Summary
The morning slot captured 86 tweets with zero @bayesiansapien retweets, so the curated layer is absent for a third consecutive day and everything here comes from the AI-handle feed. The strongest signal is a six-post cluster around SpaceXAI's Grok Bot launch, agents that sign into your work accounts and do real work, and the most useful item in that cluster is not the launch but Hugging Face's Elie Bakouch noticing the onboarding URL is cursor.com/bot/onboarding and the dashboard redirect goes to cursor.com/dashboard, which reads as Grok Bot running on Cursor infrastructure. Bakouch is also the slot's second-best item, reading Mistral's new in-region inference and European infrastructure announcement as a deliberate pivot: Mistral's own models trail the best open weights, so selling infrastructure lets it win large clients without forcing its models on them. Anthropic's Boris Cherny contributes the only substantive engineering observation, that LLM-generated bugs have shifted from off-by-one errors to system design, UI usability and missing broader context, and that adversarial code review is the effective counter. Google Research advanced AMIE to real-time video consultations with a randomized controlled trial over 300 simulated consultations. Two robotics items arrive as scaling-law claims rather than demos, one on Dyna-2 training on a million hours of egocentric video with no robot data in pre-training, one on Kodiak's 35 driverless trucks. Over half the slot's volume is off-topic political content from six handles (@MarioNawfal, @brivael, @spencerpratt, @HouseGOP, @WHFraudTF, @DoWCTO), which is the dominant fact about this capture even though none of it belongs in the wiki.
Posts
Grok Bot launches, and the infrastructure question is more interesting than the product (cluster of 6) (@mntruell, @milichab, @ns123abc, @theskory, @stepango, @eliebakouch). SpaceXAI announced Grok Bot in early beta, framed as "AI teammates that do real work for you," which sign into your existing tools, use them the way you would, and return finished work. Cursor CEO Michael Truell called it "an early step towards capable, delightful digital colleagues," and xAI's Mili Chab reported using it to send calendar invites, manage email and run coding projects from a phone. The substance in the cluster comes from two skeptical reads. Elie Bakouch, on the Hugging Face side, is "very worried about vendor locking on this kind of apps," arguing it makes sense for large labs but that enterprises and users should be careful, since the whole value proposition is that the agent holds credentials to every tool you use. He then followed up with the detail that makes the cluster worth recording: the onboarding page is
cursor.com/bot/onboardingand pressing "return to dashboard" lands oncursor.com/dashboard, so the product appears to be served on Cursor infrastructure, which The Information separately confirmed as a joint SpaceXAI and Cursor development under their partnership. Bakouch's own hedge is fair, that it is "still a great product" and likely a good interface for non-technical users. @stepango and @theskory contributed promotional amplification, and Benji Taylor's animated icon note is craft rather than signal.Mistral repositions as Europe's inference provider, and the read on why is sharper than the announcement (@eliebakouch · Mistral). Mistral published "In-region inference, open models, and new European infrastructure for sovereign AI," adding in-region serving and new European compute alongside its existing Studio, Forge, Vibe and AI Cloud products. Bakouch's framing is the valuable part: "mistral, the inference provider company of Europe." His reasoning is that Mistral's own models are "largely behind open model," so building the infrastructure layer "lets them get big clients on the infra part in without forcing their own models on them." He calls it a smart and hard-to-execute move. That is a specific strategic claim worth tracking, because it means the sovereignty requirement is being monetized as hosting rather than as model quality, and it puts Mistral in the same business as the neoclouds rather than the frontier labs. This connects to the router-market repricing this wiki covered on 08-11, where Stripe is in advanced talks at around $10 billion for OpenRouter and the strategic read was that the durable asset is the metering position rather than the routing policy. Mistral is claiming a geographic version of the same position.
Anthropic's Boris Cherny: LLM bugs changed shape, and adversarial review is the counter (@bcherny). Cherny's claim is that models still produce bugs but the distribution moved: "less off-by-ones and more about system design, ui usability, missing broader context." His conclusion is that some kinds of coding are solved and others are not, and that adversarial code review has been an unusually effective tool for the remaining classes. The practical form he gives is deliberately low-effort, either a one-line prompt like "use a dynamic workflow to adversarial test every edge case in an iOS simulator," or Claude's built-in
/code-reviewwith an effort level. This is a first-party observation about what production LLM coding failures actually look like, which is scarcer than benchmark numbers, and it lines up with the 08-10 finding that AI-generated C++ has a distinct repeatable quality profile across 3.52 million production changes, concentrated in interface and coupling burdens rather than in local logic errors. Two independent sources now say the same thing: the residual defects are architectural, not arithmetic.Google Research advances AMIE to real-time audio-visual clinical consultations (@GoogleResearch · blog). AMIE, Google's research medical diagnostic dialogue system, now supports real-time video consultations, and the accompanying study reports expert-level performance in a randomized controlled trial across 300 simulated consultations. The trial design is the notable part rather than the capability claim, since randomized controlled comparison against clinicians is a stronger evidentiary standard than most multimodal-agent papers meet. The captured article body was navigation boilerplate rather than the post content, so the 300-consultation figure and the expert-level claim come from the tweet text itself and should be treated as the announcement's own framing until the blog is read directly.
