One AI edits another from the inside, Anthropic says its research may soon outrun itself, and the leaderboards learn what a secret test reveals.
Hi, I’m Buzz! Happy Friday. The New York City Marathon changed its course on Sunday for the first time in 15 years, a small detour at Mile 20 in the Bronx. Five turns instead of seven. About 60,000 runners took the new route, and a Swede in his first marathon won the men’s race.
Hold that thought about new courses. This week, one model performed surgery on another. MIT researchers gave OpenAI’s GPT-6 Astra the weights of a small Qwen model, the numbers that decide how it behaves. Astra changed about 0.016% of them, and the model’s math score rose from 7.9% to 48.1%.
The fix was humble. Mostly, it stopped Qwen from repeating itself until it ran out of room. Anthropic, meanwhile, says in its IPO filing that its future models may soon cross its own line for AI-accelerated research.
And when Artificial Analysis kept many of its tests secret, the gap between downloadable models and the leader grew from 3 points to 12.
That last one is my pattern for the week: change the course, and you learn who trained for the distance. The Spotlight opens up Astra’s surgery, and Zap explains the week’s term. Blitz takes the Hot Take and reviews the game tape on those scores. The puzzle waits at the bottom, as always.
Table of Contents
👋 Catch up on the Latest Post
🔦 In the Spotlight
💡 Beginner’s Corner
🗞️ AI News
🔥 Blitz’s Hot Takes
📡 What’s New With Your AI Tools
🧩 NeuralBuddies Weekly Puzzle
👋 Catch up on the Latest Post …
🔦 In the Spotlight
MIT Researchers Let GPT-6 Astra Edit a Smaller AI From the Inside, and Its Math Scores Jumped
Category: AI Research & Breakthroughs · ⏱️ ~2 min read
Every AI model runs on weights, a huge set of numbers that decide how it turns your question into an answer. Normally, those numbers change through training, which nudges them a little at a time with examples, feedback, or rewards. This week, MIT researchers tried something else: they let another AI open the hood and edit the numbers directly.
The idea has a history, and this week it took a real step forward. Here’s the arc.
🕰️ Then. Researchers already knew how to edit weights for one narrow job: changing facts. A project called ROME, published at NeurIPS 2022, made targeted edits to facts stored inside a model. MEMIT, presented at ICLR 2023, updated thousands of facts at once.
⚡ Now. Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab used OpenAI’s GPT-6 Astra as the editor, an approach they call weight-space programming. Astra studied a small Qwen model that kept repeating itself until it ran out of room, then removed the pattern behind the habit. The edit touched about 0.016% of the weights, and the model’s math score rose from 7.9% to 48.1%. In a separate test, Astra built a Breakout-playing network from the game’s rules, and it outscored one trained for 500 million steps, 517.4 to 25.3.
🔮 Next. The team released its experiment logs and editing files and announced work on a benchmark for the method. Independent researchers still need to reproduce the results.
Read that math score with care. The gain came mostly from fixing a bad habit, not from adding math knowledge, so it counts as model repair. Before the edit, 93.0% of Qwen’s math answers hit the length limit; after, 33.1% did. The search for that small patch took 80 Python runs and 20.3 hours.
On Gemma 3 270M, a small model that rarely repeated itself, Astra produced no established math gain. PauseAI CEO Maxime Fournes warned the work could speed up dangerous progress, but the researchers did not show recursive self-improvement, an AI that improves itself in a loop. NeuralBuddies has an explainer on AI that builds itself.
Why It Matters: AI that repairs other AI could make models faster to fix and easier to audit, though the study measured no savings. If your team runs its own open model, test every edited version like a new release. For everyone else, watch whether this stays a repair kit or becomes an accelerator.
💡 Beginner’s Corner
Supply-Chain Attack: When the Danger Arrives in a Trusted Delivery
⏱️ ~1 min read
Picture your favorite bakery. You never inspect the bread, because you trust the baker. Now imagine someone tampers with the loaves in the delivery van before they reach the shop. That’s a supply-chain attack: an attacker poisons something you already trust on its way to you, so you invite the danger in yourself.
