GoodAI Research

A skill or a heuristic is any assumption about a problem that narrows and diversifies the search for a solution and points the search towards more promising areas. It is not guaranteed to be optimal or perfect, but sufficient to meet immediate goals.

Other names for a “skill” or “heuristic” are: behavior, strategy, ability, solution, algorithm, shortcut, trick, approximation, exploiting structure in data, and more.

Skills can also be considered biases that restrict behavior.

Some skills are simple (e.g. detecting a simple pattern such as a line or an edge) or complex (e.g. navigating through an environment).

General AI is complicated to design from scratch, especially if we want to teach it everything at once (so-called ‘end-to-end’). It is more feasible to do if the whole problem of learning and designing is deconstructed into several (less  complicated)  “sub-problems” which we know how to tackle. For example, it is clear that we want the AI to understand and remember images, so it needs the ability to analyze them and a memory to store data. We want it to be able to communicate with humans, so It will need to write, read and understand language. It will also need to learn and adapt to new things, and much more. We call solutions for each sub-problem skills.

A skill can be seen as an ability or heuristic which helps the AI solve a particular problem. Importantly, each skill can also be used for learning other skills, significantly reducing the search space for solving other sub-problems.

Skills can range from simple and concrete (like the ability to recognize faces, add numbers, open doors, etc.) to more abstract ones (like the ability to build a model of the world, to compress temporal / spatial data, to receive an error signal and adapt accordingly, to acquire new knowledge without forgetting older knowledge, etc.). Skills also provide a simple way to measure how the system works, as it is clear how to measure which system is better in understanding speech, classification, and game playing. However, evaluating general AI as a whole is still unclear.

General AI will essentially be a system that exhibits a very large set of skills. Some of those skills might be hard coded by programmers, but most will be learned. Take, for example, the evolution of humans. Evolution provided us with some hardcoded skills or predispositions, but most of what we know we need to learn during our lifetimes – from our parents, the environment, or society. Those skills cannot be hardcoded because humans, just like an AI, need to be able to adapt to unknown future situations. Sometimes it is also easier to teach the desired skill than to add it as a part of the design. On the other hand, letting the AI discover all skills by itself would be slow and inefficient. This means that our job is to identify essential skills and find the most efficient ways to transfer them to a general AI system – by hardcoding them or by teaching them. It is not necessary to find the best skills. Any skills which have the desired properties and which enable the AI to further learn and improve itself can move us closer to general AI.

Just like an AI has to use efficient methods when searching for problem solutions, AI researchers must also look for efficient shortcuts to narrow the search for the general AI architecture, optimal curriculum, etc., as we can’t effectively explore the entire space of potential solutions. We can, for example, draw inspiration from evolution, animal brains, or other systems designs. Part of the problem is also what general AI architecture and skill set is easier for us to attain now, with our current knowledge and resources. The framework, roadmap, the Institute are all part of a method for narrowing the search for the architecture and the curriculum to teach it.  Individual sub-problems can be also outsourced to other researchers and institutions.

We can ask questions like, “What is the minimal skill set that is sufficient for human-level AGI?” If we can optimize the process by cutting out all unnecessary skills, we can get to our goal faster. On the other hand, the learning algorithm alone wouldn’t be sufficient; we also need thousands or millions of learned skills for the particular domain. Without them, the AI wouldn’t be able to start solving the problems we need. For example, driving a car is not a crucial skill for a researcher AI living only in the world of internet and scientific publications, but a skill such as the ability to generalize to similar, but previously unseen situations is universal, and falls into the category of necessary skills for every general AI).

Intelligence is a problem-solving tool that searches for solutions to problems in dynamic, complex and uncertain environments. From a computational point of view, all problems can be viewed as search and optimization problems and the goal of intelligence (or an intelligent agent) is to narrow the search space in order to find the best available solution with as few resources as possible.

Intelligence achieves this by discovering skills (heuristics, shortcuts, tricks) that narrow the search, diversify it, and help steer it towards areas that are potentially more promising.

