XAI, Sentience & Safety

New XAI Papers

• 🔥 Interpretable MIL for Hematologic Diagnosis: CAREMIL framework with DeepHeme encoder achieves AUROCs of 0.999 for AML, 0.891...

Hot take from @mattshumer_: Even OpenAI is dramatically underestimating how much inference compute will be needed in coming years.

New paper Towards Long-horizon Agentic Multimodal Search targets advances in agentic multimodal search for long-horizon planning, pushing benchmarks toward real-world multimodal scenarios. Join the discussion.

Emerging XAI tools turn foundation models into explainable medical predictors.

• Genomics: Evo 2 embeddings enable SOTA variant pathogenicity...

• Deprecation alert: Anthropic is forcing users off reliable claude-sonnet-4-5-20250929 to claude-sonnet-4-6.
• No fixed versions: Model page lists...

Divergent AI safety paths emerge:

• Anthropic's Claude outperforms humans 4x on weak-to-strong supervision, hitting 97% PGR in 5 days vs. humans' 23%...

Mythos shines in multistep attacks but faces quick counters and its own flaws:

• Infiltration feats: First model to fully solve AISI's 32-step "The...

• AI auditing is hotter than Artemis II's shield upon reentry, per @KevinTFrazier quote.
• Miles Brundage (@AVERIorg) enjoyed deep dive on auditing with @ARozenshtein.
• Listen to their @scaling_laws episode for scaling AI governance insights.

Deaf-led advocacy advances equitable AI safety: CoSET's SAFE AI Task Force toolkit evaluates automated sign language tools for quality, safety, and...

Post-hoc XAI methods—feature attribution, counterfactuals, and natural language rationales—enhance trust by helping users understand hate speech moderation decisions. These complement ante-hoc approaches in stakeholder-focused explainability.

A new paper repositions XAI by arguing that justifiability—exploring several ways algorithms can be justifiable—puts mere explainability in its place as the superior goal for ethical AI.

Explainable AI (XAI) is positioned as both a technology and a type of law in financial advisory, based on foundational principles.

Anthropic's Long-Term Benefit Trust has appointed Vas Narasimhan, CEO of Novartis with over two decades in medicine and global health, to its Board of Directors. Key move to infuse biomed expertise into AI safety governance.

New XAI Papers

• 🔥 Attention Sink in Transformers: Survey covers utilization, interpretation, and mitigation of attention sinks in...

Cloud credit cycles expose a hidden financial loop: Microsoft invests $13B in OpenAI mostly as Azure credits; OpenAI trains models, boosting...

Rising interpretability techniques target model internals:

• Entropy probing reveals divergent info patterns in unified multimodal models
• SHAP...

AI tools accelerate literature search, experimentation, and drafting, sparking a structural submission explosion at AI conferences like NeurIPS/ICLR. Position paper urges: Let papers flow.

AI chatbots misdiagnose over 80% of early medical cases, per a study reposted by Gary Marcus from @FT. Stark evidence of LLM limitations in critical applications demands urgent advances in explainability for AI safety.

Model Reliability Concerns

• 🔥 Claude Opus 4.6 Regression: Claude Opus 4.6 accuracy on BridgeBench hallucination test drops from 83% to 68%,...

• Neural Computers (NCs) unify computation, memory, I/O in learned models, acting as the runtime itself—not atop existing stacks.
• Prototypes shine:...