/aiman.tech
Available→
02Lab · open notebook

The physics
of intelligence

The lab is where I question architectural defaults that became defaults by accident. Some lines are public, some are internal, all of them are live. Below are the threads in flight - Parallax leads.

L-001
State
Active · primary research line · v3
Tags
CWSAttractor dynamicsDEQ-flavouredEnergyContinual
Open page ↗

Parallax

A Central Workspace as a continuous-state dynamical system that settles into a metastable regime in an attractor landscape, perturbed by a heterogeneous pool of expert probes. Reasoning is settling. K loop is decoupled from the token loop. Candidate replacement for the residual stack itself.

Why

The whole project is a slow drift away from the assumption that cognition is the same thing as token-clocked autoregressive prediction. The CWS does K reasoning steps per context update, not per token. It does not see raw inputs - language is just one of several sensory channels. Trained with a DEQ-style contract (K-1 frozen iterations, one live final step), so reasoning depth costs no extra memory at training time. The bet: that you can get the dynamical-system flavour of cognition - settling, pondering, content-addressable basins - without giving up the throughput of modern transformers. Took three full iterations to figure out what the project actually is; v3 is where it started feeling honest.

Open questions
  • How sharp does the attractor regime need to be? Strange-attractor dynamics emerge from K-budget, not architecture - what is the right K policy?
  • Is K-DOF a graded reasoning-depth signal at language scale, or only a binary structural-recognition signal? Synthetic recall says graded; TinyStories at 6k steps says binary.
  • Can the 'replaces ResNet' framing survive 100M+ parameters and 1B+ tokens, or is the constant-memory-in-K finding only useful at small scale?
  • How do you write modality-specific motor decoders without sneaking cognition into them?
L-002
State
Active · own track
Tags
MLPStructural plasticityCompetition
Open page ↗

Dendritic Unit

A research aside. Take the activation function out of the MLP block and let the unit itself be nonlinear, the way biological dendrites are - through learned competition and coactivation between branches.

Why

An exploration into structural plasticity at the unit level. Real cortical neurons compute their nonlinearity inside the dendrites; deep learning bolts a separate ReLU on at the end. I wanted to see what happens if you trade the bolted-on nonlinearity for a learnable selection rule between affine branches. Started as a daydream, kept going because it keeps producing surprises. Not part of Parallax - lives on its own track.

Open questions
  • Does selection between branches give richer or more compositional structure than ReLU + linear at fixed parameter count?
  • How does the soft / hard temperature on the selection rule affect specialization vs coactivation in trained units?
  • Can branch specialization be read out as an interpretable factoring of the input space?
L-003
State
Internal · in-flight
Tags
AttentionLong contextArchitecture

Pi

A focused experiment exploring what kind of inductive bias could replace attention for a narrow class of long-context, structure-rich tasks.

Why

Attention pays a quadratic cost everywhere, but most long-context tasks have structure that pure attention does not exploit. Pi is a small, focused experiment - not a moonshot - asking what the right operator looks like when the structure is known.

Open questions
  • Where does attention overpay - what is the actual information geometry it is paying for?
  • What inductive bias replaces it without giving up the generality that made it useful?
L-004
State
Public · maintained
Tags
EvolutionBehaviorReal-time
On GitHub ↗

Neural Evolution

Self-driving agents in Unity, evolved from scratch. Custom NN, population-based training, configurable from the inspector. A playground for reading the geometry of behavior.

Why

Evolution is a low-bandwidth signal - exactly what makes it interesting. It forces architectural decisions to do real work, because there is no gradient to bail you out.

Open questions
  • How does network topology shape what behaviors are reachable?
  • What does a "fitness landscape" look like when the search space is architecture itself?
L-005
State
Public · supporting tooling
Tags
AgentsToolingInfrastructure
On GitHub ↗

conv-proxy

A conversational proxy with minimal intelligence by design - sits on top of agent frameworks like Openclaw without becoming the bottleneck.

Why

Agent stacks tend to grow more "intelligence" in places that should stay dumb. A thin, predictable proxy at the edges is more useful than a smart one.

Open questions
  • How do you keep a conversational layer inspectable while the agent framework changes underneath?

Want to argue with any of this

aiman@shabib.net