Hacker NewsLLM Architecture Gallery — 45+ Model Diagrams Surges to 443pts, 34cmtsSebastian Raschka·medium signalXBlueskyLinkedInCopy linkSebastian Raschka's LLM Architecture Gallery hit 443 points after posting March 15-16, covering architecture diagrams and fact sheets for 45+ models from 3B to 1T parameters. Covers dense models (Llama 3, Gemma 3, OLMo 3), sparse MoE (DeepSeek V3/V3.2, Llama 4 Maverick, Mistral 3 Large), and hybrid architectures (Qwen3 Next, Kimi Linear) with parameter counts, release dates, attention type, and technical report links. Physical high-resolution poster available.SourceSource pageSebastian Raschka↳ Follow the threadPolicy dependency / Stack layer'Rethinking Agent Security as a Networking Problem': stop asking the LLM to enforce its own policyarXivPolicy dependency / Stack layerBuild Integration, Not Fuzzing, Is What Kills LLM Dynamic Analysis on Real Autonomous-Vehicle StacksarXiv 2608.13450Policy dependency / Stack layerBest models score 70.4% on single-hop API calls but 2.4% on knowing when a tool-use policy makes a question unanswerablearXiv 2608.12282Stack layer / Threat patternFormal Specs Inferred From Tests Alone, Without White-Box Access to the ImplementationarXiv 2608.13240Policy dependency / Stack layerHinton concedes the open-weights fight at Ai4: 'I think that battle's been lost'TechCrunchPolicy dependency / Stack layerSimulator Collapse: RL Against a Single LLM User-Simulator Overfits, and Population Co-Training Recovers 14% of Held-Out SuccessarXiv 2608.12253Stack layer / Threat patternRow-Level Security Side Channels Turn Membership Tests Into Full Record Reconstruction in PostgreSQL and ElasticsearcharXiv 2608.11730Stack layer / Threat patternMicrosoft Agent Framework Python 1.14.0 adds Mistral, spins Durable Task out to its own repo, and blocks Windows junctions in skill discoveryGitHub (microsoft/agent-framework)