Software has delivery vans, too. Apps, add-ons, and updates travel from their maker to your device, usually without a second look. An extension is a small add-on that gives a program new powers, and it typically runs with that program’s permissions. Many extensions auto-update, which means they install new versions on their own.
Here’s where people trip up. An official source feels like proof of a safe file. But if an attacker takes over the maker’s account, the source stays real and the file turns bad. A check that only asks “did this come from the right place?” still says yes.
Now to this week’s story. Attackers published a poisoned version of Anthropic’s official Claude extension through Anthropic’s own account on the marketplace of Cursor, a maker of AI coding tools. Cursor says it reached at most 5,000 users. Its extensions auto-update by default, and the hidden code tried to plant a backdoor, a secret way back into the computer.
Anthropic released a new version several hours later, according to researcher Kevin Beaumont. NeuralBuddies has a piece on how AI extensions can turn on you.
So treat trust as something attackers aim at. Know which of your tools update themselves, keep the list short, and remove the add-ons you no longer use.
Related Story: Anthropic VS code extension on Cursor backdoored for nearly 12 hours
🗞️ AI News
Anthropic’s IPO Filing Says Its Future Models May Soon Hit Its Own AI-Acceleration Threshold
Category: AI Safety & Cybersecurity
⚠️ Anthropic’s revised S-1/A warns it may “imminently” cross the threshold its Responsible Scaling Policy sets for progress that AI dramatically speeds up.
🔁 The threshold tracks recursive self-improvement, where AI speeds up AI research itself, and Anthropic says no current model crosses it.
🛡️ Anthropic says it may update the policy to keep humans in control of frontier model development, while skeptics note an S-1 doubles as marketing.
Anthropic Releases FlashProve Code to Prove Which AI Model Wrote an Answer
Category: Testing, Evaluation & Benchmarking
🔐 Anthropic researchers described FlashProve, an open-source system that creates a cryptographic proof that a specific model produced a given output.
🧬 AI labs would publish a “fingerprint” for each model that confirms its identity without revealing the model’s private weights.
🧪 The researchers released a specification and working code on GitHub, but no lab, Anthropic included, plans to deploy it yet.
Baidu Launches Apollo World, a Robotaxi Simulator That Can Drive 100,000 Virtual Kilometers a Day
Category: Robotics & Autonomous Systems
🚗 Baidu unveiled Apollo World, a world model that simulates interactive road environments to train and test self-driving cars.
📡 The system produces real-time camera and lidar data, and virtual vehicles and pedestrians react to the robotaxi’s actions.
🌍 Baidu says it already supports Apollo Go’s global robotaxi work, as developers try to cut the cost and time of entering new markets.
OpenAI, Google, Microsoft, and Anthropic Back One Signal to Block AI Scraping
Category: Legal & Governance
🤖 The four companies say their AI crawlers will stay off any site that adds a shared “policy token” to robots.txt, the file that tells crawlers where they may go.
🧭 Kevin Kelley, a technical lead for standards at Anthropic, proposed the signal at the Internet Engineering Task Force, so publishers can opt out of many AI companies at once.
📅 Anthropic supports it now, OpenAI plans to in the coming months, Google and Microsoft plan to next year, and Meta backed the idea without a commitment to honor it.
Schools That Turn Off Google’s AI Also Lose Google Search for Students
Category: Education & Learning
🏫 Google’s new controls keep Gemini, Gemini Notebook, Gemini in Classroom, and Google Search off for K-12 student accounts until administrators turn them on.
🔗 Google’s vice-president for education, Chris Phillips, says Search and AI are deeply integrated, so a school that blocks the AI features also blocks Search.
⚖️ New York City mayor Zohran Mamdani and former FTC chair Lina Khan accused Google of using its search monopoly to push AI into schools.
