Open Source Tools and Platforms for AGI Development
General Artificial Intelligence (AGI) represents a transition horizon where machines' capabilities surpass specific tasks, achieving the ability to understan...
Much of the research on artificial general intelligence is done with open source tools. That carries an advantage beyond cost: anyone can check what state a project is actually in, without relying on what its front page says. This article reviews the relevant platforms and, along the way, teaches how to make that check.
The theoretical framework that does have a document behind it
The most cited formalization of the problem is Marcus Hutter's, set out in «Universal Artificial Intelligence: Sequential Decisions based on Algorithmic Probability» (Springer, 2005). The idea, briefly: unite sequential decision theory with algorithmic information theory to define an optimal agent — what he calls AIXI — choosing actions that maximize expected reward across all hypotheses compatible with what has been observed, weighted by their simplicity.
AIXI is not implementable: it requires infinite computation. Its value lies elsewhere, and it is worth understanding because it explains what theory is for in this field: it establishes an upper bound to compare against. Hutter's own page keeps the book's table of contents, close to four hundred slides from his courses, and a list of open problems with cash prizes.
The platforms, and their actual state
OpenCog
OpenCog is the most ambitious open source project explicitly aimed at AGI. Its centrepiece is AtomSpace, a knowledge store in hypergraph form on which symbolic reasoning and learning processes operate. Its bet is the opposite of the dominant one: instead of scaling learning from data, build explicit common sense and conceptual understanding.
And here it pays to be exact about its state, because this is where coverage usually fails. The AtomSpace core is declared active, stable and supported, and several repositories show recent activity. But the project itself classifies part of its code as «fossils» — abandoned — and carries a «help wanted» notice citing lack of resources. Live development toward AGI has moved to a declared successor, OpenCog Hyperon, built under SingularityNET.
Presenting it as a project in full swing would be inaccurate. Presenting it as dead would be too.
SingularityNET
SingularityNET now describes itself as «research and infrastructure for decentralized AGI», and presents itself as a founding member of the Artificial Superintelligence Alliance. Its premise is a decentralized marketplace where different AI services can be combined, and it is also where the Hyperon framework mentioned above is developed.
A note for anyone reading older material about the project: its token is no longer AGIX but FET, following the merger into that alliance. Any text still discussing AGIX is out of date.
ROS, and a misunderstanding worth clearing up
ROS appears on almost all these lists, and is almost always described wrongly. The acronym stands for Robot Operating System, but it is not an operating system. Its own site says so plainly: «The Robot Operating System (ROS) is a set of software libraries and tools that help you build robot applications.» Its official repository qualifies it as a «meta operating system».
The distinction is not pedantry. An operating system manages a machine's hardware; ROS runs on top of one — typically Linux — and provides a communication model between modules. What makes it relevant here is not that it serves to build an AGI, but its architecture: independent components communicating by messages, a modularity scheme that does interest anyone thinking about integrating heterogeneous capabilities.
And it is a good general reminder: a project's name is not its documentation.
Why a cognitive architecture is not a model
A confusion runs through this whole terrain and is worth undoing, because it explains why these projects compare badly with what dominates the headlines today.
A language model is a function: text in, text out, and everything it knows lives in its weights. A cognitive architecture like AtomSpace is a different thing: a store of explicit knowledge plus processes operating on it. The knowledge is not dissolved into numbers; it is written in structures that can be read, edited and contradicted.
That difference brings real advantages: a fact can be added without retraining anything, you can ask why the system concluded something and follow the chain, and a specific error can be fixed without touching the rest. And it brings the disadvantage that has decided the contest so far: somebody has to put that knowledge in, and the world holds more facts than anyone will write by hand.
Which is why comparing the two approaches by their score on a language benchmark is comparing badly. They do not compete on the same dimension: one scales with data, the other with human modelling work.
How to read an AGI project's roadmap
These projects publish plans years ahead, and a criterion helps there, because it is where ambition is most often mistaken for progress.
Separate what already works from what is planned. A repository with recent activity demonstrates the first; a roadmap demonstrates nothing, however detailed. Ambition is cheap to write and code is not.
And distrust the leap without steps. When a plan goes from one working component to a general capability without explaining what lies between, that gap is the whole project. Honest projects say so — OpenCog marks its dead parts and asks for help; that is worth more than any promise.
And one check that costs a minute and saves months: look at the date of the last change in the component you need, not in the project overall. A very active repository can have the exact module your work would depend on sitting frozen, and that distinction appears on no front page.
Measuring progress: where to look
Evaluating progress toward generality is an open problem, and the most useful practical reference is not in these platforms but in the tests designed for it. The most cited is the Abstraction and Reasoning Corpus, proposed by François Chollet in «On the Measure of Intelligence» (2019), where he argues that what must be measured is not what a system can already do but its «skill-acquisition efficiency». The project continues at ARC Prize, with updated results.
The capability: how to check whether an open project is alive
This is the skill the article leaves behind, and it works for any open source tool you are asked to adopt:
1. Go to the repository, not the front page. A project website updates when there is something to announce; a repository records every change with its date. Look at when the last commit landed in the components you care about, not in the project overall.
2. Look for what the project itself declares dead. Honest projects mark their obsolete parts — OpenCog calls them «fossils». That label is worth more than any external review, because it is applied by whoever knows the code.
3. Ask where development has gone. When a project has a declared successor, the interesting activity is there. Adopting the predecessor because it is the familiar name is the expensive mistake.
The deep end
- OpenCog's repositories, with their classification into active components, ongoing research, fossils and incubator. Worth looking at before citing the project.
- SingularityNET, where Hyperon is developed.
- ros.org and the ROS 2 repository, to see the definition in their own words.
- Hutter (2005), for the theoretical framework. Link straight to the book page: the site root is an old frameset with no readable content.
- Chollet (2019) and ARC Prize, for the measurement side.
This article was produced with artificial intelligence under human editorial oversight.
