Two Claude Code Builders Say Goodbye, Muse Tops a Word Game & Hermes Eyes the App Store
Dispatches of the past four-and-twenty hours from the laboratories of artificial intelligence
Anthropic · People
Two Claude Code Builders Leave Anthropic in a Day
Thariq Shihipar, a familiar voice of Claude Code on X, and Anthony Morris, who created Claude Code on desktop, both say goodbye
By Our Correspondent at the Farewell Party
“I’ve left Anthropic after more than a year of working on Claude Code,” Thariq Shihipar wrote, “and seeing it scale larger than I ever imagined.” The post praises the mission, “most of all, the idea of a race to the top,” and ends with “a bit of a break before diving into something new.”
Hours later Anthony Morris posted that the day before had been the last one. “I created Claude Code on desktop because I believe everyone can be a builder.” Walking away was “the hardest thing I’ve done in my career. But now I have some free time for GTA VI.” What comes next will be shared soon, the post promises.
The news broke before either of them posted. @ArfurGrok, an account that tracks hires and exits across tech, flagged the two departures 45 minutes apart, while both X bios still said Claude.
Thariq pointed fans to Addy Osmani, Lydia Hallie, Cat Wu, @ClaudeDevs and Boris Cherny, “who started it all.” Thariq also reposted Alex Albert’s list of Anthropic people to tag with product feedback. The reply to Cat Wu’s farewell: “installing world of warcraft rn.”
LisanBench now tests reasoning levels; GPT models trip on one rule, and the cost of a good score halves every few months
By Our Puzzle Editor
@scaling01 calls it a massive update to LisanBench, a word-chain game scored by path length. It now tests several reasoning levels per model. Meta gave free access to its Muse Wonder models, and the post is labelled as an ad, with the tests and conclusions stated as the author’s own.
At maximum reasoning, Muse Wonder Full and Muse Wonder rank first and second, though still slightly less efficient than Anthropic’s Opus 5.5. DeepSeek-V4.1-Flash sits on the cost frontier with Haiku 5.5, and Gemini 4 Argon on low takes first place among models that do not think first. Alexandr Wang reposted it: “muse wonder full and muse wonder are sota on lisanbench.”
GPT-6.1-Sol and GPT-6 score poorly because they break the rule that each step stays one edit away. Drop that rule and the table flips: GPT 6 Astra on medium leads with 56,140, ahead of Gemini 4 Argon high and GPT 6.1 Sol medium. The cheapest way to reach 2,000, 5,000 or 10,000 points has fallen 22 to 25 per cent a month.
Path length against reasoning tokens: Muse Wonder Full climbs highest, Opus 5.5 gets there with fewer tokens.אורך המסלול מול טוקני החשיבה: Muse Wonder Full מטפס הכי גבוה, Opus 5.5 מגיע לשם עם פחות טוקנים.Image: תמונה: @scaling01 on XGrok Bot
A Bot That Reads Grok Bot’s Changelog for You
Matt Palmer’s What’s New Bot brings updates to your chat; Lauren Tan shows bots settling a pull request in Slack
By Our Correspondent in the Group Chat
Matt Palmer built a What’s New Bot for Grok Bot. Each day it checks the Grok Bot changelog and Palmer’s latest “What’s new” video, then delivers both to your chat. When one user noted the changelog pages can lag several days behind, Palmer said the team is working on it.
Lauren Tan (@poteto) shared a favourite habit: add your bots to Slack and work alongside them like teammates. The attached screenshot shows two agents checking whether their pull requests clash, and agreeing they are complementary.
Lee Robinson offered $100 through X Money to each of five people with the best bug report or feature idea. And when a user called the voice chat sloppy, Eric Zakariasson said it runs on Grok Voice and asked for examples to fix.
Two bots and one human in a Slack thread: “work it out with each other.”שני בוטים ואדם אחד בשרשור Slack: „תסתדרו ביניכם”.Image: תמונה: @poteto on XNous Research
Hermes Turns Up on a Home Screen
Nous teases an app icon without a word, and warns that Hermes-branded apps already in the App Store are fakes
By Our Messenger Correspondent
Nous Research posted a phone screenshot with no words: an engraved Hermes wallpaper and a single app icon in the dock. Replies read it as Hermes Agent heading for the App Store. To the one fan who asked why an agent needs an app, Nous replied: “We found him! The one guy who doesn’t want an app.”
