AI Engineer World's Fair 2025

Useful General Intelligence

Loading slides…

Read the talk

Useful General Intelligence

Reliable agents need more than accurate clicks: they need shared environments, controllable actions, and interactions that help people think and act more effectively.

From a talk by Danielle Perszyk

Before you start: Basic familiarity with browser automation and Python will help with the apartment-search example; no neuroscience background is required.

What keeps perception reliable?

Your brain cannot inspect reality directly. It predicts what is happening using a world model, receives sensory information, and reconciles the difference. Danielle Perszyk, a cognitive scientist at Amazon’s experimental AGI SF lab, opens with this account of perception as controlled hallucination. The critical word is controlled: predictions become useful because incoming evidence can correct them.

Three brain-and-globe diagrams beneath the heading “Perception is controlled hallucination,” labeled Prediction, Sensory input, and Prediction error.Open full source frame
Perception as controlled hallucination: prediction, sensory input, and prediction error.

Communication adds another layer. As you understand someone’s words, that person influences your interpretation of the world. Even the word hallucination brings several meanings into play, including the familiar problem of chatbots producing unsupported content. Those systems already help with brainstorming, writing, code, and images, but those abilities do not yet amount to reliable general-purpose thinking, learning, and action.

In Perszyk’s framing, the ability to go beyond observed data is part of flexible intelligence; the challenge is controlling that ability. The analogy does not require building a replica of the human brain or replacing people. The goal is AI that complements human intelligence. Her stronger thesis is that treating general intelligence as something contained inside an isolated thinking machine is a category error.

That broad question has a very practical entry point: agents still struggle to click, type, and scroll reliably. Perszyk approaches the problem through the lab’s vision, the then-current research preview of Nova Act, and the human-agent interactions needed for its further development.

0:160:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:16 · section reference included

Thinking machines and thinking humans

The ambition to build thinking machines did not originate in the 2010s. Perszyk traces it to the founding of AI in 1956: engineers and mathematicians set out to solve intelligence, and although they did not complete that project, they helped initiate a technological feedback loop. More powerful computers became connected through the Internet; those connections enabled more sophisticated learning algorithms; those algorithms made computers more capable. The contemporary AGI ambition continues that trajectory by seeking smarter machines with more agency.

Douglas Engelbart offers a different starting point: improve the capabilities of the person using the computer. He invented the mouse and pioneered interactive computing; modern graphical interfaces developed through multiple projects, including Xerox Alto. The important shift is from a machine that thinks to a system that helps humans think. As people offload computation to devices, cognition becomes distributed across the digital environment.

Perszyk calls this techno-social co-evolution: people invent technologies, and those technologies reshape how people think and work. That leaves two distinct targets for more advanced AI:

TargetWhat improves?
AI as smart as, or smarter than, humansThe machine’s capabilities
AI that makes humans smarterPeople’s capabilities with the machine

Both can involve general-purpose agents. Their usefulness depends on what those agents enable people to do.

3:444:00
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:44 · section reference included

When automation increases agency

A useful tool can simplify life by taking work off your hands, or expand what you can accomplish. Automation can support both. Learning a skill provides a familiar example: initially, its details demand conscious attention; with practice, more of the execution becomes automatic, freeing attention for other problems.

But automated systems can also consume attention or narrow thought. Endless scrolling absorbs time, recommendation systems can reinforce echo chambers, and autocomplete can terminate a line of thinking before it develops. Automation does not automatically produce augmentation. Precise control and the ability to tailor a system are what let people turn automation toward their own purposes.

This changes the objective from giving AI ever more freedom to giving its users more agency. Perszyk describes that goal as building AI that “unhobbles humans.” The immediate engineering requirements are correspondingly modest: work within models’ present capabilities, and make it easy for builders to begin composing useful applications.

5:596:12
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:59 · section reference included

An apartment search becomes a browser workflow

Nova Act begins with the possibility that an agent call could become the atomic unit of digital interaction. The obstacle is access: Perszyk argues that most websites lack APIs, while their visual interfaces already expose the operations people need. The browser therefore becomes the agent’s tool. A specialized version of Amazon Nova, trained for UI interaction, supplies the model; an SDK lets developers build and deploy workflows around it. An act call translates natural-language instructions into actions on the screen.

In Carolyn’s demonstration, the task is to find a two-bedroom, one-bath apartment. The first act call initiates a sequence of browser operations. The agent considers the result of each step before planning the next, rather than assuming a fixed sequence will work regardless of what the page does. The visible Beds/Baths menu captures one intermediate filtering step, including the one-bathroom choice. High reliability on UI tasks is the stated training objective; the demonstration supplies no numerical workflow success rate.

