The Secret is Out: How we're Building AGI

Our Roadmap to Adaptive Autonomous Adult-level General Intelligence

History and Context

• Our Cognitive AI approach has its genesis in an initial 5 years of focused first-principles research into intelligence — epistemology (theory of knowledge); cognitive, developmental, and psychometric psychology; neuro-science; computer science and AI. This study culminated in our theory of AGI

• In 2002 I co-coined the term ‘AGI’ and published an outline AGI architecture

• Since then we’ve alternated between actually implementing and refining this design, and commercializing core aspects of the technology

• The commercial version has generated more than $100 mil in enterprise revenue

• In late 2024 we again shifted our focus to pure AGI development — to make the final push to full adult-level AGI

• Most members of our current 12-person team have been on the project for 8 years or more. I’m leading the company as CEO and Chief Scientist, having previously taken an ERP company from the garage to an IPO in 7 years

We believe that Big Data, statistical systems like GenAI/ LLMs cannot achieve AGI — they inherently lack key abilities essential for true intelligence. Cognitive AI, on the other hand, can meet all of the requirements.

The most fundamental requirement is the ability to incrementally validate and integrate new knowledge conceptually on the fly. This real-time, autonomously learning must be monitored and controlled by a proactive metacognition.

Additionally, the core architecture needs to deeply integrate learning, memory, and reasoning with each other, and with a common underlying knowledge representation for both knowledge and skills.

Perception, thought, action, language and higher-level symbolic reasoning operate within a fully integrated neuro-symbolic architecture (INSA).

From AGI to Narrow Commercial, back to AGI

Around 2019 — long before LLMs — our previous dedicated AGI prototype delivered several impressive demos. However, shortly afterwards we pivoted to commercializing our technology. In order to rapidly achieve robust competence for call center applications we added many hand-crafted language, logic, and business rules, thus stepping into the Narrow AI Trap (with eyes wide open).

The Direct Path to AGI

Just over a year ago we launched a focused project to achieve full adult-level general intelligence in the fastest, most direct way. This roadmap is founded on the vast experience and technology we’ve accumulated over the years. Specifically, we took our latest commercial proto-AGI engine and stripped out all hand-crafted rules and specialized code. We also added vision and other perceptual input back.

This new engine then served as the foundation for developing our AGI target:
An AGI brain that lives, learns, and operates on a single off-the-shelf computer. A system can learn by itself to (at least) do any cognitive task that a smart, educated human remote-worker is capable of learning — and to achieve this with no additional training material, help, or time.

A key insight and guiding principle is that the system must be able to learn and adapt autonomously.

To provide a solid foundation for adaptive language and reasoning ability we begin by teaching the system like a child — with grounded examples and exercises. As it learns at an increasing ‘age’ level (roughly following child development from 3 years up) it learns more and more autonomously, until it can ‘hit the books’ largely unaided.

About half of our team are ‘AI Psychologists’ (a profession I invented) who create curricula carefully designed for our AI’s specific needs at every stage of development. They also design and build tests to probe the understanding and capabilities of the system. Working closely with the engineering team we rapidly iterate towards increased autonomous intelligence.

Some Design Philosophy Details

Humans don’t need 10 trillion words or gigawatts of energy to learn language. We also aren’t equipped with special complex algorisms. Evolution managed to get us to be intelligent with relatively simple mechanisms of pattern learning, matching, prediction, completion, association and generalization, as well as memory, aping and some reinforcement learning. These are also the primary mechanisms of our system. No backprop.

Naturally, AI systems have many advantages over biological ones: The potential of photographic memory and higher speed, 24/7 full focus, much better logic abilities, built in math and other tools, as well as instant access to online information.

The Roadmap to Real AGI Deployment

Having trained our system to roughly 3 and 4-year-old language and reasoning ability (with less than 10 man-years of effort and zero cloud training cost!) we’re now in the middle of expanding to 5-year-old cognition.

Our specific plan is to rapidly iterate the learning/ cognition development cycle all the way up to adult level. This involves increased language competence, more abstract comprehension, reasoning and problem solving ability, as well as improved meta-cognition including theory of mind.

Each development milestone is validated via our own benchmarks as well as industry-standard cognitive tests, including novel task learning and completion.

In addition to cognitive development we also need to scale vision acuity, overall capacity, external system integration, and knowledge acquisition. For these tasks we can leverage existing technology and outside skills.

As the system gets closer to fully autonomous learning (proactively asking for help as needed) we will initiate exclusive partnerships with leaders in various industries to commercially deploy our AGI. At this stage the system will be able to teach itself to a graduate-level STEM education which will form the baseline for more specialized applications. We will publicly release our technology once transparent, safe, reliable learning and cognition have been fully validated.

Remaining Challenges

At the moment by far the biggest challenge is our team’s small size. We have clearly identified tasks for about 50 additional people to achieve our goal within 18 to 24 months. While there are still a great number of specific problems we’ll need to solve, we don’t see any fundamental theoretical or technical issues. We’re definitely not at a research stage but rather in full development mode. We expect that remaining engineering challenges will be solved with just additional man-power.

Opportunity

Our approach to AGI is unique. It is based on a solid theory of what makes human intelligence so powerful and does not suffer any of the limitations of big data statistical systems.

We look forward to releasing trustworthy productive AGI to the world to boost individual human flourishing. The rewards aren’t just financial, but importantly providing a path away from the destructive attention-capture economy of current AI towards a future of increased human agency and positive abundance.

We’re currently looking for a strategic partner aligned with our vision.