Quick Hits

• @percyliang announced that Marin 535B-A23B started training this week with "the whole process is open": 18.75T tokens across 11 GB200 NVL72 nodes for roughly three months (2.7e24 FLOPs), with post-training to follow. @0xSero calls it "the coolest thing I've seen all week."
• @b_nnett spotted that Cursor shipped Grok Bot 0.18.0 with runtime source maps enabled and reconstructed the source on GitHub. @LinearUncle reports the system prompts, tool definitions, and model routing are all readable, and says the newest build already strips the maps.
• @jun_song says Xiaomi unveiled local inference hardware running a 120B and a 3B model simultaneously on its new O3, O100, and D100 chips, and argues Chinese hardware makers are betting big on open weights and local inference.
• @dhh says @SpencerGBull's OmaPilot plugin, which runs a local Qwen 27B model, will ship in Omarchy 4.1 as the "fully agentic OS" play.
• @tobi implemented Cursor's "Git at Scale" blog architecture as a single open-source Rust binary over a weekend, pointed at any S3-style object store; @JorgeO claims it "scales better than GitHub," which is his characterization, not a benchmarked fact.

One Frontier Run Trains in the Open, Another Stays Deliberately Vague

The contrast in model news is stark. On one end, @percyliang laid out the full "voyage plan" for Marin 535B-A23B: 80% pretraining and 20% midtraining on 18.75T tokens, running about three months on 11 GB200 NVL72 nodes for 2.7e24 FLOPs. Before the hero run, the team trained a four-rung scaling ladder from 1.6B-A61M on 48B tokens up to 27.7B-A1.2B on 926B tokens, both to debug issues and to forecast the big run. He concedes it is "by far our biggest run" and they are "definitely expecting the unexpected." @0xero's bookmark note, from @0xSero, highlights the key point: the full training process is public and live.

On the other end, @martin_casado tweeted that he is "mind blown from a new model I just got access to" and thinks it may be "the most significant drop this year," with zero specifics. @mark_k speculates it could be a stronger LLM or even a world model, but that is pure guessing from two vague posts.

Grok Bot's Source Leaks Through Leftover Source Maps

@b_nnett found that "the Cursor team shipped Grok bot (0.18.0) with runtime source maps enabled" and was "surprised nobody noticed until now," posting a reconstructed version with downloads on GitHub. @LinearUncle, amplifying it in Chinese, calls this a rare chance to study a production agent's internals, with system prompts, tool definitions, and model routing logic all readable. His practical warning: move fast, because in his own testing the latest version no longer ships source maps. Two caveats worth holding onto: this reflects one shipped build, and "reconstructed" code should be read skeptically until you inspect it yourself. Still, for anyone studying how a commercial agent harness actually routes between models, it is an unusual window.

Local Inference Gets Dedicated Hardware and an OS Story

The local-compute thread ran through three posts. @op7418 reported (in Chinese) that Xiaomi announced a local AI host built on three new chips, the O3, O100, and D100, supporting dual 120B and 3B models with fast/slow system switching. @jun_song reads it as evidence that Chinese hardware makers see open weights and local inference as the future and are betting accordingly.

On the software side, @SpencerGBull released the first alpha of OmaPilot, an agent plugin for Omarchy Linux, and is recruiting testers to "iron out the Omarchy harness." @dhh enthusiastically confirmed it runs off a local Qwen 27B model and that a version ships in Omarchy 4.1. Rounding out the Omarchy ecosystem, @OmarchyMac shipped Omarchy Mac Quattro, adding Quickshell theming, Lua-based Hyprland config, full disk encryption, improved Apple Silicon support, Btrfs with Snapper snapshots, and faster one-step installation.

The Lean Agent Stack: Short Prompts, Skills, and a 72% Speedup

@xiaomovps published a profile of Pi Agent's creator, Mario Zechner (@badlogicgames): an Austrian game-framework veteran who built libGDX, sold RoboVM (later shut down by Microsoft), and wrote Pi in two nights in late 2025 as a personal tool with just four default tools (read, write, edit, bash) and a compressed system prompt. Per the profile, his own testing found CLI integrations often cheaper and more stable than MCP, and Pi was later acquired by Armin Ronacher's company Earendil, with Zechner keeping technical direction.

@trevin supplied the day's most actionable confession: his Nous Research Hermes browser tasks felt slow because his config was silently still using the legacy browser_* tools. After switching to the Browser Use 3.0 CLI with a persistent cloud/CDP browser, his local benchmark across a dozen tasks ran 72% faster, and he shares the exact prompt to audit your own setup.

Skills also featured twice. @steveruizok suggests a deceptively simple one: ask an agent to "describe the UX of my application completely," which he says surfaces inconsistencies, bugs, and design errors that never came up before. @FareaNFts lists 17 skills to install on a fresh Hermes setup, arguing "a raw Hermes install is smart. A skilled Hermes setup is unfair." And for running agents away from the desk, @kunchenguid shows SSHHIP 1.24's herdr integration, letting you navigate agent sessions from your phone via a command dial, because "mobile keyboard is the worst."

Reading List: Context Layers, Eval Granularity, and an RL Syllabus

@sachinrekhi argues the highest-leverage AI productivity move is not chasing model releases but building a "Compounding OS," or company brain: standardize on an agentic platform, build a shared org-wide skill library, and make all company context machine-legible. @prukalpa endorses it and describes speaking about Atlan's version of this at an AI engineer event.

@realmadhuguru's eval series reached part 7 with a Goldilocks principle: evaluate the intermediate jobs, not just the final answer. His financial-agent example shows stage-level scores (client understanding at 92%, data analysis at 70%) pointing you to exactly where to dig.

For skill-building, @AndrewYNg posted an AI Engineering Skills Map for building and deploying AI applications, which @hi_ZachParent likes as a starting point for FDE-style technical skillsets while noting it omits deciding what is worth building and scoping with clients. @cwolferesearch is publishing a complete guide to RL for LLMs, synthesizing the RLHF Book, Sutton and Barto, OpenAI's Spinning Up, Sebastian Raschka's from-scratch books, John Schulman's notes, Lilian Weng's policy gradient writeup, and more; @loganthorneloe calls the source list a goldmine for LLMs generally.

Lighter links: @Meaningness praises @jon_stokes' essay "Verifiable Domains Will Eat The World" as brilliant at several levels of meta. @vegardstikbakke recommends @hamiltonulmer's interactive post on serving fast dashboards with range requests and Hyparquet. @DeRonin_ calls @frederickjames' "Financial freedom from mobile apps" tutorial the all-in-one article for launching an app. And @pbakaus says one of his less-popular OSS projects gets more personal use than impeccable_ai, after @austriker27 asked who was using it; the project itself goes unnamed in the posts.

Practical Takeaway

The strongest actionable thread is @trevin's: stale agent configs silently tax you, and his move to Browser Use 3.0 CLI cut task time 72% in his own benchmarks. If you run Hermes or any agent with browser tools, paste his audit prompt into a session this week and verify which toolset your config actually loads, then re-benchmark. If your stack still feels heavy afterward, Zechner's CLI-over-MCP finding suggests testing a leaner integration against your real workload before adding another MCP server.