At the boundary of theory and experiment.

Frontier AI Research Collective approaches intelligence fundamentally differently—pursuing creative ideas with first-principles rigor and mathematical elegance.

Read our manifesto
Abstract
FAIRC

Frontier AI Research Collective

Understanding before application, explanation before optimization.

  • Tavish Mankash
  • Keshava Prasad
  • Vardhaman Kalloli

Frontier AI Research Collective · affiliated with LEAP · supported by SIMBO.AI

fairc.org

Abstract

FAIRC is born out of a redefinition of Artificial Intelligence research. Our mission is to pursue AI with the curiosity and rigor of the sciences that came before it: seeking understanding before application, and explanation before optimization.

Keywords: language modelling · neuroscience-inspired intelligence · knowledge representation · reinforcement learning · mechanistic understanding

1The question

AI has advanced rapidly, yet so much of it remains centered around applications, while fundamental questions remain untouched. We still understand remarkably little about how intelligence can emerge from limited experience, or what mechanisms allow intelligent systems to learn, generalize, and reason.

Statement of vision

To answer questions no one is answering, and to ask questions no one is asking.

Figure 1 is a small instance of the first question, running on this page. A network with 337 parameters sees n labelled points drawn from two interleaved spirals [1] and must decide the label of every other point in the plane. With 128 examples it recovers the spirals. With 8, many different functions fit the data equally well, and retraining from a new seed shows how arbitrary the choice between them is. When, and why, a learner extends its experience in the right direction is the kind of question we mean.

Training examples, n
Examples
128
Step
0
Loss
—
Held-out accuracy
—
Click adds
Figure 1. Learning from limited experience. A 2–16–16–1 network trained in your browser with Adam [2] on the cross-entropy L(θ)=−1n∑ilog⁡pθ(yi∣xi)\mathcal{L}(\theta) = -\tfrac{1}{n}\sum_{i} \log p_\theta(y_i \mid x_i), where fθ(x)=pθ(1∣x)f_\theta(x) = p_\theta(1 \mid x). Shading shows fθf_\theta across the plane; the solid line is the decision boundary fθ=12f_\theta = \tfrac12, and the dashed lines are the 0.15 and 0.85 level sets. Held-out accuracy is measured on 600 points the network never sees. Click the field to add examples of your own.

2Frontiers

At FAIRC, we pursue multiple frontiers in foundational Artificial Intelligence, including Language Modelling, Neuroscience-Inspired Intelligence, Knowledge Repre­sen­ta­tion, and Rein­force­ment Learning. Our scope is broad by design. We pursue any line of inquiry that advances the mechanistic understanding of intelligent systems: from the mathematics of representation to the biology of cognition, and from the dynamics of learning to embodied agents.

Each plate below renders a canonical object from one of these frontiers. Three are live simulations; the first is a schematic of attention patterns reported in the interpretability literature. All of them respond to the pointer.

Plate I

Language Modelling

Causal self-attention, softmax(QK⊤/d )V\mathrm{softmax}\big(QK^{\top}/\sqrt{d}\,\big)V [3], over a sentence from our manifesto. Row ii shows how token ii distributes its attention over the tokens before it. The plate cycles through four stylised head types: previous-token, attention sink [5], induction [4], in which the second “is” attends to what followed the first, and uniform mixing. Hover to read a weight; click for the next head.
Plate II

Neuroscience-Inspired Intelligence

Forty leaky integrate-and-fire neurons [6], each obeying τ v˙=−v+I(t)\tau\,\dot v = -v + I(t) and resetting after a spike. Twelve share a rhythmic 5 Hz input and fire together as an assembly (red), the kind of co-active group Hebb proposed as a unit of thought [7]. The trace beneath shows population rates. Click to stimulate the assembly.
Plate III

Knowledge Representation

A small semantic network [8] over concepts from our manifesto, stored as triples (h,r,t)(h, r, t) and laid out by a spring–electric simulation [9]. Hover a concept to read its relations; drag it to rearrange the graph.
Plate IV

Reinforcement Learning

Value iteration [10] on a gridworld with slippery moves, V(s)←max⁡a∑s′P(s′∣s,a) [r+γV(s′)]V(s) \leftarrow \max_a \sum_{s'} P(s' \mid s,a)\,\big[r + \gamma V(s')\big] [11]. Value spreads backward from the reward one sweep at a time; arrows mark the greedy policy, which an agent then follows. Click to move the reward; shift-click to add or remove a wall.

3Output

Research, systems, and tools we make public. Code and systems live on GitHub; models, datasets, and experimental records live on Hugging Face.

4Collective

FAIRC is a small collective of people who want to understand difficult things together. We are small by design: an idea should be able to change what the collective takes on next.

  • Tavish Mankash
  • Keshava Prasad
  • Vardhaman Kalloli

References

  1. [1]Lang, K. J., & Witbrock, M. J. (1988). Learning to tell two spirals apart. Proceedings of the 1988 Connectionist Models Summer School.
  2. [2]Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization. International Conference on Learning Representations.
  3. [3]Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems 30.
  4. [4]Olsson, C., Elhage, N., Nanda, N., Joseph, N., et al. (2022). In-context learning and induction heads. Transformer Circuits Thread.
  5. [5]Xiao, G., Tian, Y., Chen, B., Han, S., & Lewis, M. (2024). Efficient streaming language models with attention sinks. International Conference on Learning Representations.
  6. [6]Lapicque, L. (1907). Recherches quantitatives sur l’excitation électrique des nerfs traitée comme une polarisation. Journal de Physiologie et de Pathologie Générale, 9, 620–635.
  7. [7]Hebb, D. O. (1949). The Organization of Behavior: A Neuropsychological Theory. Wiley.
  8. [8]Quillian, M. R. (1968). Semantic memory. In M. Minsky (Ed.), Semantic Information Processing. MIT Press.
  9. [9]Fruchterman, T. M. J., & Reingold, E. M. (1991). Graph drawing by force-directed placement. Software: Practice and Experience, 21(11), 1129–1164.
  10. [10]Bellman, R. (1957). Dynamic Programming. Princeton University Press.
  11. [11]Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.). MIT Press.

From the manifesto

Dissatisfaction is the thrust; discovery is what answers it.

Read the manifesto