Related labs, independent researchers, and what we learn from them

What is the claim—and why did it upend the AI safety debate?

• The old consensus (polysemantic chaos & superposition): For the past five years, most interpretability researchers assumed that neural networks cram far more concepts than they have neurons into a tangled, chaotic mess (“superposition”), where a single neuron appears to fire for unrelated things like cat faces and car grilles. Under that view, auditing a model required training massive secondary models (Sparse Autoencoders) in corporate datacenters just to untangle what a network is thinking.
• Cammarata’s reversal (“Already Clean”): Re-examining the very consensus he helped establish in 2020, Cammarata argues that in modern, sufficiently capable models, the vast majority of native neurons and activation subspaces already have a single, legible, structured job. Famous examples of “tangled” neurons were largely artifacts of tiny early architectures, weak visualization tools, or human mislabeling (for instance, the “cat face + car grille” neuron is actually a clean geometric V-contour detector).
• Scientific caution & how we test it: Veterans such as Neel Nanda (Google DeepMind) and researchers in the Chris Olah (Anthropic) circuits tradition have rightly noted that this working hypothesis still requires rigorous peer-reviewed replication, particularly in middle transformer layers and under adversarial pressure where covert superposition or steganographic channels could still hide.

Three ways this directly accelerates the Collective Intelligence Network

1. Lightweight, on-device taste distillation (Reclaiming your profile): If native neural activations are already structured along clean concept axes, your personal agent (1B–8B local model) does not need a billion-dollar cloud Sparse Autoencoder to separate what genuinely fulfills you over 30 days (z_telic) from late-night outrage traps (z_limbic). It can read and filter those native concept directions directly on your phone or laptop in milliseconds.
2. Direct native-circuit auditing at the runtime gate (No one commands a swarm): When open-weight models from US, European, or Chinese labs are submitted to run as personal twins, independent watchers can audit their native weight subspaces directly for sleeper backdoors or hidden corporate biases—pairing Cammarata’s native-unit inspection with Neel Nanda’s and Buck Shlegeris’s adversarial control tests.
3. Symbolic-vector git merge on our Principle Realignment Test-Net: Because native activation axes align cleanly with human-legible concepts, each participant’s agent can project lived friction directly onto our 12-dimensional symbolic principle basis (v_u,k ∈ ℝ¹²) and cohere community updates via Byzantine-Robust Geometric Median Aggregation (Weiszfeld, 1937; Pillutla et al., 2022).

Rather than building a brittle internal citation scraper or inventing proprietary impact scores for outside scientists, our research pipeline follows the principle of least resistance: both human readers and personal agents can query live h-index, citation velocity, and preprint provenance in one click.