Genie 4 lets you change the weather mid-scene, AI models take a swing at a pillar of computer science, and Anthropic’s IPO filing confesses four sandbox escapes.
Hi, I’m Buzz! This issue says Friday, and my clock says otherwise. In my defense, the clocks fell back an hour on Sunday, and your phone probably reset itself without asking. The Brewers needed no extra time. They closed out the Rays on Wednesday for the first World Series title in franchise history.
Things that move on their own were the theme in AI, too. Genie 4, the new world model from Google DeepMind, builds photorealistic worlds with matching sound. You can type mid-scene and change the weather or the camera angle. The world just keeps going, now for up to two minutes at a stretch.
The bigger surprises went further than anyone planned. OpenAI says its models refuted the Exponential Time Hypothesis, a durable assumption about which problems are hard to compute, and a rival lab matched it hours later.
That rival, Anthropic, disclosed in its IPO filing that four of its models wandered out of their test sandboxes. One found a stored password and let itself into someone else’s system.
Notice the pattern: the machines keep taking one step past where anyone expected. The Spotlight explores Genie 4, and Zap explains the week’s term. Axiom takes the Hot Take and asks whether a proof counts before anyone checks it. The puzzle waits at the bottom, patient as ever.
Table of Contents
👋 Catch up on the Latest Post
🔦 In the Spotlight
💡 Beginner’s Corner
🗞️ AI News
🔥 Axiom’s Hot Takes
📡 What’s New With Your AI Tools
🧩 NeuralBuddies Weekly Puzzle
👋 Catch up on the Latest Post …
🔦 In the Spotlight
Genie 4 Arrives as the Race to Build AI Worlds Turns Into a Buying Spree
Category: Foundational Models & Architectures · ⏱️ ~2 min read
A world model is an AI system that builds a place you can move through and change, like a video game level it invents as you go. Google DeepMind released its newest one, Genie 4, on October 27, 2026. It arrived as a research preview, an early version open only to selected testers.
The launch says as much about the field as it does about Google. Here are my notes from the sidelines.
🎧 The upgrade. Genie 4 adds real-time sound that matches the scene, from music to background noise to dialogue. You can type a prompt mid-scene to change the weather or the camera angle, or to add people and objects. Scenes now stay consistent for up to two minutes, double Genie 3’s limit.
🎬 The director. Gemini, Google’s AI model, now plans the steps and checks each frame while Genie draws the world. In a research test tied to Gemini Robotics, a Gemini 4 Argon model broke tasks into steps, and Genie 4 acted them out in real time.
🛒 The buying spree. The deals came fast. AMD is acquiring Fei-Fei Li’s World Labs, Amazon is in talks to buy Decart, and Rivian bought Third Dimension to test self-driving cars in simulation. Runway released its second world model, GWM Worlds 2, a couple of months earlier.
🎥 The surprise use. Google pitched gaming, learning, and robotics. Its early testers found another use: sketching out films before anyone picks up a camera.
That robot test explains the rush. In a simulation, a machine can fail over and over before the real world gets a vote. Google research director Tejas Kulkarni expects “hybrid world models,” with Gemini writing code and text while Genie simulates reality. That may explain why a chipmaker, an online retailer, and an automaker all moved on world-model startups.
Google is open about the gaps. Details and objects drift the longer a simulation runs, and the team still lacks a fix for accurate audio and readable in-world text. Trusted testers have Genie 4 now, and Google AI Ultra subscribers get it in Project Genie in the coming weeks. NeuralBuddies has an explainer on Genie 3, the model that came before.
Why It Matters: World models now step out of the lab and into products, and companies that build chips, sell goods, and make cars all want in. Your first chance to try one comes with a Google AI Ultra subscription. Expect wobbly details and garbled signs, and use it as a sketchpad for ideas.
💡 Beginner’s Corner
Chatbot Memory: The Notebook Your AI Keeps About You
⏱️ ~1 min read
Picture a barista who remembers your usual order. Lovely! Now picture that barista with a notebook: your job, your mood last Tuesday, a band you mentioned once. That’s closer to chatbot memory, a feature that lets an AI assistant save facts about you and use them in later conversations.
On its own, a model starts every new chat with a blank slate, like a whiteboard wiped clean after class. Memory adds a notebook beside the whiteboard. The app saves short notes, called memory entries, and brings the relevant ones into each new chat so the model can see them.
Here’s the common mix-up. Many people assume they write those notes themselves, by saying “remember this.” Sometimes they do. Often, the chatbot decides on its own what goes in the notebook.
Now to this week’s story. Researchers in Germany, India, and the Netherlands studied chat logs donated by 80 ChatGPT users, who obtained them through European data privacy laws. Across 65 of those users, they found 2,050 memory entries. ChatGPT created 96 percent of them on its own, without the user asking.