One of the most useful skills is the capacity to gradually acquire new skills – which helps in exploiting accumulated knowledge in order to speed up the acquisition of additional skills, the reuse of existing skills, and recursive self-improvement. This way, the intelligent agent slowly creates a repertoire of skills that are essentially building blocks for new, more complex skills.

An intelligent agent operates with limited resources (time, memory, atoms, computation cycles, energy, etc.), which is another constraint put on intelligence, favoring skills that use fewer resources.

Gradual and guided learning also helps narrow the search, because at each step, an intelligent agent has to search for a new solution only within a small and useful area, decreasing the number of candidate solutions, thereby reducing the complexity of the search space. On the other hand, if there was no gradual or guided learning and the agent were expected to find a solution to a complex problem too far from its current capabilities, it might never find the solution.

Teaching the AI through gradual and guided learning, where we fine-tune individual learning tasks in order to teach the AI desired skills (behaviors), will allow us to have more control over the behaviors it will use later to solve novel problems. The AI’s behavior will, therefore, be more predictable.

In this way, we can imprint positive human biases into the AI, which will be useful for future value alignment (between AI and humans) – one of the important aspects of AI safety.

If we have a hard task, a good way to solve it is to break it down into smaller problems which are easier to solve. The same is true for learning. It is much faster to learn things gradually than try to learn a complex skill from scratch. One example of this is is a hierarchical decomposition of a task and gradual learning of skills from the bottom of the hierarchy to the top.

For instance, if you have a newborn child and you give it a task to learn how to get to the airport, the chance that it will learn to do it is really small because the space of possible states and actions is just too large to explore in a reasonable amount of time. But if you teach it gradually with small tasks, for instance how to crawl and then walk, you increase the chances of success, as it can use these skills to try to get to the airport.

We want to build systems that learn gradually. Furthermore, we want to guide their learning in a correct way. Guided learning means showing the system what things make sense to learn and in what order. This reduces the necessity for exploration even further.

Basically, you show the child that it makes sense to learn how to walk and open doors, and only then to try to get to the airport.

Another benefit of gradual learning is that it can be more general. We do not have to specify a single global objective function (the main goal of AI) at the beginning, because we are rather teaching universal skills, which can be used later for solving some new tasks.

In the case of the child, we basically start teaching it to walk and open the door, even if we don’t know it will need to get to the airport later, or to become a dentist, etc.

If we teach skills gradually, we have better control over the knowledge which is learned by the system. Later, if we specify a goal for the system, it is more likely that in order to fulfill it, it will try to use these already learned skills rather than inventing new behavior from the scratch. It means that in this way, we reduce the chances that the system would invent any unwanted or harmful strategy.

This is similar to teaching the child how to walk and open the door, and then to go to the airport. It is more likely that it will try to solve the task by walking and opening the door, rather than trying to learn a completely new skill (like flying) from scratch, because it would be just more difficult.

Performance benefits:

• Optimizing a model that has few parameters and gradually building up to a model with many parameters is more efficient than starting with a model that has many parameters from the beginning. At each step, you only need to optimize/learn a small amount of new parameters.
• There is no need to know the size of the network a priori
• Network size can correspond to the complexity of given problems (there are no neurons or weights to prune)
• Starting with a small network is faster (than the other way around)
• Reuse of existing skills is made possible

A skill or a heuristic is any assumption about a problem that narrows and diversifies the search for a solution and points the search towards more promising areas. It is not guaranteed to be optimal or perfect, but sufficient to meet immediate goals.

Other names for a “skill” or “heuristic” are: behavior, strategy, ability, solution, algorithm, shortcut, trick, approximation, exploiting structure in data, and more.

Skills can also be considered biases that restrict behavior.

Some skills are simple (e.g. detecting a simple pattern such as a line or an edge) or complex (e.g. navigating through an environment).

One way to compare the intelligence of various intelligent agents is to measure and compare their generality, complexity of problems that they can solve, and efficiency of all of their heuristics.

