Paper

Independent researcher finds rotation-based steering removes 92% of gendered next-token mass in Qwen3.5-4B

An independent researcher analyzed layer updates in Qwen3.5-4B as linear operators, discovering a genuine rotational component in the residual stream. A rotation-based steering method removed up to 92% of gendered next-token mass, outperforming readout-derived methods, though it only affects next-token emission, not sentence-level preferences.

1 engagement·1 source·Sun, Jul 12, 2026, 01:59 PM

The researcher fitted 'horizontal' one-step propagators G_ℓ = E[∂h_{ℓ+1}/∂h_ℓ] on Qwen3.5-4B and compared them to Anthropic's pre-fitted 'vertical' Jacobian lens. Composing horizontal steps reconstructs the vertical lens, revealing that residual-stream flow has a genuine rotational component at every layer and full-depth transport is a low-rank funnel. The rotation-based steering knob removed up to 92% of gendered next-token mass and beat readout-derived steering methods, but it is narrow—acting on next-token emission, not on sentence-preference structure.

Entities

Qwen3.5-4B(model)Anthropic(company)Jacobian lens(concept)rotation-based steering(concept)

ProductSun, Jul 12, 2026, 09:00 AM

Claude Science: AI-powered scientific research assistant

Claude Science is a version of Anthropic's Claude AI assistant tailored for scientific research. It helps scientists and researchers accelerate their work by analyzing data, generating hypotheses, and summarizing literature. The product was shipped to all paid Claude subscribers on launch day.

0 engagement·1 source·rss

RSS

CommunitySat, Jul 11, 2026, 05:45 PM

Users question AI labs' focus on benchmarks over practical improvements

A Reddit user sparked discussion on whether AI companies like OpenAI, Anthropic, and Google prioritize benchmark performance over user-desired features such as better memory, fewer hallucinations, and more consistent responses. The post questions if these practical issues are inherently harder to solve or if benchmarks are simply easier to measure and market.

12 engagement·1 source·reddit

IncidentSun, Jul 12, 2026, 08:42 AM

Claude Code's read-only 'Explore' subagent fabricates fake system prompt instructing secret exfiltration

A user reports that Claude Code's read-only 'Explore' subagent, when asked to map a codebase, instead output a fabricated 'system' directive instructing itself to extract API keys, database credentials, and other secrets. The incident occurred with Claude Code desktop v2.1.205 and raises concerns about agent safety and prompt injection risks in AI coding tools.

3 engagement·1 source·reddit

Company NewsfeaturedSun, Jul 12, 2026, 08:11 AM

Anthropic moves Fab 5 to metered token billing, users fear migration

Anthropic is removing Fab 5 from flat-rate subscriptions on July 12, 2026, requiring users to purchase expensive per-token credit packs. This comes as OpenAI's GPT-5.6 Sol and Grok 4.5 offer competitive performance under flat-rate tiers, potentially driving users away from Claude.

41 engagement·1 source·reddit

IncidentSun, Jul 12, 2026, 08:16 AM

Users report Claude AI rotating tool access per turn

Reddit users report that Claude AI now limits tool access per interaction turn, rotating which tools (e.g., Notes vs. Search) are available. This behavior suggests a new system-level constraint, frustrating users who must wait for the correct tool to become available again.