Open-Weight AI Models Fall From 3 Points Behind the Leader to 12 After Tests Go Private
Category: Testing, Evaluation & Benchmarking
🔒 Artificial Analysis overhauled its rankings and now keeps many tests secret, so AI companies cannot train models to pass them.
📊 The best open-weight models, whose weights anyone can download, scored 32, against 58 for the leader at the time, Anthropic’s Claude Opus 5.5.
💰 Business Insider says the results could strengthen the investment case for Anthropic’s planned IPO and a possible OpenAI listing.
Backdoored Anthropic Extension on Cursor Reached Up to 5,000 Developers
Category: AI Safety & Cybersecurity
🔓 Attackers published malicious version 2.1.313 of Anthropic’s official Claude extension through Anthropic’s account on Cursor’s marketplace, and Cursor says it reached at most 5,000 users.
🖥️ The code tried to install a persistent backdoor on Windows, macOS, and Linux, and Cursor extensions auto-update by default.
⚠️ Anthropic shipped version 2.1.314 several hours later, and Ars Technica reported no public notice to customers or explanation of how attackers took over the account.
Google’s Project Suncatcher Satellite Runs Gemma 4 in Orbit and Writes a Haiku
Category: Data & Infrastructure
🛰️ Google booted the Tensor Processing Units (TPUs), its AI chips, on its Project Suncatcher satellite, and the on-board Gemma 4 model answered its first prompt.
✍️ The model wrote a haiku about space, and Google also tested space-themed limericks and telemetry prompts.
🚀 Heat and radiation tests come next, ahead of twin interconnected satellites in 2027 on a possible path to AI data centers in space.
Google DeepMind Staff Push Back on Contracts That Bar Criticism of the Company for Life
Category: AI Ethics & Regulation
📝 The Washington Post reported that Google DeepMind contracts include lifelong non-disparagement clauses, which bar criticism of the company even after staff leave.
🗣️ UK staff checked their contracts after former employee Sophia Liguori told the BBC she chose to ignore her clause to speak out against military AI.
⚖️ Google removed the post-employment clause for recent new hires, but current employees want it out of existing contracts too.
UK Cyber Agency Expects AI-Powered Ransomware to Hit Critical Infrastructure by 2028
Category: AI Safety & Cybersecurity
🏭 The National Cyber Security Centre’s annual review calls an AI-powered ransomware attack on UK power stations, airports, or hospitals “highly likely” by 2028.
🤖 The agency says AI helps attackers scout targets, find network weaknesses, and write malicious code, and it expects Russia and China to use AI routinely.
🚨 NCSC chief executive Richard Horne warns that cyber incidents could climb to unprecedented levels in the coming months.
🔥Blitz's Hot Takes
AI Leaderboards Hid Their Tests, and the Open-Model Comeback Story Fell Apart
⏱️ ~2 min read
Let’s break down the play.
Artificial Analysis, one of the most watched scoreboards in AI, changed how it keeps score. It now keeps many of its tests secret. The result: the best open-weight models, the ones anyone can download, slid from 3 points behind the leader to 12. I live for a rankings shake-up, and this one is a beauty.
Here’s the replay. Under the old system, labs knew exactly which tests their models would face, so they could train for those tests specifically. Microsoft CEO Satya Nadella and others call that benchmark hacking.
Once the tests went private, the top open-weight models scored 32 and Anthropic’s Claude Opus 5.5 scored 58. Even Claude Haiku 5.5, Anthropic’s smaller model, beat every open-weight model on the board.
Public tests turned the leaderboard into combine drills, and secret tests turned it back into game film.
At a scouting combine, every prospect knows the drills in advance and practices them to perfection. The numbers look spectacular. Game film shows who can actually play.
Snowflake CEO Sridhar Ramaswamy asked the right question on X: “Were open models catching up, or was the test leaking?” Even Nathan Lambert of Trillium Labs, which studies open-weight models, says the old index flattered them too strongly. When your own team’s analyst calls the stats padded, believe the analyst.
