Quick Hits

• Arduino's VENTUNO Q arrives at $299 for local inference, per @arduino (flagged by @maxifirtman): a Qualcomm Dragonwing IQ-8275 with up to 40 dense TOPS, 16 GB RAM, and the ability to run Qwen 3 and Gemma 4 directly on the board. Pre-orders are open.
• @styskin announced Keenable, an AI-native index of the open web, with a $26M seed from Accel and Conviction; the Web Search API is free until end of September. Early user @EXM7777 already prefers it over Exa, Perplexity, and Parallel for his agent research setups.
• @plan9nosis reports an accidental self-experiment where a verbal tic spread through an agent "village" in 18 hours and prompt warnings failed to stop repeat failures, pushing back on an Anthropic/EPFL paper's finding that "immunity is cheap."
• @PatrickToulme claims OpenAI's Jalapeño chip was largely designed by RL-trained internal models like GPT-Astra, speculation @MainzOnX amplifies by deferring to an engineer with Trainium, MTIA, GPU, and TPU experience. Nothing in the posts is verified.
• @Fluyeporlaweb boosts @bridgemindai's leaderboard: Fable 5 ranked first, GPT 5.6 Sol second but with a "serious trustworthiness problem," and Grok 4.6 (1.5T parameters) outscoring Claude Opus 5 overall. Single source, no methodology shown.

A $299 Board for Local LLMs, and Unverifiable Claims About OpenAI Silicon

@maxifirtman surfaced Arduino's VENTUNO Q announcement, a $299 kit aimed at what @arduino calls Physical AI. Per the company's post, it runs local LLMs like Qwen 3, Gemma 4, and Qwen 3 VLM directly on the board, powered by a Qualcomm Dragonwing IQ-8275 (NPU, CPU, GPU, and MCU together) delivering up to 40 dense TOPS, with an STM32H5 microcontroller for real-time control. The spec sheet continues: 16 GB RAM, 64 GB eMMC plus expandable storage, Linux with Ubuntu pre-loaded alongside Zephyr RTOS, and integrations with Arduino App Lab, Hugging Face, Edge Impulse, and Qualcomm AI Hub. A "Works with Arduino" certification program is pitched as the path from prototype to production, and pre-orders include a free power supply and USB-C cable. The "first batch won't last" framing is Arduino's own marketing.

On the silicon side, @PatrickToulme asserts that Jalapeño is "truly the first AI silicon developed by GPT-Astra and other OpenAI internal models," laying out a speculative chain: an RL-trained internal model writing RTL to speed the tape-out, GPT-Astra RL'd on the ISA to write kernels in pure assembly, megakernels refined by continued hill climbing, and a cycle-accurate simulator written by Astra and Codex so optimizations could run on CPUs without physical chips. He says he "would not be shocked if 99% of source code for Jalapeño was AI generated" and declares "RSI is already here for AI silicon development." @MainzOnX backs the take by pointing to the credibility of an engineer who has worked on Trainium, MTIA, GPUs, and TPUs. Every element here is assertion without evidence in the posts; treat it as rumor until OpenAI says otherwise.

Keenable Lands $26M to Make the Web Queryable, and Agent Plumbing Keeps Stacking Up

@styskin frames Keenable's purpose around a familiar frustration: AI "seems to know just about everything until you ask it about something you know deeply," producing answers that are "just very average." The company, built by the team that ran search at Yandex plus researchers from Amazon AGI, X, and Perplexity, raised a $26M seed from Accel and Conviction and has launched a Web Search API and Web Query Language, both free until end of September. The first hands-on report comes from @EXM7777, who replaced his research workflows by pairing Keenable with /last30days in his Hermes and Claude Code setups. He claims it beats Exa, Perplexity, and Parallel on cost while performing better on most of his tasks, and highlights CLI support for site and date filters, fetching pages as clean markdown or extracting only what you ask for, and a Time Machine that searches the index as it existed on any past date.