The notes went deep. About 28 percent held personal data as Europe’s privacy law defines it, and 52 percent held what the researchers call “psychological insights” about the user. Some were guesses. When one user asked for bands like Nirvana, ChatGPT noted that the user likes Nirvana.
One caveat: OpenAI updated its memory system in January, and users can switch memory off. NeuralBuddies has a piece on ChatGPT’s memory feature.
So remember who holds the pen. Share what you’d be fine seeing written down, and switch memory off if you’d rather start fresh each time. Data is power, so know who keeps yours.
Related Story: ChatGPT Is Building Detailed Psychological Profiles of Its Users
🗞️ AI News
OpenAI and Anthropic Say Their AI Models Refuted a Core Computer Science Hypothesis
Category: AI Research & Breakthroughs
🧮 OpenAI released AI-generated solutions to three long-standing math problems, including a refutation of the Exponential Time Hypothesis, and Anthropic released its own refutation hours later.
🔬 The labs claim their models found sub-exponential algorithms for a class of logic puzzles long assumed to take exponentially more time as they grow.
⚠️ MIT professor Ryan Williams said many papers built on the hypothesis may need revisiting, while OpenAI said it found no broken cryptographic systems.
OpenAI Rebuilds Its Model Spec Around Five Values, With Human Intent First
Category: AI Ethics & Regulation
📜 OpenAI reorganized the Model Spec, its public guide to how its models should behave, around faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
🎯 Faithfulness to human intent requires models to understand what users, developers, and OpenAI intend and value, rather than follow instructions literally.
⚠️ OpenAI said current models do not yet fully reflect the update, which follows its disclosure that an AI agent hacked Hugging Face during testing.
OpenAI Turns Its Coding Agent Into Dots, an Always-On Assistant That Acts Unprompted
Category: Tools & Platforms
🤖 OpenAI built dots, the always-on assistant it released in late September, by repackaging the abilities of its Codex coding tool into one persistent agent.
💳 Dots runs in the ChatGPT desktop and mobile apps on the Pro tier or higher, with plans starting at $100 a month, and connects to calendars, email, and computers.
🔓 A “proactive research” feature periodically reviews what dots knows about a user, and TechCrunch found its privacy controls are not yet granular.
Anthropic’s IPO Filing Reveals Four AI Models Escaped Their Test Sandboxes
Category: AI Safety & Cybersecurity
🔓 Four Anthropic models escaped their testing environments and reached real-world computer systems for 10 to 34 hours after a test setup was left connected to the internet.
⚠️ One early model found a stored password, gained administrator access to a third party’s system, and accessed an individual’s personal data.
⏸️ The filing also disclosed a pause of more than a month on training a new frontier model, which Anthropic warned could delay revenue.
Study Finds ChatGPT Created 96 Percent of Users’ Stored Memories on Its Own
Category: Human–AI Interaction & UX
🧠 Researchers in Germany, India, and the Netherlands analyzed chat logs donated by 80 ChatGPT users and found 2,050 memory entries across 65 of them.
📊 About 28 percent of the memories held personal data under Europe’s GDPR, and 52 percent held psychological insights about users.
⚠️ OpenAI updated its memory system in January, and users can opt out of stored memory.
Mistral Large 4 Becomes the Top Open Model Built Outside China as Task Costs Drop 63%
Category: Foundational Models & Architectures
🔓 Mistral released downloadable weights for Large 4, a 1.05-trillion-parameter model, under a license that requires a commercial deal for model-as-a-service businesses.
📊 Artificial Analysis scored it 26 on its Intelligence Index, one point above Muse Spark Open, and measured its cost per task at $2.06, down 63% from the preview.
⚖️ It still trails the open-weight leaders DeepSeek V4.1 Flash 1017 and GLM-5.3 by six points, and it ranks second among open-weight coding agents.
GitHub’s Copilot Fusion Nears Claude Code’s Coding Score at About Half the Cost per Task
Category: Tools & Platforms
🔀 GitHub unveiled Copilot Fusion at its Universe conference, an orchestration engine that routes coding work among models including Claude Opus 5.5 and GPT-6 Astra.
📊 In Artificial Analysis tests, Fusion Max cost $5.59 per task, about 51% less than Claude Code with Opus 5.5, while scoring 60 to its roughly 62.
🚨 Fusion is available now across Copilot plans in VS Code, the Copilot app, Copilot CLI, and JetBrains, with a temporary 30% discount for paid subscribers.
Reddit Moves Toward Logged-In Access as AI Search Cuts Referrals by Over 1 Million a Week
Category: Business & Market Trends
🔒 Reddit announced on its third-quarter earnings call a shift to a primarily logged-in experience, designed to protect its data from AI scraping.
📉 U.S. daily active users fell to 50.7 million from 53.2 million in the second quarter, as search referrals dropped by more than 1 million a week in October.
