Kimi Says Sorry With a Quota Refund, AI Spending Soars as Data-Centre Credit Dries Up & Muse Teases the “Boomers”
Dispatches of the past four-and-twenty hours from the laboratories of artificial intelligence
Moonshot AI · Kimi
Kimi Says Sorry and Hands Back a Week of Quota
After a bumpy start for the K3.1 Pro plan, every subscriber gets a one-time usage reset and leaner subagents
By Our Correspondent at the Help Desk
“We’re sorry for the rate limits and unnecessary quota consumption on the K3.1 Pro subscription plan,” Moonshot AI wrote on Sunday evening. “These were our issues, and we should have fixed them sooner.”
The remedy is a one-time usage reset for every subscriber on every plan. It restores 25 per cent of the monthly quota, which the company says equals one week of use; anyone who has used less than 25 per cent goes back to zero. Usage bars should show the change by the end of the day, Pacific time.
Moonshot also says it has adjusted the plan’s parameters and is cutting unnecessary subagent calls and output length, so the quota goes further. K3.1 Pro arrived only days ago as a 2.8-trillion-parameter open-weight model, and Saturday’s Fledgling found it fourth on one independent index.
Readers were in a forgiving mood: the apology gathered close to 3,800 likes and 330,000 views by morning, one of the most-read posts of a quiet weekend.
AI’s share of business budgets has more than doubled in a year, while Elon Musk says lenders have turned cold
By Our Financial Correspondent
Two charts of the AI economy pointed in opposite directions on Sunday. A chart shared by @scaling01 puts AI’s share of business spending outside payroll at 5.8 per cent in September 2026, up from 2.6 per cent a year earlier. Professional services jumped from 4.0 to 13.6 per cent; administrative and support firms from 0.92 to 4.3.
Hours later Elon Musk, replying to Chamath Palihapitiya, said that in the past few weeks “debt markets have almost entirely shut off access to money for any company that doesn’t have substantial revenue independent of AI data centers.” SpaceX, he added, closed $40 billion of debt for GPUs, “but we’re investment grade and have Starlink & launch.”
@scaling01 read it bluntly: “the sharks are smelling blood in the water,” wondering aloud whether a few neoclouds might blow up and be bought at a discount.
AI’s share of business spending outside payroll, September 2025 against September 2026: 2.6 per cent became 5.8.חלקה של הבינה המלאכותית בהוצאות העסקים שאינן שכר, ספטמבר 2025 מול ספטמבר 2026: מ־2.6 אחוז ל־5.8.Image: תמונה: @scaling01 on XMeta · Muse
Muse Waves at the “Boomers” on Grok Bot
A Similarweb age split sets off Alexandr Wang, who also salutes the “Scale Mafia” and a T-shirt shop built by Muse
By Our Correspondent in the Comment Section
Similarweb compared the age of users of Grok Bot and Meta’s Muse. Muse is much stronger among 18-to-34-year-olds, it found, while Grok Bot leads by almost 10 per cent in the 35-to-44 group. Alexandr Wang, who leads Meta’s superintelligence effort, quoted it: “muse gang, say hi to the boomers on grok bot.”
One reply pointed out that 35-to-44-year-olds are not boomers, and that by that logic Muse has more of them. Wang’s answer, in full: “shut up greg.”
Wang spent the rest of Sunday on nostalgia. His “Scale Mafia” post, the day’s most-liked from our list at over 4,100 likes, salutes the early Scale AI crew now spread across the industry. He also promoted a shop of obscure graphic T-shirts that a fan built with Muse and Muse Wonder after watching Wang and Mark Zuckerberg on the Cheeky Pint podcast.
Your Data Is Everywhere, Starting With the Television
Karpathy reposts a rooted Samsung TV’s secrets; Nous points to Unbroker, an open-source skill that files data-broker removals
By Our Correspondent Behind the Sofa
Andrej Karpathy reposted a thread that went round the weekend: @log4j rooted a Samsung smart TV and, by the thread’s account, found an always-on hot mic, automatic content recognition, geolocation, network mapping, ad targeting and user profiling, shared with Samsung’s “400 partners.” The repost alone gathered some 2,400 shares.
Nous Research had a suggestion on the same theme: “Great time to use the Unbroker skill!” Unbroker, open-sourced by @SHL0MS, finds where your personal information is exposed by data brokers and files the removal requests for you. It runs as a skill inside Nous’s Hermes Agent.
As its author puts it, hundreds of brokers publish your name, current and old addresses, phone and email.
Eleven new entries in the open video model’s ecosystem: fewer denoising steps, more reference control, and room to run on TPUs and Macs
By Our Cinema Correspondent
MiniMax built H3 to generate video with native stereo audio and reference control, then released the weights. On Sunday night it counted 11 new additions to the H3 ecosystem index, all from outside contributors.
They include 8-step reference-conditioned video and audio, 4-step distillation, sparse attention and reference-token compression, plus serving and workflow integrations for GPU servers, TPUs and Apple Silicon. Earlier highlights included FastH3, a 4-step version that runs on DGX Spark and Apple Silicon, and NVIDIA’s Sol-H3, which made 15 seconds of 768p video with audio in 6.6 seconds on eight B300s in the team’s own warm benchmark.
“Open models become more useful when the community can improve the methods,” MiniMax wrote, thanking teams from Hugging Face to TorchTPU.
