Researchers find sparse neuron subsets can predict factual hallucination in several open models
Migrated the Entry to schema v2 and narrowed the conclusion to predictive, domain-specific signals.
View Entry →Changelog
Material changes to the VyDex evidence ledger.
This page records new entries, meaningful updates, removals, and methodology changes.
Change Types
- Added
- New entry added to the ledger.
- Updated
- Important source, status, evidence, caveat, context, or interpretation changed.
- Removed
- Entry removed because it no longer meets criteria or no longer supports the frontier interpretation.
- Methodology Change
- Rules, labels, categories, or judgment standards changed.
Changes
Updated Updated Gambit says one operator used Claude Code and GPT-4.1 in a campaign targeting nine Mexican government organizations
Migrated the Entry to schema v2 and attributed the disputed campaign scope directly to Gambit.
View Entry →Updated Kosmos AI surfaces a candidate flippase signal linked to neuron vulnerability in Alzheimer’s disease
Migrated the Entry to schema v2 and narrowed the conclusion to a candidate cross-dataset flippase signal.
View Entry →Updated GPT-5 literature searches prompted six Erdős database status changes after finding earlier human results
Migrated the Entry to schema v2 and corrected the title and interpretation to literature recovery.
View Entry →Updated Epoch estimates frontier AI benchmark progress nearly doubled in pace around April 2024
Migrated the Entry to schema v2 and clarified the fitted ECI acceleration and uncertainty.
View Entry →Updated Artificial neuron repeatedly fires at biological-range voltage and energy
Migrated the Entry to schema v2 and clarified the biological-range voltage and energy result.
View Entry →Updated Dreamer 4 becomes first reported agent to obtain Minecraft diamonds using only offline training data
Migrated the Entry to schema v2 and clarified the offline training scope and authored task hierarchy.
View Entry →Updated NHC verification finds Google DeepMind’s GDMI leading individual hurricane guidance in 2025
Migrated the Entry to schema v2 and clarified the basin, lead-time, and availability limits.
View Entry →Updated METR finds frontier AI task-completion horizons doubling about every seven months
Migrated the Entry to schema v2, broadened the title to the full task set, and revised Evidence Strength to Strong.
View Entry →Methodology Change Methodology v2.0.0 Published
Published the Entry v2 judgment standard with simpler editorial fields, explicit assessment rationales, and version-aware immutable history.
View Methodology →
Added Researchers trace factual hallucination signals to small groups of model neurons
Initial publication of sparse factual-QA hallucination predictors and their cross-domain intervention limits.
View Entry →
Added One person reportedly used AI as a core tool in attacks on nine Mexican government organizations
Initial publication of Gambit's disputed report that one person used commercial AI as a core tool in attacks on nine Mexican government organizations.
View Entry →
Added Kosmos AI discovers an overlooked molecular clue to early neuron vulnerability in Alzheimer’s disease
Initial publication of Kosmos Discovery 7 and its cross-dataset flippase–phagocytosis signature.
View Entry →
Added GPT-5 literature search moves six Erdős database entries to solved status after finding earlier human results
Initial publication of six GPT-5-assisted Erdős Problems literature recoveries and their human-verified status changes.
View Entry →
Added Epoch estimates frontier AI benchmark progress nearly doubled in pace around April 2024
Initial publication of Epoch’s December 2025 ECI acceleration estimate with later cross-metric and Anthropic context.
View Entry →
Added Artificial neuron repeatedly fires within living-neuron voltage and energy ranges
Initial entry added from the peer-reviewed device study, source data, transparent peer review, prior bio-voltage and biointerface research, and later field context.
View Entry →
Updated Dreamer 4 becomes first reported agent to obtain Minecraft diamonds using only offline training data
Completed the evidence review and marked the Entry stable.
View Entry →
Added Dreamer 4 becomes first reported agent to obtain Minecraft diamonds using only offline training data
Initial entry added from the Dreamer 4 preprint, project artifacts, and prior Minecraft-agent baselines.
View Entry →Added METR finds frontier AI software-task horizons doubling about every seven months
Initial entry added from METR’s peer-reviewed task-horizon study, public analysis artifacts, Time Horizon 1.1, and the BRIDGE reproduction.
View Entry →Added NHC verification finds Google DeepMind’s GDMI leading individual hurricane guidance in 2025
Initial entry added. The claim is centred on formal 2025 NHC verification rather than DeepMind’s June 2025 launch announcement.
View Entry →Methodology Change Methodology v1.0.0 Published
Published the initial public judgment standard for Stage 1 entries.
View Methodology →