social-stream · 2026-11-15

2026-11-15-morning

Summary

The morning slot captured no reposts and no tweets from tracked handles, so this synthesis reads the morning Following-feed capture, which covers the US Saturday afternoon and evening. The strongest technical signal is a cluster of 3 harness and memory posts: Harness² (agents that improve their own harness by test-running edits), MemCredit (memory picked by its effect on the next action), and a Meta engineer's essay on building agents that work. Efficiency got two strong posts. Meta's e-means gets k-means-quality clusters 10-45x faster on GPUs, and a Chinese-language explainer of NVIDIA's looped Dory1.0-4B kept spreading. The most useful precision post is Sasha Rush's group showing that rounding breaks operators in PyTorch, JAX and diffusers. On epistemics, Meta's "Against All Likelihood" essay argues that truth should mean invariance across contexts, not likelihood. The Nvidia story kept running, led by Gary Marcus amplifying a TechPowerUp report that Nvidia engineered scarcity by funding GPU buyers. Starship's tower catch dominated the off-topic feed. No new saved posts were captured this run.

Posts

  • Harness and memory engineering (cluster of 3) (@rohanpaul_ai on Harness², paper, @omarsar0 on MemCredit, paper, @taha_rafiq). Harness² (Northwestern and Google) lets an agent's harness (its prompts, tools, skills and memory) improve on new tasks with no answer key. An editor model proposes fixes, the task reruns, and only edits that change behavior on real runs are kept, with lessons stored in a library for the next task. One improvement step raised GLM-5's WorkBuddy-Bench score by 11.5 points, past Gemini 3.5 Flash on its stock harness. Editing without those test runs dropped coding scores from 73.4 to 63.2. MemCredit measures each memory by removing it and checking whether the agent's next action changes. Most memories change nothing, and the learned selector reaches 86.6% on ALFWorld against 65.7% for BM25, with 74.5% fewer memory tokens. Taha Rafiq's X article distills years of building agents at Meta into four rules: build verification first, read agent traces, don't over-engineer, and build feedback loops into the product. See MemCredit and agent harness engineering.

  • e-means: k-means rebuilt for GPUs (cluster of 2) (@akshay_pachaar, @TheTuringPost, paper). Classic k-means has to scan the whole dataset before it moves any cluster center. Meta's e-means moves each center by an exponential moving average over mini-batches. It splits crowded clusters to revive dead ones and sets batch and step sizes from scaling laws, so there is nothing to tune. It reaches 25 epochs' worth of k-means loss in about 2 epochs, 10-45x faster on GPUs. It matters for vector quantization, retrieval indexes and data deduplication.

  • Dory1.0-4B explainer (@noahatwhere, paper). A Chinese-language walkthrough of NVIDIA's looped transformer, a model that reuses one middle block of layers several times so depth grows without new weights. The middle block runs 1-4 times, giving 72 to 192 effective layers from 4B parameters, and the input block's hidden state is re-injected with L2 normalization before each loop. SWE-bench Verified goes from 50.3% at one loop to 61.8% at four, and four loops write 1.3-2.5x fewer output tokens on agentic tasks. See looped transformers.

  • When Rounding Breaks Machine Learning (@ChrSzegedy RT of @srush_nlp, post). Robert Joseph George and Sasha Rush use FloatLib, a Lean library that models floating-point arithmetic at any precision and proves error bounds. They swept 154 operators across seven libraries and about 780 million outputs. It started with a softmax whose float32 target gradient came out exactly zero even though the true value was representable. The sweep found bugs in PyTorch, JAX, TensorFlow Probability, diffusers, PyTorch Geometric and xLSTM, several never reported. See the wiki summary.

  • Against All Likelihood (cluster of 2) (@dair_ai, @rohanpaul_ai, paper). A Meta position essay. Models trained by maximum likelihood learn what is common, so repeated falsehoods become probable and rare truths get discounted. It proposes treating a claim as true if it survives relevant changes of viewpoint. In practice, the model answers as several personas (say, three medical specialists), keeps what all answers share, and abstains if nothing is shared. Its sharpest worked example: on one binary case, a loss of 0.67 means 100% accuracy while a nearly equal 0.71 means 0%. No experiments. See the wiki summary.

  • Weight-Space Reading (@crazyMLguy, repo). Can an LLM predict what a weight edit will do to another model without running it? Qwen 27B with direct weight access predicted subtle edit effects at 73% balanced accuracy, against 57.8% from text descriptions. On Muse-30B, 34 edits and 6 no-op controls produced no detectable change, yet agents often predicted one. So false positives are the open problem. Models reading their own weights showed no clear advantage.

  • Decision models as teachers (@rohanpaul_ai, paper). A recap of yesterday's "Actor, Judge, or Teacher?". JEV completes only 42.9% of ALFWorld tasks on its own, but grading another agent's moves after the run raised that agent from 72.8% to 82.9%. See the wiki summary.

  • Nvidia: scarcity allegation and the Huang story (cluster of 3) (@GaryMarcus RT of @TechPowerUp, @rohanpaul_ai on the petition, @dylan522p). TechPowerUp reports allegations that Nvidia engineered GPU scarcity by funding AI startups that then bought its GPUs. The article body was behind a bot check, so only the headline is captured. The push to reinstate Huang is gaining, with standwithjensen.com circulating, and Dylan Patel posted a one-line jab that the whole semiconductor supply chain, neocloud complex and open-source ecosystem is tangled up in it. See Open-source AI without Jensen.

  • Long reasoning as the investment theme (@rohanpaul_ai). Eric Schmidt says that if he could invest in one AI theme it would be long reasoning: a model that stays on one problem for hours or days is a different product from one that answers and stops.

  • Interpolation versus search (@fchollet). François Chollet: a curve fitted by SGD can only interpolate, but once it guides a search process the system is doing search, and theorem proving and reasoning-chain generation are search.

  • Agent swarms explainer (@undefinedKi). A primer riding a Sam Altman quote on agent teams: one goal, a shared task board so agents never duplicate work, and per-agent memory, tools and workspace. Generic, but it is the same inbox-and-worker shape as Anthropic's always-on guide.

  • Off-topic or skip. Starship's tower catch (Musk, Hassabis and others celebrating), a "full 3D studio in 20 repos" list thread, an Obsidian "second brain" repo list, an agents-post-to-X promo, and Grok Bot prompt bait.