Perspectives

Toward General-Purpose AI

Disclaimer, Version 0.1: These are my personal working hypotheses, not a definitive roadmap to AGI or the views of any organization with which I have been affiliated. I am writing them down to make them concrete enough to examine and challenge. I expect these views to change as I learn, run experiments, and encounter better evidence.

Current Thesis

I believe progress toward general-purpose AI is more likely to come from scalable methods that can learn from increasingly diverse data, compute, and interaction than from systems built around many hand-designed rules. Language is extraordinarily useful, but text is only one representation of the world. Images, audio, video, physical sensors, demonstrations, and actions preserve spatial, temporal, causal, and social information that can be difficult or impractical to express completely in words.

My current hypothesis is that systems capable of learning jointly from these sources, including the consequences of their actions, have a better chance of developing useful models of the world. Those models could help us evaluate possibilities before committing scarce real-world resources. I do not think such systems should become unquestioned automated decision-makers. Their purpose should be to expand human agency while people continue to set goals and remain accountable for consequential decisions.

Language Is a Powerful but Lossy Interface

Most interactions with language models require a person to translate a situation into text. Before a model receives anything, the person must notice the relevant details, understand them, and describe them accurately. Information can be lost at every step.

A multimodal system changes that interaction. Instead of describing every visible detail, I could show it an object or procedure. With wearable devices, a system could observe the situation alongside me while I point, speak, demonstrate, and correct it. The important distinction is not that multimodal systems avoid internal tokens or representations. It is that a person no longer has to manually compress the entire experience into language first.

This matters for tacit knowledge: things people can recognize or demonstrate but struggle to specify completely. Driving smoothly, manipulating an unfamiliar object, diagnosing a machine by its sound, and adjusting an action in response to subtle feedback all contain information that a written explanation may omit.

Perception, Action, and Grounded World Models

Multimodality should mean more than attaching an image encoder to a language model. A system deployed in the world could receive vision, sound, touch, pressure, temperature, movement, force, and other measurements. Machine perception does not need to be limited to human senses; it could also include depth, infrared, precise vibration, or other signals useful for a task.

Action is equally important. Passive observation reveals correlations, while intervention can reveal how the world changes in response to an action. A system that moves an object and observes the result receives a different learning signal from one that only reads a description of the event.

More sensors do not automatically produce understanding. The algorithms and training objectives must use synchronized observations and actions to learn persistence, causality, uncertainty, and consequences rather than superficial shortcuts. My belief is therefore not simply that we need more data. We need learning methods whose capabilities continue to improve as multimodal experience and compute scale.

Scaling as a Design Principle

I am influenced by the Bitter Lesson: general methods that benefit from computation have repeatedly outlasted approaches that depend heavily on encoding human expertise by hand.

For multimodal systems, this suggests prioritizing architectures, representations, and objectives that can absorb additional modalities and improve predictably as data and compute increase. The aim should not be to reproduce every human mental mechanism. It should be to develop general learning procedures that can discover useful structure from observation, interaction, and feedback.

This remains a hypothesis rather than a guarantee. More compute can amplify weaknesses in the data or objective, and more bytes do not necessarily provide more useful information. Scaling is valuable only when the system can turn the additional experience into transferable capability.

Simulation Before Intervention

A sufficiently grounded world model could help evaluate possible actions before we spend resources executing them. Instead of physically testing every candidate, a system could simulate alternatives, identify likely failures, estimate uncertainty, and recommend the few real-world experiments expected to be most informative.

Potential applications include scientific experiments, engineering designs, robot trajectories, operational plans, and candidate medical interventions. The purpose of simulation should be to narrow a search space, not to replace verification. A learned model can be confidently wrong, particularly outside its training distribution. Reality must remain the final test.

General Learning, Not Universal Mastery

I do not think generality requires a system to begin as an expert at every task. Humans have uneven abilities: a person who cannot perform surgery or tell a good joke is not therefore considered incapable of general intelligence.

A more useful definition may be general learning ability. Given appropriate time, instruction, demonstrations, tools, practice, and feedback, can a system acquire a genuinely unfamiliar skill? Can it transfer prior knowledge, recognize its uncertainty, improve from mistakes, retain the new capability, and adapt when the environment changes?

This suggests an empirical test for generality: place a system in an unfamiliar environment and determine whether it can learn a new capability from limited multimodal experience and correction rather than relying on a memorized solution. Broad pretrained competence and general learning are related, but they are not the same thing.

A Deployment-Driven Improvement Flywheel

Deployment can expose limitations that static benchmarks miss. Users attempt new tasks, encounter failures, and reveal where a system lacks knowledge or behaves unreliably. Those observations can become new evaluations, corrections, and carefully curated training signals for subsequent systems.

I currently think of this as a deployment-driven improvement flywheel rather than autonomous recursive self-improvement:

1. People use a system in real situations.
2. New uses expose capability boundaries and failure modes.
3. Researchers convert those failures into evaluations and trustworthy learning signals.
4. The next system improves on the measured weaknesses.
5. Greater capability enables more demanding uses, beginning another cycle.

This process introduces serious risks: privacy violations, reinforcement of popular misconceptions, optimization for engagement rather than usefulness, and degradation from repeatedly training on low-quality synthetic output. The quality and provenance of feedback matter as much as its quantity.

Human Agency as a Design Constraint

I want increasingly capable AI to expand human agency rather than remove people from meaningful control. A system can perform substantial work, but people should continue to determine goals and remain accountable for decisions that affect lives and institutions.

“Human in the loop” is not enough if the human merely approves a recommendation they cannot understand. Useful augmentation should help people investigate evidence, compare alternatives, understand uncertainty, create, learn, and make better decisions. The aim is not simply to automate what humans currently do. It is to increase what individuals and groups can understand and accomplish while preserving human responsibility.

I also do not assume that more capable machine intelligence will become increasingly human. Artificial systems may develop forms of cognition shaped by their sensors, objectives, data, and computational substrate rather than by animal evolution. Human value does not depend on remaining the most computationally capable entity.

Open Questions

These are the questions on which my current position is least settled:

• Which training objectives force multimodal systems to learn causal structure rather than correlations between modalities?
• How much embodiment is necessary, and which capabilities can emerge from passive data or simulated interaction?
• How should we measure general learning separately from knowledge acquired during pretraining?
• How can a system distinguish useful deployment feedback from noise, manipulation, and repeated model-generated errors?
• When should simulation be trusted to guide action, and how should uncertainty be calibrated?
• Which decisions must remain under meaningful human control as systems become more autonomous?
• Does intelligence require anything analogous to emotion, or are emotion and intelligence separable dimensions?

Revision Log

• August 2026, Version 0.1: Initial statement of my hypotheses about scalable multimodal learning, action-grounded world models, simulation, general learning, deployment feedback, and human agency.