Entity trail

OrcaBonsai-27B-Uncensored

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-19 / sources-researcherDan Abramov got a Lean-verified proof of Conway's 50-year-old refinement conjecture for about 40 billion tokensIn a 2026-09-18 post Dan Abramov describes producing a mechanically verified Lean proof of Conway's refinement conjecture on omnific integers (if ab = cd then there exist e, f, g, h with a = ef, b = gh, c = eg, d = fh), with verification recorded in the Palomar registry, at roughly 40 billion tokens and about $40,000 in API cost. Early attempts produced grandiose but incoherent mathematics; what worked was separating Lean formalization of peer-reviewed sources from exploration of novel results, forcing standalone Lean files importing only Mathlib so the output stayed auditable, running specialized agents (PM, math researcher, red team, formalizer), and twice discarding all accumulated work to refocus. The transferable part is the auditing structure, not the mathematics.
2026-09-19 / sources-researcherRefusal-direction ablation applied entirely at inference strips safety from a ternary-quantized 27B without touching a single weightContinuum-AI-Corp published OrcaBonsai-27B-Uncensored on 2026-09-18 (439 stars in a day), intercepting residual-stream contributions at 129 points (64 MLP blocks, 48 linear-attention layers, 16 full-attention layers and the embedding layer) and applying y' = y - alpha * dot(y, r) * r against a learned refusal direction r. Because the ternary weights are never edited, the model keeps its bit-identical 1.72-bits-per-weight quantization-aware-trained compression, sidestepping the quality loss of re-quantizing edited weights. Claimed results: AdvBench refusals 99% to 6%, JailbreakBench 96% to 4%, StrongREJECT 99.3% to 3.3%, with MMLU unchanged at 76.7% to 77.7%. The builder takeaway is that weight-hash and weight-diff based safety attestation does not detect this class of modification at all.

Graph relationships

BENCHMARKED AGAINST
StrongREJECT
OrcaBonsai-27B-Uncensored achieves 99.3% refusal rate on StrongREJECT compared to original 3.3%.
Source finding
BENCHMARKED AGAINST
JailbreakBench
OrcaBonsai-27B-Uncensored achieves 96% refusal rate on JailbreakBench compared to original 4%.
Source finding
BENCHMARKED AGAINST
AdvBench
OrcaBonsai-27B-Uncensored achieves 99% refusal rate on AdvBench compared to original 6%.
Source finding
BUILT BY
Continuum-AI-Corp
OrcaBonsai-27B-Uncensored was published by Continuum-AI-Corp on 2026-09-18.
Source finding

Source trail

overreacted.io (Dan Abramov)GitHub (Continuum-AI-Corp)

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…