Verses.ai's Genius Agents. For AGI: "Deep learning is not enough"

Offtopic
1

[…]
Deep learning is not enough.
[…]

For me FAQ:

What makes Genius unique compared to other AI systems?

Genius Agents have the ability to Reason (predict), Plan (optimize)*, and Learn (adapt)*.

Inference. Genius is well suited for handling fuzzy problems spaces that have uncertainty, complexity, and/or limited sample data.

Explainability. Unlike blackbox Neural Nets, the data structure in Genius Core and the decisions that Genius Agents make are auditable and human readable.

Adaptive. Whereas today’s AI models are pre-trained on massive amounts of data and then static, Genius Agents learn continuously through experience.

Network Effects. Genius is designed to foster a diverse ecosystem of intelligence whose value increases exponentially as the network grows.

* Future capabilities

Interesting statement:

Hallucinations are a feature, not a bug.

I have no relation to this: The public beta is now open, I did hesitate to post what I read before it, and except for part from one of the the job ads (the other has lesser qualifications, mentions Python, neither mentions Julia or other language), the rest is from my draft, I composed before my holiday:

Essential Qualifications:

  • Demonstrated track record at publishing in top echelon ML venues such as NeuIPS, ICML, ICLR, CVPR, etc on topics such Deep Learning, Reinforcement Learning, ML Theory & Foundations Unsupervised Learning & Generative Models
  • […]
  • Must have a Masters degree, have a PhD or be pursuing doctoral studies

Public Beta Preview on June 20, 2024

[…] Dr. Friston has been recognized as the world’s top neuroscientist and by Wired Magazine as “The Genius Neuroscientist Who Might Hold The Key to True AI.”
[…]
We are developing Genius, a framework that enables automated generation of general intelligent agents.
[…]
Applying the “genius” of Nature lets us develop a new kind of software that is efficient, self-learning, and adaptable in real-time, all while being transparent and accountable.

HOW

In November 2023, our international research team of computational neuroscientists and engineers made an historic breakthrough in the development of AGI. Our alternative approach to machine intelligence, pioneered by Dr. Friston, is called Active Inference. […]

After successfully decoding the underlying mechanisms of intelligence as it functions in the brain and biological systems, we developed a pioneering intelligent framework in software based on these mechanisms. Using the same probabilistic algorithms observed in the brain, we’ve demonstrated the ability to design agents that reason, plan, learn, and adapt without training. We believe this biological approach, following nature’s blueprint for intelligence, is the critical breakthrough needed for the realization of General (AGI) and Superintelligent (ASI) Agents.
[…]
Our Future Of Global AI Governance report offers a first-of-its-kind perspective on global artificial intelligence governance which combines the legal expertise of the world’s largest law firm, Dentons, with our expertise in next-gen intelligent systems, and guidance on socio-technical standards from the Spatial Web Foundation.

WHY NOW?

Recent breakthroughs at VERSES have led to promising evidence of matching and surpassing the performance of state-of-the-art approaches based on deep learning while using orders of magnitude less data and compute.
[…]
AGI’s definition, direction and destination should be public. Given recent opinions shared by OpenAI’s CEO, Sam Altman, regarding the lack of confidence in the viability of LLMs to achieve AGI or ASI, we believe our approach offers a compelling solution.

Our invitation to collaborate stems from OpenAI’s charter, issued in its pursuit of non-competitive AGI development, which states,

If a value-aligned, safety-conscious project comes close to building AGI before we do, we commit to stop competing with and start assisting this project.

Through the open letter, VERSES invites OpenAI to explore our novel approach and collaborate.

WHAT’S NEXT?

The end goal is not AGI. The end goal is what we can achieve as a civilization with AGI. […]

At the heart of VERSES: Active Inference

We are investigating this hypothesis using Active Inference, an approach to modeling intelligence based on the physics of information.

[…] so-called “Artificial General Intelligence” (AGI). Beyond that, […]

Our approach focuses rather on designing a diversity of Intelligent Agents

Active Inference Development Stages

S0: Systemic Intelligence “This is contemporary state-of-the-art AI; namely, universal function approximation […]”,
S1: Sentient Intelligence [We are arguable here, as they define it (possibly the planning part missing), since this was written.]
S2: Sophisticated Intelligence “… This stage corresponds to “artificial general intelligence” in the popular narrative about the progress of AI.” Another doc: #Ability to learn and adapt to new situations"
S3: Sympathetic (or Sapient) intelligence
S4: Shared (or Super) intelligence “Ability to work together with humans, other agents and physical systems to solve complex problems and achieve goals”

The natural path to general intelligence:

Designing explainable artificial intelligence with active inference: A framework for transparent introspection and decision-making
https://arxiv.org/pdf/2306.04025

I would like unsupervised, or semi-supervised, but at least:

Supervised structure learning
https://arxiv.org/pdf/2311.10300

The importance of order
First, the order in which data are encountered—or presented, i.e. the scheduling or curriculum [12, 13]—matters. This follows from the fact that we are dealing with state-space models that have to learn the transitions or dynamics, in terms of a discrete number of paths

In other words, learning of the dynamics or physics is only possible if data are presented in the order in which they are generated. This means that there is some requisite supervision of structure learning; in the sense that the process generating training data has to respect their ordinal structure. Clearly, this is not an issue if the data are generated by the process being learned. However, it suggests that it is not possible to do structure learning or any form of disentanglement [14, 15] in the absence of ordinal structure.

The importance of being discrete
The second issue is the commitment to discrete state-space models. This speaks to the secondary agenda of this paper; namely, to foreground the utility of discrete state-space models in relation to the (implicit) generative models used in most of deep learning. Here, we are reading deep learning as synonymous with the use of backpropagation of errors and requisite differentiability [16]. Differentiability restricts models to continuous state-spaces of the kind that support embedding.

2

Yeah, and in other news, food and energy will be henceforth cost-free for everyone.

1 Like
3

I think you’re objecting to “without training” (do not take it to literally), i.e. that would be impossible (and I agree), but I think there “training” is a code-word for (massive) “pre-training” (would likely also apply to “finetuning”) in deep-learning models, and/or “supervised training/learning” (as opposed to un- or semi-supervised), as opposed to “continual learning”:

Adaptive. Whereas today’s AI models are pre-trained on massive amounts of data and then static, Genius Agents learn continuously through experience.

Of course “learning” is/implies some kind of “training”, not just in software, but also in the brain, “train[ing]” is a frequent word in the in the open access Nature Communications paper “Experimental validation of the free-energy principle with in vitro neural networks” (the link to it is in blue in the sentence preceding the one you quoted):

Pharmacological downregulation of gamma-aminobutyric acid (GABA)-ergic inputs (using a GABAA-receptor antagonist, bicuculline) or its upregulation (using a benzodiazepine receptor agonist, diazepam) altered the baseline excitability of neuronal networks. These substances were added to the culture medium before the training period and were therefore present over training.

“training data” also implies “training”, as from the other paper from the professor I quoted:

learning of the dynamics or physics is only possible if data are presented in the order in which they are generated. This means that there is some requisite supervision of structure learning; in the sense that the process generating training data has to respect their ordinal structure.

[Another interpretation of what you wrote, AGI (which they predict still years of) brings on the end of (economic) scarcity.]