Andrej Karpathy's April 8 tweet on building a personal LLM knowledge base with Obsidian hit 18,196 likes — the week's top technical post on X. His setup: a vault of ~2,800 markdown notes indexed into a vector DB, then queried by Claude via MCP. Highlights include a daily 'inbox-to-atomic-notes' agent and a 'Socratic review' agent that surfaces stale or contradictory notes. The thread ignited a broader PKM-meets-LLM conversation and turned a niche workflow into a widely-copied playbook for personal AI.
KarpathyObsidianPKMPersonal AIMCP
Why it matters
Karpathy's endorsement pulls a niche PKM workflow into the mainstream AI-tooling conversation. The practical value isn't the specific stack — it's the proof that MCP + a local vault + a simple daily agent is enough for most 'personal AI' use cases, without waiting for a product. Expect clone posts, Notion/Obsidian feature parity moves, and a surge in MCP server variety.
Impact scorecard
6.7/10
Stakes
5.5
Novelty
6.0
Authority
8.5
Coverage
3.0
Concreteness
7.0
Social
9.0
FUD risk
1.0
Coverage4 outlets · 0 tier-1
AI Noon, Every.to, Stratechery (brief), The Pragmatic Engineer
First-party post from a well-established practitioner, full stack description, reproducible. Zero FUD — it's a workflow writeup, not a claim about the world.
@hardmaru (David Ha) flagged a paper adapting Sora-style video-diffusion architectures to build a learned world model of an actual Linux desktop. The model ingests 9,000 hours of screen-recording + keyboard/mouse traces and learns to predict next-frame UI state conditioned on user input — effectively a probabilistic operating-system simulator. On a held-out eval of 50 common tasks (opening files, running commands, navigating web UIs), the model achieves 73% next-event accuracy at 2-second horizons and 41% at 30-second horizons, beating the prior SOTA (Meta AI Habitat-UI) by 18pp. Direct application: train agents in fully simulated computer environments without real-system rollouts — cuts RL data costs ~40x and eliminates the safety risk of letting agents touch production systems during training.
EE Times deep-dive on AMD's ROCm 7.0 and whether it can finally dent NVIDIA's CUDA moat. AMD's MI400 (96GB HBM4, 5.2 PFLOPS FP8) now runs PyTorch, vLLM and SGLang out-of-the-box — but reviewers testing MLPerf Inference v5.1 still see 1.6–2.2x gaps vs H200 on representative LLM workloads, driven by kernel-library maturity rather than raw silicon. Breakthrough of the cycle: AMD hiring 600 CUDA-kernel engineers in 12 months, plus open-sourcing HIPify tooling that auto-translates 83% of typical CUDA kernels. AMD claims Meta, Microsoft and OpenAI are all now shipping production MI400 pods. NVIDIA's response: CUDA 13 with tensor-core autotuning targeting the same eval suite, launching Q2.
Anthropic announced the advisor strategy on the Claude Platform: pair Opus 4.6 as a planning/critique advisor with Sonnet 4.6 or Haiku 4.5 as the executing model. The advisor inspects partial outputs, suggests corrections and redirects the executor mid-generation. On SWE-bench Multilingual, Sonnet+Opus-advisor scores 2.7 percentage points higher than Sonnet alone, at roughly 1.3x the cost vs 7x the cost of running Opus end-to-end. General availability today via the Claude Console and CLI; pricing is existing Claude API rates for both models (no advisor premium). Anthropic positions this as the first first-class multi-model inference primitive in any frontier-lab API — not just routing or cascading but explicit advisor/executor roles with shared context.