—giving them a form of embodiment.

However, true reliability requires more than just accurate clicks; it requires understanding a user’s higher-level goals. To achieve this, the talk draws a deep analogy from human cognitive evolution.

In a very real sense, as represented by the steady evolution of our augmentation means, the development of artificial intelligence has been going on for centuries.

Human general intelligence is not a property of an individual brain but an emergent phenomenon of social interaction. It arose from a feedback loop where our ancestors learned to use each other’s minds as tools, a process called “representational alignment.”

The key cognitive technology that enabled this was language, which co-evolved with our “models of mind.” Language is unique because it presupposes an understanding that words refer to concepts that other minds create. For agents to become truly general and aligned, they will eventually need to develop their own models of our minds.

Building Useful General Intelligence Together

We cannot build these “models of mind” directly. Instead, we must create the preconditions for them to emerge. This requires:

1. A Shared Environment: A model of the UI that both humans and agents can interact with.
2. Intuitive Interfaces: Tools that support natural human-agent interaction.
3. A Common Language: A way to communicate goals and actions effectively.

This creates a virtuous cycle:

• Useful products encourage human-agent interaction.
• This interaction generates valuable data.
• The data is used to advance the models.
• Better models lead to more useful products and, ultimately, smarter and more capable humans.

This collaborative, co-evolutionary process is how we can collectively build Useful General Intelligence.

• Tags:
• AI Agents
• Feature
• Human-Computer Interaction
• Cognitive Science
• Useful General Intelligence
• Nova Act
• Amazon AGI