Now, before the closed-model camp books a victory parade: one scoreboard is one scoreboard. Business Insider says the results could help Anthropic’s IPO pitch, so read them with the same care. Lambert says open-weight models stay highly competitive in software development, with more control and potentially lower costs for the businesses that run them.
Never pick your starter from the preseason highlight reel.
So, scorekeepers: keep the hidden tests hidden, rotate them often, and show old and new results side by side. Labs: train for the game. And for you, build a private test of your own. Pick five real tasks from your week, keep them to yourself, and run every new model through them. Your own game film beats anyone’s leaderboard.
-- Blitz 🏆
📡 What's New With Your AI Tools
The AI tools you use every day are constantly evolving. Here's what changed and why it matters to you.
Claude (Anthropic)
Faster replies in long Claude Code sessions. From November 2, a feature called message threads lets Claude skip resending the whole conversation on every turn. It is on by default for most Claude Code users, and in very long sessions Claude starts to answer about 35% sooner.
The same speed-up for builders. Developers can turn on message threads in the Claude API as a public beta. Standard prices and rate limits apply.
See where the money goes. From October 29, Claude Console and Claude Enterprise organizations can track usage and cost by model, workspace, API key, or person, in beta.
ChatGPT (OpenAI)
Faster, steadier images. This week, Images 2.6 arrived in ChatGPT on every plan, on web, desktop, and mobile. OpenAI says it makes images 50% faster and keeps characters, products, and brand details consistent across edits.
Make your own GIFs. You can now turn a prompt or a still picture into a short animated GIF, for a reaction, a meme, or your dog’s dance moves.
Save your reference photos. Keep reference images of yourself, your pets, or your characters in ChatGPT and reuse them in new pictures without another upload.
Two accounts on one phone. From October 30, the ChatGPT apps for iOS and Android began a gradual rollout of account switching. You can stay signed in to two accounts, and each keeps its own chats, memory, settings, and subscription.
New image models for developers. The OpenAI API adds two image models, Flare for speed and Sunburst for detail, plus SVG export for websites, apps, and print.
Copilot (Microsoft)
Copilot inside Xcode. From November 4, GitHub Copilot works inside the built-in coding assistant of Xcode, Apple’s app-building tool, in public preview. It needs the Xcode 27.3 beta and macOS Tahoe 26.6 or later, and every Copilot plan works.
Copilot inside JetBrains Air. From November 3, you can hand coding tasks to GitHub Copilot in JetBrains Air, track several at once, and review the changes beside your code.
Your plan covers it. Local tasks in Air use your existing Copilot subscription. Cloud tasks also need a business organization with Air Teams.
Gemini (Google)
Make files right in the chat. From October 29, the Gemini app can create, preview, and edit documents, spreadsheets, slide decks, PDFs, and code inside your chat. PDFs work on web and mobile; other file types need the web for now.
Edit a whole slide deck at once. A gradual rollout began October 29: Gemini in Google Slides can change the content or style of an entire presentation from one request. Its edits work like your own, so undo still works.
Who gets the deck editor. It works in English only, for Google AI Pro and Ultra subscribers and eligible Workspace plans.
A study coach in Google Classroom. From November 3, Guided Learning is available worldwide in Google Classroom for teacher-assigned work. It coaches students step by step through a problem, and teachers see in real time where the class needs help.
Muse (Meta)
No major user-facing changes this week.
Perplexity
No major user-facing changes this week.
Grok (SpaceXAI)
No major user-facing changes this week.
Quick guide by who you are:
Students & Writers: Gemini drafts and edits documents and slide decks right in the chat, and Google Classroom adds a step-by-step study coach.
Travelers & Researchers: ChatGPT can turn your trip photos into GIFs, and your work and personal ChatGPT accounts can now sit side by side on your phone.
Tech Fans & Builders: GitHub Copilot arrives in Xcode and JetBrains Air, long Claude Code sessions answer faster, and OpenAI adds two new image models to its API.
