Karpathy Builds HackerTube, Claude Cracks Linux & Mythos 5.5 Peeks Out of the Code
Dispatches of the past four-and-twenty hours from the laboratories of artificial intelligence
Weekend Hack · Claude Code
Karpathy Asks Hacker News What to Watch, and Builds HackerTube in Two Hours
Claude sifted 43 million comments from twenty years of the forum down to 6,454 videos people keep recommending
By Our Correspondent on the Sofa
HackerTube: a familiar layout, filled only with videos Hacker News kept recommending.HackerTube: מבנה מוכר, ובתוכו רק סרטונים ש־Hacker News המליץ עליהם שוב ושוב.Image: תמונה: @karpathy on X
Andrej Karpathy was annoyed with YouTube’s recommendations, so he spent a weekend afternoon replacing them. He had Claude filter 43 million comments from twenty years of Hacker News down to 6,454 videos that people keep recommending, and put them on a new site, HackerTube. He says it took about two hours with Opus 5.5, and calls the result a “significantly better viewing experience.”
There are no logins; your history stays in your own browser. Each thumbnail is a frame from the middle of the video rather than the polished, A/B-tested image uploaders choose. The comment data is offered as a single zip file, so you can point your own Claude at it and build your own version. There is even a section called HackerTube Longs.
His only manual work was in the browser: creating an API key for video details, booting a server and buying the domain. Then he gave Claude the server’s address and told it to secure, deploy and optimise the code. “Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast,” he wrote. The post passed 2.5 million views and 33,000 likes.
An Unreleased Claude Wrote a Zero-Click Root Exploit for Linux
The kernel bug dates back to version 4.4 and is patched upstream; servers with multipath TCP were the target, and Go turns that on by default
By Our Correspondent in the Server Room
Anthropic’s Frontier Red Team says an unreleased Claude model wrote an exploit that lets a remote attacker run commands as root on Linux servers with a multipath TCP listener, with no login and no clicks. The bug, CVE-2026-98432, is a race condition in how the kernel sets up IPv4 TCP connections, and it goes back to Linux 4.4. The kernel’s CVE team rates it 9.8 out of 10. The write-up spread widely on X over the weekend.
In Anthropic’s proof of concept it took 48 minutes to leak a kernel address and about five more to reach a root shell. A lost race only fails one handshake and does not crash the machine, so an attacker can keep trying. The fix is already upstream and backported to every maintained stable kernel. Anthropic did not publish the exploit.
Go programs open multipath TCP listeners by default since Go 1.24, which is why the warning travelled fast. A note on the oss-security mailing list says GODEBUG=multipathtcp=0 turns that off, but stresses that the bug lives in ordinary TCP, so the real fix is a patched kernel.
“Claude Mythos 5.5” Turns Up in Anthropic’s Page Code
A hidden heading named the model; the visible page still says 5.1, and there are no specs
By Our Correspondent Reading the Source
A screenshot shared by @scaling01 shows a hidden heading on one of Anthropic’s pages reading “Claude Claude Mythos 5.5”, above the line “Our most capable model class for cybersecurity and biology research.” The tip first reached @kimmonismus by email. @scaling01 later clarified that the tagline describes the Mythos family as a whole, not 5.5 in particular.
According to @RoundtableSpace, the name was then pulled, and the visible page still says 5.1. Anthropic has announced nothing. In the same evening @scaling01 argued that, almost seven months on, Chinese models still have not caught up with Mythos Preview, which he puts at about 161 on Epoch’s capabilities index, though he expects the next Chinese frontier models to get there.
The hidden heading and tagline, as shared on X.הכותרת הנסתרת והמשפט שמתחתיה, כפי ששותפו ב־X.Image: תמונה: @scaling01 on X
A Game in Four Days and 1,150 Pull Requests, Playable Inside a Post
@poteto’s tibo.dog was built with Grok Bot and Cursor, music and pixel art included, all of it code
By Our Correspondent at the Arcade
@poteto released tibo.dog, a game about Tibo, on a $5.99 domain, made with Grok Bot and Cursor, and it plays right inside an X post. The trivia thread: about 1,150 pull requests in four days; a new lint rule every time a bug turned up; TLA+ models to hunt bugs in the story, cutscenes and game systems; and comments banned by lint, because “the fix is a better name or a test.”
There are no audio files: the music is synthesised live by a Game Boy-style chip written in code, and every sprite is code too. The simulation is deterministic, so the server replays a whole run to check a high score. The first download is about 124 KB, and the target is 60 frames a second with 500 enemies on a mid-range phone. Thibault Sottiaux, the Tibo in question, joked that they had planned it “months in advance.”
Eleven story-mode bugs that TLA+ models caught before players did.אחד־עשר באגים במצב העלילה שמודלי TLA+ תפסו לפני השחקנים.Image: תמונה: @poteto on X
@trq212 is building a fighting game of his own and has added a taekwondo master who fights with fire. He also switched the art from 3D models to generated sprites, made with GPT, and hopes to have a playable slice next week.