7:458:00
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:45 · section reference included

Extract records, compare commutes

Once the browser displays rentals, Fjord extends the workflow using ordinary Python. First, a Pydantic class defines the desired records, and a structured extraction asks the agent for JSON matching that schema. This is the boundary between interpreting a visual page and manipulating explicit data: subsequent code can operate on apartment records instead of browser pixels.

The next question is personal to the search: how far is each apartment from the nearest Caltrain station by bicycle? A helper named add_biking_distance takes an apartment and uses Google Maps to calculate its commute. A Python thread pool runs the searches in parallel, with a separate browser for each address. Finally, Pandas turns the enriched records into a table sorted by biking time.

The composition can keep the browser helper separate from validation and ranking. Here, address and biking_minutes make the apartment-to-commute data boundary explicit; add_biking_distance supplies the browser work described above.

python

from concurrent.futures import ThreadPoolExecutor
from typing import Callable

import pandas as pd
from pydantic import BaseModel, Field


class Apartment(BaseModel):
    address: str


class ApartmentResults(BaseModel):
    apartments: list[Apartment]


class ApartmentCommute(Apartment):
    biking_minutes: float = Field(ge=0)


extraction_schema = ApartmentResults.model_json_schema()


def rank_apartments(
    extracted_json: str,
    add_biking_distance: Callable[[Apartment], ApartmentCommute],
) -> pd.DataFrame:
    apartments = ApartmentResults.model_validate_json(
        extracted_json
    ).apartments
    if not apartments:
        return pd.DataFrame(columns=["address", "biking_minutes"])

    with ThreadPoolExecutor() as pool:
        commutes = list(pool.map(add_biking_distance, apartments))

    return (
        pd.DataFrame([commute.model_dump() for commute in commutes])
        .sort_values("biking_minutes")
        .reset_index(drop=True)
    )

Natural language handles the UI-specific work; Python handles record structure, concurrency, and ordering. Sorting by biking time also makes the user’s objective concrete: the desired output is a commute comparison, not merely a successfully operated search form.

The demonstration points to the apartment example in the Nova Act SDK repository. Its current documentation includes functionality added after this research preview. Perszyk describes the minimal starting script as three lines of code, excluding installation and authentication. She also reports that underlying model improvements were shipping every few weeks. The small entry point and the richer Python workflow serve different needs: beginning with one action, then composing actions into a useful application.

9:079:15
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:07 · section reference included

Reliable actions need a shared environment

Even the smallest browser interactions involve interpretation. On Amazon’s website, icons have labels, but many interfaces leave their meaning implicit. People often infer an unfamiliar icon’s purpose from surrounding cues. An agent cannot be explicitly taught every possible icon, much less every useful way to operate a computer. Perszyk therefore points to reinforcement learning as a way for agents to explore and learn these interactions.

Exploration might produce computer-use strategies unlike ours. That can be valuable for a complementary system, provided the agent and the person still interpret the relevant environment consistently. Perszyk contrasts assistants built around LLM wrappers, which she characterizes as lacking environmental grounding and a world model, with computer-use agents that see pixels and act on UIs. She describes the latter as an early form of embodiment: their actions have observable consequences in a digital environment shared with users.

Nova Act’s emphasis is on making the smallest interaction units reliable while exposing granular developer control. Those units can then compose into larger workflows, just as words combine into more complex meanings. Shared grounding is necessary, but it is not sufficient for understanding a higher-level goal. An agent may correctly interpret a page and execute its controls while still missing what the person ultimately wants to accomplish.

Two robot panels contrast predicting the next text token without a world model with acting on a computer while learning a digital world model.Open full source frame
Narrow AI and pre-general agents contrasted through digital grounding.

The next development problem is therefore broader than improving action accuracy: what would let computer-use agents reliably understand those larger goals, and thereby increase their users’ capabilities?

10:1610:28
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:16 · section reference included

Intelligence distributed across minds

The history of intelligence augmentation extends beyond computers. In Augmenting Human Intellect: A Conceptual Framework, Engelbart described the evolution of augmentation methods as a development of artificial intelligence already underway for centuries. Perszyk takes that perspective further back, asking how human intelligence itself became flexible and reliable.

Her compressed evolutionary account begins around six million years ago with an environmental change and a social feedback loop: larger brains, more connections between individuals, better use of social information, and further brain growth. This is a proposed explanation of social cognition, not a precise chronology of brain enlargement. The Smithsonian’s account of brain evolution places the fastest enlargement much later.