The GoodAI Research team is made up of a few “architecture groups” each working on its own general AI prototype. Each group is designing their own curriculum (School for AI). However, we are aiming to align all of the teams in order to focus on problems similar to the training and evaluation talks from the Gradual Learning round of the General AI Challenge.

Our AI Safety team is studying: how we can advance safely with our technology, how to mitigate threats to our team and humankind as a whole, how to we can create an alliance of AI researchers committed to the safe development of general AI, developing our futuristic roadmap, and more.

The General AI Challenge team formulates the problems for each round of the General AI Challenge, and manage the day to day activities of the Challenge.

GoodAI’s Futuristic Roadmap is our vision for the future and the specific step-by-step plan we will take to get there. The roadmap outlines challenges we expect to come across in the course of general AI  development and our efforts to keep AI safe, and how we will mitigate risks and difficulties we will face along the way.

Our futuristic roadmap is a statement of openness and transparency from GoodAI, and aims to increase cooperation and build trust within the AI community by inspiring conversation and critical thought about human-level AI technology and the future of humankind. While our R&D roadmap is focused on the technical side of general AI development, this futuristic roadmap is focused on safety, society, the economy, freedom, the universe, ethics, people, and more.

​You can find one of the most recent roadmaps here.

GoodAI stands apart from other AI companies because of our roadmap, framework, and big-picture view. We pursue general AI with a long-term, 10+ year vision, and remain dedicated to this goal. We will not be distracted by narrow AI approaches or short-term commercialization, though we are certain to find useful applications for our general AI technology along the way.

Our roadmap, framework, and experimental implementations are at a very early stage and should be taken as works in progress.  We are focused on the gradual accumulation of skills and recursive self-improvement. We do research in growing network topologies and modular networks and train and teach our AI in our School for AI.

We are optimizing the process of building and educating general AI.

Our mission is to build general AI as fast as possible, but this is not a race.

It’s not about competition, and not about making money.

At GoodAI, we want to create a positive future for everyone. Developing general AI will be the most helpful thing in human history, and we want to help make this dream come true.

How does our work contribute to the fields of AI and general AI research?

There is a significant lack of unified approaches to building general-purpose intelligent machines. Comparable to the biological sciences, most researchers, universities and institutes still operate within a very narrow field of focus, frequently without consideration for the ‘big picture’.

We believe that our approach is a way to step out of this cycle and provide a fresh, unified perspective on building machines that learn to think. We hope to achieve this in a number of ways, each of which are equally relevant and essential for tackling different aspects of the building process:

• Our framework provides a unified collection of principles, ideas, definitions and formalizations of our thoughts on the process of developing general AI. This allows us to amalgamate all that we believe is important to define as a basis on which we and others can build. It will act as a common language that everyone can understand, and provide a starting point for a platform for further discussion and evolution of our ideas.
• Our roadmap is a principled approach to clearly outlining and defining a step-by-step guide for obtaining all abilities and skills that a human level intelligent machine needs to possess. This includes their definitions, as well as the gradual order and way in which to achieve them through curricula of our ‘School for AI’.
• Our School for AI provides learning curricula — a principled, gradual and guided way of teaching a machine. This approach differs significantly from current approaches of narrow-focused and fixed datasets. We believe that gradual and guided learning are essential parts of data-efficient learning that are paramount to quick convergence towards a level of intelligence that is above current standards.
• To compare and contrast existing approaches and roadmaps and foster more effective distillation of knowledge about the process of building intelligent machines, our AI Roadmap Institute is a step towards an impartial research organization advancing the search for an optimal protocol for  achieving general artificial intelligence.
• Last but not least, our software infrastructure is comprised of our large-scale and highly parallel Arnold Simulator, able to handle extremely dynamic network topologies, as well as various learning environments. It was developed specifically for the numerous curricula of our School for AI, and serves as an ideal platform for transforming our conceptual ideas to practical implementations with tangible results.

Using the language of our principles, the above are simply a set of heuristics for steering our search for general AI that we believe are important and will help us achieve a significantly faster convergence towards developing truly intelligent machines.