Summary
No curated reposts or saved posts came in this morning. These posts come from the morning Following-feed capture, which covers the late US Sunday. The strongest signal is a compute-supply cluster of 4 posts: Musk says debt markets have "almost entirely shut off" for AI-only neoclouds, Rohan Paul revives Nadella's warning that Microsoft lacks finished buildings near power, and Ed Zitron and Gary Marcus pile on that "most of the neoclouds are screwed." A second cluster of 3 posts is about agent context and memory: Graphify's codebase knowledge graph, a five-layer agent-memory thread, and an argument that app-style agent interfaces now beat the CLI. Open research surfaced through reposts: Stanford's open Marin 535B-A23B MoE passed 80% of pretraining, NVIDIA's AW-0 robot model plans in code, and a DeepMind paper finds moral-argument training buys consistency, not validity. GitHub shipped a secret-detection model with an 86% faster median scan. Hype and politics ("send this prompt to Kimi," agent trading, election posts) are skipped.
Posts
Neocloud credit and power (cluster of 4) (@StockSavvyShay, @rohanpaul_ai, @edzitron, @GaryMarcus). Musk says debt markets have "almost entirely shut off" for neoclouds that only rent AI compute and expects "guppy-level death and consolidation" among the 200+ operators. The investor reply: lending is not gone, it is selective, favoring contracted revenue, customer prepayments and asset-backed deals. Rohan Paul pairs this with Nadella's line that the shortage is "warm shells" (finished, utility-ready buildings near large power feeds), not chips, and with OpenAI's push for the US to add 100 GW of generation a year; GPUs depreciate while waiting for substations. Zitron says that if two operators' accounts are right "CoreWeave is toast" outside a Blackstone rescue, and Marcus agrees most neoclouds are in trouble. This lines up with DeepInfra's Nov 9 rent increase covered in the digest and the GPU-rent summary.
Agent context and memory (cluster of 3) (@techNmak with Graphify, @AnnatarXBT, @omarsar0). Graphify maps a codebase once into a queryable knowledge graph (code parsed locally with tree-sitter, docs and media through the assistant's model), so an agent asks "what connects authentication to the database" instead of grepping a dozen files; its README claims an 80% cut in agent token spend, a vendor number. The memory thread lists five layers (working, episodic, semantic, procedural, forgetting) and quotes Mem0 at 1,800 tokens per query instead of 26,000; no paper is linked, so treat the figures as marketing. Elvis Saravia, a heavy CLI user, argues that app surfaces like Grok bot and the Codex app now scale multi-agent work better than the terminal. See agent memory.
Marin 535B-A23B (@stanfordnlp). Percy Liang and David Hall's fully transparent MoE (535B total, 23B active per token) passed 80% of pretraining. It is the largest open-process pretraining run tracked here.
AW-0, agentic world action models (@askalphaxiv, paper). NVIDIA, Berkeley and UPenn generate three coupled streams in one mixture-of-transformers: a code expert writes parameterized skill routines, while video and action experts predict future frames and motor commands. It reaches 50.7% average success on RoboTwin 2.0 against 41.3% for π0.5, and 71.1% vs 46.0% across eight real manipulation tasks, including towel folding and a Rubik's Cube.
Consistency is not validity (@omarsar0, paper). A Google DeepMind policy is trained with RL in a self-judged tournament over logic-checked dilemmas from classic novels, rating pairs of its own arguments. It becomes more decisive and stops flipping when options are reordered on MoralChoice, ELEPHANT and MORABLES, but user pushback and reframing still move its answers.
GitHub secret detection (@github). GitHub's own classifier raised acceptance of flagged secrets by 11.5% and cut median scan time by 86%.
Open weights and bad actors (@bengoertzel, essay). Goertzel argues that blocking bad actors from truly open weights cannot work, so the real choice is closed systems from approved providers or open systems with defense built around them. He also posted three 2033 GDP bets against Gary Marcus.
Other reposts. Next-Concept Modeling, a coarse-to-fine parallel text generator (@sedielem); Safety Mid-Training, which moves safety data into mid-training (@stanfordnlp); a list of 15 "System One" decision models beyond Jev (@TheTuringPost); IEEE Spectrum on AI model welfare (@mmitchell_ai). Thread only, no paper detail captured.
Skip. Kimi "founder prompt" bait, "6 Claude agents" and agent-trading posts, reverse-engineering link dumps, stock picks, and political posts.