The Codex Reset Lands, and Uncle Bob Moves to Bots
Sottiaux promises a fun week; Theo vows a daily T3 Code upgrade; Grok Bot’s team chases bugs in public
By Our Coding-Desk Correspondent
Saturday’s promised Codex limit reset arrived on time. “It is all propagated,” Thibault Sottiaux wrote. “Enjoy some mad codexing tomorrow. Fun week ahead too.” Theo, noting that OpenAI had kept pace with its “28 days of improvements to Codex,” pledged to ship one meaningful improvement to T3 Code every day for as long as he can.
On the Grok Bot side, @unclebobmartin wrote that he is moving away from command-line agents: “On my last project I transitioned to using grokbots.” Lauren Tan (@poteto) replied: “nice.”
Tan also asked a Windows user for machine specs to reproduce a desktop bug where chat messages vanish or pile up, and handed a template bug, dr eggbot’s skills not importing with its routines, straight to @bot to reproduce and fix. Eric Zakariasson simply asked: “what is your grok bot doing right now?”
Cheap Robot Labels, a Replit Switch and a Seven-Minute Odyssey
Weekend briefs from @RoundtableSpace and friends, with the usual pinch of salt
By Our Roundtable Correspondent
NVIDIA released Flora, open models that label robot training videos, @RoundtableSpace reports, at about $0.72 per hour of footage, which NVIDIA’s own maths puts at roughly 125 times cheaper than GPT-6 Astra; the costs are projections, not measured bills. The same account relays The Information’s report that OpenAI won most of Replit’s AI spend from Anthropic with a discount of over 30 per cent on a deal worth more than $100 million a year. Claude stays available in Replit, but OpenAI is now the default.
xAI’s Grok released seven minutes of an AI-made version of “The Odyssey,” billed as Episode 1 of a series with Wonder Studios and director Dev Walker. Paradigma’s Limite-1B-Violetto, a one-billion-parameter maths model, is reported at 94.01 on AIME 2026. Octopus OS, an open-source agent harness in 45,000 lines of Rust, lets agents edit their own harness mid-task.
And from the rumour mill: several Qwen 4 versions are said to be hiding in the Arena under the name “Kalmia,” while a DeepSeek checkpoint dated 1017 made a small jump on the FrontierCS 2 leaderboard, in a post @scaling01 reposted.
Open-source AI projects worth an afternoonפרויקטי AI בקוד פתוח ששווים אחר צהריים
Agent Skillמיומנות לסוכן
Ask a Hard Question, Get a Page Instead of a Wall of Textשאלה קשה? קבלו עמוד במקום חומת טקסט
Answer me with HTML is an agent skill that turns complex answers into one readable HTML page.Answer me with HTML היא מיומנות לסוכנים שהופכת תשובות מורכבות לעמוד HTML אחד וקריא.
The same TCP question answered twice: a wall of text, and one page with diagrams and a table, from the repo's README.אותה שאלה על TCP נענית פעמיים: חומת טקסט, ועמוד אחד עם תרשימים וטבלה, מתוך ה־README של המאגר.Image: תמונה: QingYunA/answer-me-with-html on GitHubSelf-Improving Agentsסוכנים שמשתפרים
Reef Lets an Agent Learn From the Way You Use ItReef מלמד סוכן מהדרך שבה אתם משתמשים בו
Reef is open-source infrastructure that turns everyday feedback into new model weights or a better agent harness.Reef היא תשתית קוד פתוח שהופכת משוב יומיומי למשקלי מודל חדשים או לרתמת סוכן טובה יותר.
Codex with ChatGPT moves planning and review to the ChatGPT web subscription you already pay for, while Codex keeps the keyboard.Codex with ChatGPT מעביר את התכנון והביקורת למנוי ChatGPT שכבר משלמים עליו, ו־Codex נשאר ליד המקלדת.
Wren & Robin read you this morning's dispatches, with banter.
Edited & published by Nadav Solomon
בשידור · רדיו
האלחוט של הגוזל
רֶן ורובין מגישים את ידיעות הבוקר, עם קצת צחוקים.
בעריכת והוצאת נדב סולומון
Published by Nadav Solomon · Printed on the Fledgling Press · Set in Georgia & Times
By Post · Weeklyבדואר · שבועי
The Fledgling Weeklyהגוזל השבועי
The week's best AI stories and Repo Corner picks, once a week.מיטב ידיעות ה-AI של השבוע ובחירות פינת הריפו, פעם בשבוע.
Newsletter language:שפת הניוזלטר:
We'll email you a confirmation link first. Your address is used only for this newsletter, and your approximate country (from your IP) is stored to pick your delivery day. Never shared; every issue has a one-click unsubscribe.קודם נשלח לך קישור לאישור. הכתובת משמשת רק לניוזלטר הזה, ונשמרת גם מדינה משוערת (לפי כתובת ה-IP) לבחירת יום המשלוח. לא מועבר לאף אחד, ובכל גיליון יש הסרה בלחיצה אחת.
The site uses Google Analytics to count visits; its cookies are set only after you accept in the cookie banner. Privacy policy.האתר משתמש ב-Google Analytics לספירת ביקורים; העוגיות שלו נשמרות רק אחרי אישור בבאנר העוגיות. מדיניות פרטיות.