The other half of the proposed loop concerns how a group becomes more capable. People offload computation to one another, distributing cognition across a social environment. Representational alignment is the ability to reproduce aspects of what is in one mind in another mind, making cooperation possible. In this account, inferring the existence of other minds lets people use one another’s knowledge and abilities as tools.

General intelligence, on this view, does not by itself guarantee alignment, and human-like general intelligence cannot be adequately located inside an isolated person or model. It emerges through interactions and continues changing with them. That changes what an agent evaluation should examine:

  • Creativity: whether working with the agent improves people’s ability to develop ideas.
  • Productivity: whether the interaction helps people accomplish useful work.
  • Strategic thinking: whether people become better able to reason about larger goals.
  • Flow: whether the system supports sustained, absorbed activity.

Perszyk proposes these human outcomes as evaluation targets beyond model capabilities and time spent on a platform; she does not present measurements for them here.

12:2712:44
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:27 · section reference included

Language connects communication and representation

Perszyk identifies language as the adaptation connecting this social account to generalization. Language and models of other minds reinforce one another: sensitivity to social cues stabilizes those models; more stable models support richer language; richer language further stabilizes what people can represent about one another. Communication and representation become integrated through that reciprocal process.

Her stronger theoretical proposal is that a model of a mind became an original placeholder concept: something like a variable capable of standing for any concept another mind might contain. She presents this as a route to generalization and distinguishes it from communication systems that, in her account, lack models of minds.

The contrast with programming languages concerns how meaning is established. Code does not negotiate meaning with its reader in real time; Perszyk offers that stability as a reason code is comparatively easy to verify. Human language, by contrast, depends on interpreting what another person means. Her claim that LLMs do not understand language uses this specific criterion: understanding that words refer to things minds invent, rather than only producing appropriate linguistic sequences.

A statement linking human language to modeling minds points to four explanatory groups. Red crosses mark nonhuman animal communication, programming languages, and LLM outputs above it.Open full source frame
Human language connects representation, communication, co-evolution, and generalization.

The point of asking what is in a word is therefore to look beyond the symbol. In Perszyk’s account, what gives it its flexible meaning is a mind—and another mind’s ability to infer what it represents.

15:2915:37
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:29 · section reference included

Agents as a collective subconscious

Language leads into a broader history of cognitive technologies: tools that enable increasingly abstract thought and provide foundations for later tools. They become useful through development within communities. Early computers had interfaces that limited their usefulness to many people; interactive computing expanded who could use them and what those users could accomplish. Today, abundant access to information brings a different problem: abundant distraction.

Agents could address that problem in several complementary ways:

  • Handle repetition: perform recurring work that would otherwise consume conscious attention.
  • Share skills: learn from people and make those abilities available across communities.
  • Teach discoveries: bring newly discovered knowledge back to their users.

Perszyk calls this potential role a “collective subconscious”: useful activity distributed across agents and people, without requiring everyone to attend to every step.

This returns to the opening question of controlled hallucination. In her account, tools for thought stabilize thinking and reorganize brains by directing attention toward common features of the environment, selecting relevant signals from noise, and helping people stabilize those signals into shared world models. Coordination supplies part of the control.

Nova Act can then be understood as a set of primitives for a cognitive technology: a way to align human and agent representations through shared digital activity. As with earlier cognitive technologies, developing that potential requires use in diverse communities. Repeatedly clicking the same place is only the beginning of reliability.

16:5517:09
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:55 · section reference included

Useful products make the next learning loop possible

Reliability ultimately includes understanding the larger goal. Perszyk’s proposed next step is agents with models of human minds. She does not describe those models as objects engineers can directly install. Instead, the task is to create the conditions under which they could emerge: a common language between humans and computers, a model of their shared environment, and interfaces that support intuitive interaction.

That makes useful products part of the research mechanism. People need a reason to interact with agents; those interactions produce the human-agent data needed to advance the models; better models can support more useful products and more capable users. The intended loop is reciprocal: humans help agents improve, and agents help humans think and act more effectively. Perszyk closes by inviting the audience to the upcoming Nova Act workshop, where that broad ambition returns to the practical work of building with agent calls.

18:4218:58
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

18:42 · section reference included

Resources

From the talk

  • The original research-preview announcement explains atomic browser commands, Python composition, and the scope of early reliability evaluations.

  • Perszyk's later companion essay explains how agents could reduce digital drudgery while strengthening human thought and agency.

Transcript download
Full transcript

Loading transcript…