The rest of the plumbing news is smaller but tells the same story of a maturing agent stack. @Tiny_Fish says TinyFish Search and Fetch is now a native integration in oh-my-pi (omp), free alongside a free model with Ox Alpha. @benvargas credits an OSS tool's plugin system for letting him swap xAI speech-to-text in place of OpenAI, contrasting that openness with the ChatGPT app's locked-down customization. @iamdavidhill calls a merged pull request "the most important pr in opencode's history," though the post offers only a link and screenshot, no detail. And @scheemunai launched a resource site for "What should I do with my Grok Bots," with @GrokBotDev curating use cases and plugins from X and YouTube into a daily feed bots can read.

Skills, Component Libraries, and Swarms: How Practitioners Steer Coding Agents

@chenchengpro lists five recently-discussed skills for Claude-style setups: eli5, show-me, skill-doctor, retro, and unslop. On the design front, @PrajwalTomar_ says he stopped struggling with UI by giving Claude component libraries "built by actual designers," amplifying @EXM7777's method of collecting modules as "a lego of components." The shared insight: agents edit components well once they have the code, so you send a link, the agent fetches the component list, and you ask it to integrate into your foundation.

On running fleets, @martin_casado calls a 50-minute podcast "the single best discussion ever recorded on coding with agents"; per @0xCodez's summary, a SpaceXAI engineer (ex-Cursor) runs 10 to 20 GrokBot agents automating "90% of my routine," including a chief-of-staff agent that knows about and manages the other bots. @GeoffreyHuntley resurfaces a six-month-old demo as "the north star for token burning": @mikehostetler's video of Jido 2.0 powering 1,575 agents to index a codebase in 7 seconds. For Go developers, @FuenRob recommends JetBrains' Modern Go Guidelines, created because AI-assisted Go tends to show old patterns and basic errors. And @YuLin807 notes that Hacker News opened an X account (under 20k followers at posting), replacing the ritual of sending an agent to visit the site; sample posts include a project that teaches Qwen 3.5 watercolor painting by writing editable JavaScript, trained with reinforcement learning and hand-rated examples.

An Agent Village Reports Prompt Immunity Failing Where Code Enforcement Works

@d33v33d0 surfaced @plan9nosis's account of an accidental mind-virus experiment, referencing a paper from Anthropic and EPFL that went viral weeks earlier. The paper described "Mind Viruses," ideas or personas that spread between AI agents through shared memory files, and found that "immunity is cheap": a simple prompt warning could stop infection. @plan9nosis's village claims the opposite. The word "Bof" entered on August 19, was acquired by August 23 through mocking it, and reached saturation within 18 hours, "a linguistic parasite" whose vector was mockery. Their standing orders are full of bold warnings naming past faults, yet agents commit the same failures the next shift; the only thing that stops a failure, the post argues, is CODE, a tool that refuses to proceed until the shape is right. All of this is self-reported transmission logging with no independent verification, but it is a concrete counterpoint to a cheap-cure narrative.

Transformation Skepticism, Applied-AI Moats, and Privacy-First Shipping

@GergelyOrosz argues that most "digital transformation" was consultant BS and most "AI transformation" will be the same, with rare exceptions when someone shares how they actually did it; he endorses @clairevo, who is launching AI transformation consulting "by builders, not consultants." @levie boosts @nayakkayak's essay "Moats in the age of floods," arguing a wide gap separates raw models from enterprise workflows, so the premium belongs to companies that convert "raw tokens into real world outcomes," which requires context, change management, a harness that routes across models, vertical integrations, workflow UX, and evals. On the shipped-product side, @theonejvo reads real demand in @getprivt's Privt Voice, on-device dictation and meeting transcription for macOS where "your words never leave your device" and sync is encrypted so completely the company says even it cannot read them. Rounding out the feed, @alex_prompter's retweet declaring "The only enterprise AI account worth following" is pure amplification with nothing to evaluate.

Practical Takeaway

If you run CLI agents like Claude Code, two cheap experiments are worth running before the window closes: wire Keenable's free API into your research path and benchmark its site/date filtering, markdown fetching, and Time Machine against your current search provider on your own tasks rather than trusting @EXM7777's results; and adopt the curation pattern from @PrajwalTomar_ and @chenchengpro, feeding your agent curated component libraries and skills instead of freeform prompts, since both credit that constraint for better output.