Source Audit: KTransformers and Heterogeneous CPU-GPU MoE Inference
Auditing kvcache-ai/ktransformers: dissecting CPU-GPU offloading, AMX/AVX-512 kernels, and how heterogeneous memory architectures run 671B MoE models.
Fodev JEO
LATEST INVESTIGATIONS
Source audits & field notes
Source Audit: Memori and the Trade-Offs of SQL-Native Agent Memory
Auditing MemoriLabs/Memori: analyzing relational schema interception, ACID durability, and why standard SQL databases challenge dedicated vector stores.
Source Audit: Sprite Fusion Pixel Snapper and Deterministic AI Pixel Cleanup
Auditing Hugo Duprez’s Rust/Wasm Pixel Snapper: why deterministic grid snapping and color quantization solve diffusion pixel art flaws where LLMs fail.
Source Audit: Anthropic 100k Study and Multi-Agent Productivity Boundaries
Auditing Anthropic’s 100,000-conversation productivity report against METR randomized developer trials and 2026 multi-agent benchmarks under equal token budgets.
Source Audit: Fara-7B Synthetic Trajectory Engine and Critical-Point Halting
An empirical audit of Microsoft Fara-7B and FaraGen: why synthetic trajectories and critical points matter for on-device computer-use agents.
Gemini Native Visual Generation: Architecture and Typography in Google Image Models
A source audit of Google DeepMind’s native multimodal image models, evaluating unified visual tokenization, typographic precision, and diffusion transformer scaling.
Source Audit: Andrej Karpathy LLM Council and Deliberation Mechanics
An architectural audit of karpathy/llm-council: examining 3-stage anonymized peer review, Borda count aggregation, and synthesis bottlenecks.
Structured Outputs and Schema-Constrained Prompts in Multi-Agent Workflows
Source audit of structured input schemas, JSON mode constraints, and prompt orchestration across multi-agent pipelines with Gemini and open LLM APIs.
Graph-Vector Hybrid Memory for Agents: Beyond Naive Vector RAG
Source audit of hybrid graph-vector memory architectures for autonomous agents, evaluating multi-hop relationship traversal and persistent episodic storage.
Engineering in the Multi-Agent Era: Invariant Testing and Harness Design
Analysis of evolving engineering skills when transitioning from writing code syntax to designing agent harnesses, interface contracts, and invariant verification.
인기 태그
Search article titles, categories, and tags. Full text is searched when needed.
Type to search articles.