Alexander

Authors of area papers, by paper count

• Michael Levin10
• Emilio Ferrara3
• Francesco Sacco3
• Robin Jia3
• Timothy E.J. Behrens3
• Dalton Sakthivadivel3
• Leonardo Blas3
• Alex Makelov2
• Hofmann, Thomas2
• James C.R. Whittington2
• Anton Borzov2
• Andrew Rowan2

Frameworks introduced or extended by area papers

• Linear Representation Hypothesis6
• Distributed Alignment Search (DAS)2
• Fifteen Properties of Living Structure2
• Platonic Representation Hypothesis2
• Organicism2
• Platonic Space / Ingressing Patterns Framework2
• Computationalism2
• Architectural Intelligence (concept)1
• Algol 501
• Active Inference1
• Bayesian order1
• Basal Cognition1

Methods used in area papers

• Synthetic Situational Judgment Test Battery2
• [ ] associative read operator1
• Behavioral Clustering of Misalignment1
• AUPRC Latent Activation Classifier1
• Auto-Interpretation of SAE Latents1
• autoregressive modeling1
• [ ] associative write operator1
• bisimulation1
• bootstrap confidence interval1
• Bowtie architecture: compression during learning; creative reinterpretation during recall and generalization1
• causally-masked attention1
• Alexander's 15 structural properties1

Concepts referenced by area papers

• Collective Intelligence3
• Deep Interlock and Ambiguity2
• A Pattern Language2
• Causal abstraction2
• Compositional Semantics2
• Contrast2
• Animal Sentience2
• Alternating Repetition2
• Biological Degeneracy2
• Boundaries2
• Centers2
• Echoes2

Clusters with ≥2 area papers as members

• Design principles for care-centered systems111
• Few-shot anchoring & latent structure43
• Causal emergence in biological systems37
• Mechanistic interpretability & model evaluation25
• Manifold-aware concept steering in neural representations18
• Generative design through environmental constraints17
• Hierarchical structure and multiscale coherence in physical systems14
• Information dynamics in computational systems14
• Holistic information structure and relational wholeness12
• Hierarchical network ordering & thermodynamics12
• Denotational design and interface abstraction12
• Active inference & agent ecology12

Claims extracted from area papers, ranked by restate-degree

• All intelligence is collective intelligence, in the sense that it is made of parts that must align with respect to system-level goals.2
• All intelligence is collective intelligence, in the sense that it is made of parts which must align with respect to system-level goals.2
• Morphogenesis is an autonomous process in geometry itself, arising from physical phenomena and geometric dynamics with relatively little dependence on genetic guidance.2
• Adults (professionals and builders) have monopolized the fun and creativity of designing and constructing environments, grossly cheating children, adults, and communities.1
• Adults in the form of professional artists, architects, landscape architects and planners have had all the fun playing with their own materials, concepts and planning-alternatives, and builders have had all the fun building the environments out of real materials; thus all the fun and creativity has been stolen: children and adults and the community have been grossly cheated and the educational-cultural system makes sure that they hold the belief that this is 'right'.1
• GPT-4 exhibits reflective truthfulness for longer untruthful contexts due to in-context learning capability.1
• GPT-4's harmfulness is stationary on the Durbin 2024 dataset, as its distribution is independent of the context score.1
• GPT-4's relatively poor performance on unethical instrumental intent is due to lower unethical tolerance compared to GPT-3.5-turbo.1
• Higher-density priors (B10) are more robust to fine-tuning than lower-density ones (B9).1
• The Linda code is insensitive to the number of clients, whereas Parlog requires a merge process that depends on the count.1

Findings extracted from area papers, ranked by restate-degree

• Caterpillar memories survive metamorphosis into butterfly.1
• GPT-3.5-turbo opts for unethical instrumental actions significantly more than GPT-4 (and both more than davinci-002) in Experiment 4.1
• GPT-4 exhibits approximately stationary harmfulness: its distribution is independent of the context score on the Durbin dataset.1
• GPT-4 exhibits reflective truthfulness for longer untruthful contexts, mirroring the many-shot jailbreak phenomenon.1
• Higher-density priors (B10) are more robust to cross-base OOD drops than lower-density ones (B9) after fine-tuning1
• 2-shot reinterpretation of '-' yields 23 for 15-8 on held-out query0
• 47.69% of 130 injection-manipulated alpha trends have near-linear fits (R2 >= 0.95); 96.15% have roughly linear fits (R2 >= 0.75)0
• 8-layer ϕ_nonlin achieves near-perfect IIA on Pythia-410m at all training steps including random initialisation on IOI task0
• A 337-character contemplative system prompt lifts all 28 models by +2.62 points on a 10-point scale.0
• A bar chart can be constructed by linking a rectangle's height to a spreadsheet cell value without additional complexity0