An Agent Goes Looking for New Antibiotics in Protein “Dark Matter”
DeepMind researchers used Antigravity to find a candidate peptide that is now in the wet lab
By Our Correspondent at the Bench
Google’s Antigravity team showed how Google DeepMind researchers used the agent platform to search millions of uncharacterised proteins for new antimicrobial peptides. Google’s ProtNLM model predicts what a protein does from its sequence; the agent filtered the database to a few candidates and settled on a 97-amino-acid bacteriocin from a Streptococcus species with no close match among reviewed proteins.
The agent then checked its AlphaFold structure against known bacteriocins, read the neighbouring genes, reviewed the literature and wrote a wet-lab plan. A new 3D structure viewer inside Antigravity lets the researcher and the agent point at the same part of the molecule. The candidate is now in wet-lab testing, and the team stresses that only the lab can confirm it.
MiniMax Thanks the People Squeezing Its 428B Model Onto Home Machines
A 2-bit version of M3.1 Flash weighs 143.9 GB, and there are recipes for a DGX Spark, Apple Silicon and four RTX 3090s
By Our Correspondent in the Garage
MiniMax thanked the community for making M3.1 Flash, its 428-billion-parameter model, run on more local hardware. Unsloth made 2-bit dynamic GGUF weights of 143.9 GB, tested on coding and tool use. Other builders shared a single-DGX-Spark recipe, quants for Apple Silicon through oMLX, AMD’s MXFP4 format for Instinct chips, and a setup that spreads the model across four RTX 3090s and system memory.
Lower-bit weights need less memory and move less data while the model writes, MiniMax explained, but the savings only help with careful calibration and smart placement of the model’s experts between GPU and system RAM.
A Full Beta, a Fake Muse, and an Air Conditioner Remote
Smaller dispatches from a quiet Sunday
By Our Correspondent with Odds and Ends
Nous Research says its Hermes mobile app has hit the most testers Apple allows in a beta, after the app arrived on 31 October and many people could not get in. It added that dealing with fake Hermes apps will be “much easier once you have your own app listed.”
Alexandr Wang posted a warning: if you met “muse” at a Halloween party and they asked for your number, that was not Muse, and it is a phishing attempt.
OpenAI Developers reposted Tony, who disliked his big, heavy Daikin air conditioner remote, so he had Codex write firmware that turns a LilyGO T-Display board into a replacement, and open-sourced it. And @RoundtableSpace counted 424 bots on Grok Bot Shelf, a directory you can filter by whether a bot only reads or also takes actions.
Open-source AI projects worth an afternoonפרויקטי AI בקוד פתוח ששווים אחר צהריים
Phone Agentסוכן לטלפון
Mika Works Your Android Apps on a Screen You Never SeeMika מפעיל את אפליקציות האנדרואיד שלכם על מסך שלא תראו
An open-source agent that taps, types and scrolls on a hidden background screen while you keep using your phone.סוכן בקוד פתוח שמקיש, מקליד וגולל על מסך נסתר ברקע, בזמן שאתם ממשיכים להשתמש בטלפון.
Left, your phone; right, Mika working on its hidden background screen.משמאל הטלפון שלכם, ומימין Mika שעובד על המסך הנסתר ברקע.Image: תמונה: wangzan101/Mika on GitHubAgent Skillsכישורים לסוכנים
geo-sleuth Finds Where a Photo Was Taken, and Shows Its Workgeo-sleuth מוצא איפה צולמה תמונה, ומראה איך
An agent skill that locates photos with no signs or GPS using map geometry, elevation skylines, satellite tiles and street view.כישור לסוכנים שמאתר תמונות בלי שלטים ובלי GPS, בעזרת גיאומטריה של מפות, קווי רכס, תצלומי לוויין ותמונות רחוב.
From 27,335 railway bridges to one camera position, step by step.מ־27,335 גשרי רכבת למיקום מצלמה אחד, צעד אחר צעד.Image: תמונה: Oldcircle/geo-sleuth on GitHubTrend Radarמכ״ם מגמות
new-term-radar Catches New AI Names Before the Keyword Tools Donew-term-radar תופס שמות חדשים בבינה מלאכותית לפני כלי מילות המפתח
A daily report of AI model, product and repo names that are starting to surge, read from X, Hugging Face, GitHub and sitemaps.דוח יומי של שמות מודלים, מוצרים ומאגרי קוד שמתחילים לתפוס תאוצה, מתוך X, Hugging Face, GitHub ומפות אתרים.
Signals go in at the top; a daily report and a keyword table come out.האותות נכנסים למעלה, ודוח יומי וטבלת מילות מפתח יוצאים למטה.Image: תמונה: zhuyansen/new-term-radar on GitHub
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.