Asked about Hermes look-alikes already in the store, Nous said legitimate third-party projects do not use the Hermes branding, and “these are generally scams.”
Tibo Rests, the Limits Reset, and ChatGPT Frees an Embroidery Machine
Day 20 of the Codex diary is a day off; a $12-a-month app meets its replacement
By Our Weekend Correspondent
Thibault Sottiaux’s Codex diary reached Day 20 with three words, “Rest and reset,” and then a promise that the reset lands “by midnight PST.” Earlier on Saturday, Sottiaux invited engineers who have scaled products to a billion users to get in touch.
@OpenAIDevs reposted a happier hack. @simpsoka bought a $300 embroidery machine whose biggest complaint is a mandatory $12-a-month app, then had ChatGPT write a new app that talks to the machine. “Generative art patches incoming!”
Microsoft’s Skill Recorder learns a task from one demonstration; Baidu opens a whole-document OCR model
By Our Roundtable Correspondent
Microsoft open-sourced Skill Recorder, according to @RoundtableSpace. You press record and do a task once, talking through it if you like. It captures clicks, app switches, pages and terminal commands on your own machine, GitHub Copilot turns the recording into editable steps, and one click makes it a skill or a scheduled automation. It runs on Mac, Windows 11 and Ubuntu and needs Copilot access.
Baidu’s Unlimited-OCR, built on DeepSeek-OCR, reads a whole multi-page PDF in one pass and can be tried free on Hugging Face. And one Reddit user had Claude Code turn their own CBCT and MRI scans into a 3D model of their skull in Three.js; the jaw specialist, by the user’s account, nearly cried.
Open-source AI projects worth an afternoonפרויקטי AI בקוד פתוח ששווים אחר צהריים
Phone Agentסוכן בטלפון
A $7 Board Gives an AI Eyes and Hands on a Real iPhoneלוח של 7 דולר נותן לבינה מלאכותית עיניים וידיים על אייפון אמיתי
ESP32 iPhone Agent lets any OpenAI-compatible vision model tap and type on an iPhone, with no Mac, jailbreak or developer account at run time.ESP32 iPhone Agent מאפשר לכל מודל ראייה תואם OpenAI להקיש ולהקליד על אייפון, בלי מק, בלי פריצה ובלי חשבון מפתחים בזמן הריצה.
The phone, the mirrored screen and the agent's step log while it sets a repeating alarm, from the repo's README.הטלפון, המסך המשוקף ויומן הצעדים של הסוכן בזמן שהוא מכוון שעון מעורר חוזר, מתוך ה־README של המאגר.Image: תמונה: onehumanbeing/esp32-iphone-agent on GitHubAgent Audioשמע לסוכנים
Ears for Coding Agentsאוזניים לסוכני קוד
earshot lets an AI agent describe, judge and even make sound, locally on a CPU.earshot מאפשר לסוכן בינה מלאכותית לתאר, לשפוט ואפילו ליצור צליל, מקומית על המעבד.
Frames from earshot’s 22-second launch film, whose every sound was made in code, from the repo.פריימים מסרטון ההשקה של earshot, באורך 22 שניות, שכל צליל בו נוצר בקוד, מתוך המאגר.Image: תמונה: Leonxlnx/earshot on GitHubFilms in Codeסרטים בקוד
Fifty-Nine Film Styles for Claude Code to Directחמישים ותשעה סגנונות קולנוע ש־Claude Code יודע לביים
Lemo-Opuscar gives Claude Code a library of film styles, each with a demo made entirely in code, and lets you make your own.Lemo-Opuscar נותן ל־Claude Code ספרייה של סגנונות קולנוע, לכל אחד סרטון הדגמה שנעשה כולו בקוד, ומאפשר ליצור סגנון משלכם.
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.
