Summary
Everything today landed in the morning Following-feed capture, which covers the late US Sunday. The afternoon and evening slots were empty, and there were no curated reposts or saved posts. The strongest cluster is compute supply: Musk says debt markets have "almost entirely shut off" for AI-only neoclouds, Rohan Paul revives Nadella's "warm shells" warning that power-ready buildings, not chips, are the bottleneck, and Zitron and Marcus pile on. It matches DeepInfra's coming rent increase, so this is real signal, not just doom posting. A second cluster on agent context and memory (Graphify's codebase graph, a five-layer memory thread) is useful framing but rests on vendor numbers. The single-slot standouts are Stanford's open Marin 535B-A23B MoE passing 80% of pretraining and NVIDIA's AW-0 robot model that plans in code. Bait prompts, agent-trading posts and politics are noise.
Posts
- Neocloud credit and power (cluster of 4) (@StockSavvyShay, @rohanpaul_ai, @edzitron, @GaryMarcus) [morning]. Musk predicts "guppy-level" consolidation among 200+ neoclouds, while investors reply that lending is selective, not gone. Nadella's power-shell shortage and OpenAI's 100 GW-a-year ask frame it. See GPU rents.
- Agent context and memory (cluster of 3) (@techNmak · Graphify, @AnnatarXBT, @omarsar0) [morning]. Graphify maps a codebase into a queryable graph and claims 80% fewer agent tokens, a vendor number. The memory thread quotes Mem0 at 1,800 vs 26,000 tokens per query with no paper; Saravia argues app surfaces now beat the CLI for multi-agent work. See agent memory.
- Marin 535B-A23B (@stanfordnlp) [morning]. Fully transparent MoE (23B active per token) passed 80% of pretraining. Largest open-process run tracked here.
- AW-0 agentic world action models (@askalphaxiv · paper) [morning]. Code, video and action experts in one mixture-of-transformers. 50.7% on RoboTwin 2.0 vs 41.3% for π0.5, and 71.1% vs 46.0% on real tasks.
- Consistency is not validity (@omarsar0 · paper) [morning]. DeepMind self-judged moral-argument RL makes the model decisive and order-stable. User pushback still moves its answers.
- GitHub secret detection (@github) [morning]. In-house classifier lifts flagged-secret acceptance 11.5% and cuts median scan time 86%.
- Open weights and bad actors (@bengoertzel · essay) [morning]. Argues the real choice is approved closed systems or open systems with defenses built around them. Also posted 2033 GDP bets against Marcus.
- Next-Concept Modeling (@sedielem) [morning]. Coarse-to-fine parallel text generation. Thread only, no detail captured.
- Safety Mid-Training (@stanfordnlp) [morning]. Moves safety data into mid-training. Thread only.
- System One decision models (@TheTuringPost) [morning]. List of 15 models beyond Jev. Light.
- AI model welfare (@mmitchell_ai) [morning]. IEEE Spectrum piece. Commentary.
- Bait and promo [morning]. Kimi "founder prompt," "6 Claude agents," agent trading, link dumps, stock picks, political posts. Skip.
Full day context: daily digest.