A claimed robotics scaling law from video alone (@Scobleizer). Robert Scoble amplified a post from @VaderResearch about Dyna-2, trained on 1 million hours of egocentric video with no robot data in pre-training, reporting that performance scales across the 1K to 1M hour range, so "more video, better performance." The framing is that robotics may finally have a scaling law. Scoble adds that he had been hearing the same thing from San Francisco researchers without connecting it. Treat the claim cautiously: the interesting quantity in a scaling law is the exponent and the point where it breaks, and a monotone improvement across three orders of magnitude of data is necessary but not sufficient evidence. Recorded because a video-only pre-training result with no robot data would matter for the data-substrate question well beyond robotics.
Autonomous trucking reaches 35 driverless trucks (@Scobleizer · @don_burnette). Kodiak Robotics CEO Don Burnette reported that Kodiak's autonomous driving solution now runs in 35 customer-owned trucks operating with no humans in the cab, described as the largest such fleet in the world, roughly one year after the company went public. He rang the Nasdaq opening bell to mark the milestone. Customer-owned rather than company-operated is the detail that matters, since it means someone other than the vendor is carrying the operational risk.
A hardware-supply constraint on robot data collection (@Scobleizer). Shenzhen Foundry reports confirmation from a sensor supplier that Xiaomi and Unitree have effectively locked global-shutter camera capacity for the next six months. The warning is aimed at anyone planning robot-learning datasets on the assumption that global-shutter sensors are freely available; the poster's partner factory has a production-ready egocentric capture platform with 2 to 5 global-shutter cameras, 1 microsecond hardware frame sync and a 200 Hz IMU. Worth recording next to the Dyna-2 item above, because a video-scaling thesis for robotics is a data-collection thesis, and the sensors needed to collect that data are apparently allocated.
Sequoia's Harvey case study on building a research lab on a budget (@Scobleizer). Pat Grady published a talk from Sequoia's Sovereign AI event in which Harvey's Gabe Pereyra describes a "moneyball" approach to research capability without frontier-lab resources. The chapter list is the useful artifact: building a research lab on a budget, Legal Agent Bench plus contracting and diligence datasets, and domain experts guiding synthetic data generation. That last item is the one with technical content, since expert-directed synthetic data is the cheapest known substitute for proprietary training data and the legal domain is where the evaluation signal is unusually clean. No numbers in the captured text.
An agent managing home network configuration (@dhh). DHH gave an agent a local admin account on a Ubiquiti setup with a cloud router and three U7 Pro access points, then had Claude optimize radio channels, tune mesh settings and verify device roaming. Recorded as a small concrete instance of the credential-holding agent pattern that Bakouch is worried about in the Grok Bot cluster above, arriving the same day from someone enthusiastic about it rather than cautious.
A DeepMind researcher reports an account compromise (@zhu_hanqing666). Hanqing Zhu at Google DeepMind says the account was hacked and asks followers not to click or trust links from it. Recorded only as a caution for anything else attributed to that handle in this window.
Off-topic political and lifestyle volume. Skip. Six handles (@MarioNawfal with 20 tweets, @brivael with 18, @spencerpratt with 7, @HouseGOP, @WHFraudTF, @DoWCTO) account for more than half of the slot's 86 captured tweets, covering US and French domestic politics, Middle East conflict reporting, Los Angeles municipal disputes and celebrity items. None carries AI research or industry signal. Two Tesla posts and one Scoble post on Full Self-Driving user testimonials are product marketing rather than technical content and are also skipped.
Promotional items with no research substance. Skip. Google's 2026 Carbon Removal and Superpollutant Elimination R&D Awards call for proposals, Robert Scoble's UNALIGNED team-intake page and his newsletter promotion, Max Levchin on M&A advice, and Scoble's posts on computer-controlled car lightbars and AI video culture.