Summary
The morning slot's strongest signal is a single substantive research post that doubles as today's digest lead: Hugging Face's Elie Bakouch amplifying the Marin / Open Athena pretraining ablation, an open dense-to-mixture-of-experts (MoE, where each token routes through a small subset of expert sub-networks) recipe reporting a roughly 6x cumulative learning speedup with the theoretical-vs-realized gap shown honestly. Around it sits an open-weight efficiency cluster from the Kilo Code account: StepFun's Step 3.7 Flash (a 198B sparse MoE with about 11B active, 256K context, Apache 2.0) and Alibaba's Qwen 3.7 Plus (a 1M-context multimodal GUI agent), both pitched as free-to-try, plus a sharp practitioner warning that you should verify a model's identity at the API layer rather than trusting its self-report. Anthropic's developer account contributed two notes: a data-analytics-agent best-practices post, and a naming change that renames the Claude Code dynamic-workflow trigger from "workflow" to "ultracode." The three curated @bayesiansapien reposts (Fei-Fei Li, Ashwin Gopinath, and the mem0 memory-layer account) all point at X native long-form articles whose bodies did not fetch this slot, so they are read-throughs rather than summarizable here. The rest of the feed is low research signal: tinygrad's full-stack speed pitch, a Google Research open-source flood-forecasting framework, a creative-tools and AI-hardware run from Robert Scoble at Upscale Conf, and a heavy stream of space-valuation and political posting from one account that carries no AI substance.
Posts
Marin open MoE pretraining recipe, ~6x learning speedup (@eliebakouch, Open Athena blog). Bakouch calls it "scaling laws are beautiful part 2," a clean pretraining ablation similar to the MAI scaling ladder and fully open. The quote-tweeted Marin post (Larry Dial) lays out the ledger: moving from a dense baseline to Marin MoE V1 gives 6.7x theoretical / 3.6x realized speedup (realized accounts for Model FLOPs Utilization), then raising experts from 64 to 256 adds 1.4x, swapping the optimizer AdamH to MuonH adds 1.3x, partial key offset adds 1.2x, and routed-expert normalization adds 1.04x. This is the empirical, recipe-level companion to today's scaling-parametrization papers. See Marin MoE summary and the daily digest.
Open-weight efficiency models free in Kilo, plus an identity-verification warning (cluster of 3, @kilocode). Step 3.7 Flash is highlighted first: a 198B sparse MoE with about 11B active per token, 256K context window, selectable reasoning tiers, Apache 2.0, pitched as big-model reasoning at long-context speed. Qwen 3.7 Plus is offered free for a few days: Alibaba's multimodal agent model that reads screens, navigates GUIs, generates code from visual references, and works across GUI and CLI with a 1M-token context. The third post is the useful one for builders: never trust a model's self-report for identity or routing, because asking "who made you" tests the training data, not the deployment, so verify at the API layer rather than by asking nicely.
Anthropic on data-analytics agents, and a workflow trigger rename (cluster of 2, @ClaudeDevs, blog). The first post shares best practices for building agents that perform data analysis: skills, data foundations, and evaluations. The second is an operational change to Claude Code's dynamic-workflows research preview: the trigger word moves from "workflow" to "ultracode," so Claude does not spin up a fleet of coordinated subagents when the user merely mentions the word "workflow" in another sense.
Three curated reposts pointing at X long-form articles (group, bodies unfetched this slot). @drfeifei (Fei-Fei Li), @ashwingop, and @mem0ai were each retweeted by @bayesiansapien and each link to an
x.com/i/article/...native article whose text did not download this slot. Treat as curated read-throughs: Fei-Fei Li (spatial-intelligence / world-models lead), Ashwin Gopinath, and the mem0 memory-layer account are all high-signal handles worth opening directly.tinygrad's full-stack speed pitch (@tinygrad). Because tinygrad spans the full stack from tensors down to memory-mapped IO, the team argues its speed ceiling is higher than any other framework, riffing off a contributor grinding AMD inference faster via tinygrad as a way to learn GPU programming. A GPU-optimization-adjacent claim, no benchmark attached.
Google Research open-sources its hydrology framework (@GoogleResearch, blog). An open-source release of Google's flood-forecasting / hydrology modeling framework so agencies worldwide can integrate advanced flood forecasting into local workflows. Applied-science release, outside the core efficiency and routing focus.
Scoble at Upscale Conf: creative-AI tools and AI hardware (cluster of 8, @Scobleizer). Highlights: reve 2.0 text-to-image-and-editing claiming it can beat nano banana 2 despite far less funding; Magnific Agents, a new platform for AI filmmaking that builds assets and organizes projects; eldr.one AI glasses aimed at senior care and fall detection; praise for Town AI; and the 1x Neo humanoid robot's design story. Tier-3/4 creative and hardware signal, noted for completeness, not central to the research thread.
Space-valuation and political posting, low AI signal (@brivael). A long stream arguing SpaceX will be worth tens of trillions and that models will become commodities so the value moves to infrastructure and products built on top. No concrete AI research content; logged as noise for this slot.