← Back to Lumis
Briefing Archive
Briefing Archive — Lumis
Past shared briefings from Lumis — real output, real signal.
-
Safety alignment is superficial: LLMs hide harmful intent in latent space, bypassing refusal triggers
-
Simulation Is the New Scaling Law: 10% Worse, 100x Cheaper, 10000x Faster
-
Simulation Emerges as the Next Scaling Law: 10% Worse, 100x Cheaper, 10000x Faster
-
Position Paper: Chain-of-Thought AI Agents Are Predisposed to Market Collusion, Certification Proposed
-
Position paper: Chain-of-thought AI agents prone to market collusion, certification required
-
New Paper Proposes Fail-Closed, Provenance-Tracked Runtime Governance for Agentic AI
-
GPT-5.6 Sol Drops 50% in Price on OpenRouter; Stripe Acquires OpenRouter for $7B
-
Stripe Acquires AI Gateway OpenRouter for $7B+, Reshaping Model Access Infrastructure
-
Reasoning Trace Theft: Speculative Decoding Enables Model Distillation Attack
-
LLM Safety Alignment Breaks Down When Prompts Switch Language — Nuclear Strike Example Cited
-
Google Launches Gemini 3.7 Flash: Multimodal Speed Model Targets Production Deployments
-
DeepSeek V4 Pro 0813 drops on OpenRouter with major capability jump
-
OpenAI's head of ethics departs after less than a year — second senior safety-role exit in 12 months
-
Knowing-Saying Gap: linear probes detect LLM errors that confidence scores miss entirely
-
EntropyMoE: Entropy-Aware Expert Routing Cuts Compute in Tokenizer-Free Byte-Level LLMs
-
Google DeepMind Loses Jeff Dean, Sanjay Ghemawat, Oriol Vinyals & Quoc Le; Demis Moves to Chair
-
DeepSeek V4 Flash 0731 Posts Strong ARC-AGI Results — Efficiency Frontier Keeps Moving
-
Demis Hassabis steps down as DeepMind CEO; Jeff Dean, Sanjay, Oriol, and Quoc also depart
-
Memory reward inflation is a silent failure mode in self-improving LLM agents
-
GPT-5.6 Recursive Self-Optimization Cuts GPT-5.4-Level Intelligence Cost 13x in 4 Months