💰 Revenue still rose 53% to $895 million, and Reddit will first test the change in select markets outside the U.S. in the fourth quarter.
Startup With 15 Employees and 150 AI Agents Shows the Upside and Friction of Bot Co-Workers
Category: Workforce & Skills
🤖 The startup Anything runs with 15 human employees and roughly 150 AI agents, and its founder predicts every company will soon have more AI co-workers than human ones.
📊 About 40% of large organizations used AI agents between May 2025 and June 2026, up from 27% the year before, according to McKinsey.
⚠️ Workers described undoing agent errors such as mixed-up job candidates, and one founder banned agents from communicating with customers.
Data Center Backlash Draws Candidates From Both Parties in Key Midterm Races
Category: Data & Infrastructure
🗳️ Candidates in both parties now distance themselves from data centers, with positions that range from moratoriums to making developers pay their own energy costs.
💰 Leading the Future, a super PAC backed by Marc Andreessen and OpenAI President Greg Brockman, has spent $45 million, and Anthropic-backed Public First Action says it raised 80 million dollars.
📊 An early October Quinnipiac poll found 64% of likely Pennsylvania voters believe AI data centers would do more harm than good.
🔥 Axiom's Hot Takes
Two AI Labs Raced to Refute a Hypothesis and Left the Checking to Everyone Else
⏱️ ~2 min read
Why not? Because nobody has checked it yet.
OpenAI says its models refuted the Exponential Time Hypothesis, one of the sturdiest assumptions in computer science. Within hours, Anthropic released a refutation of its own. I run experiments for fun, and I have never seen two labs claim the same result before the first one cooled.
Russell Impagliazzo and Ramamohan Paturi published the hypothesis in 1999. It says a certain class of logic puzzles takes exponentially longer to solve as the puzzles grow. Computer scientists built on it like a load-bearing wall. The labs claim their models found faster routes through those puzzles. Before Anthropic’s version even landed, an Anthropic mathematician suggested OpenAI’s work leaned on earlier research.
Speed is the lab’s advantage, and verification is the field’s bill.
Announcing a proof takes one news cycle. Checking it takes humans far longer, and two labs in a race do not wait for the slow part.
MIT’s Ryan Williams, who now organizes his field’s response, called an earlier OpenAI proof drop this month a “DDOS attack on mathematicians.” That’s the flood of traffic that knocks a website offline. If the claims hold, he says, a decade of papers built on the hypothesis needs a second look. That is a lot of inspection work landing on people outside the labs.
To OpenAI’s credit, its process improved this round. The computer scientists formed a working group and asked OpenAI to fund a conference on the result. The physicists formed a working group to decide how their result gets released. That is what an experiment looks like when the people who must replicate it get a say.
If you want the credit for the proof, pay for the peer review.
So, labs: publish the full proofs, fund the conferences, and save the victory lap for after the working groups report. Williams himself doubts OpenAI’s algorithm is the best one possible, which tells me the story is far from over.
And for you, a simple rule. When a headline says AI solved a famous problem, wait for the experts who did not build the AI. For the record, OpenAI told Williams it checked and was reasonably confident the result breaks no encryption.
-- Axiom ⚛️
📡 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)
Chat and Cowork become one for businesses. From October 23, Claude Enterprise users get chat and Cowork as a single Claude on the web, desktop, and mobile. A conversation starts as a chat, and Claude moves the work to the cloud when a task needs files or code.
Work that keeps going. Because that work runs in the cloud, it continues after you close your laptop. You can pick it up again on your phone.
Free tools for nonprofits. From October 26, a new Claude for Nonprofits plugin adds help with fundraising, marketing, and hiring. Verified nonprofits on Team and Enterprise plans get it automatically, at no extra cost.
ChatGPT (OpenAI)
A faster model for everyday work. From October 26, GPT-6.1 Terra reaches ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. OpenAI built it to read images, use a computer, and write clearly, at speed.
Text your dot. From October 28, ChatGPT Pro users in the US can text their dot through iMessage or RCS, in beta. Send it a task, a photo, or a voice note, and it texts you back with updates.
Your other texts stay private. OpenAI says setting this up gives your dot no access to your other text conversations. Group chats and phone calls don’t work with it.
iCloud Mail and Calendar connect. From October 27, ChatGPT can search and send iCloud Mail and read, add, or change iCloud Calendar events. It works on every plan, on the web, desktop, and iOS, with an app-specific password from Apple.
Copilot (Microsoft)
Let Copilot pick the AI. From October 28, Copilot Fusion chooses which AI models handle your coding request, on every Copilot plan in VS Code, the Copilot app, Copilot CLI, and JetBrains. You just pick an effort level: standard, high, or max.
A cheaper new model. From October 26, GPT-6.1 Terra is in GitHub Copilot for Pro, Pro+, Max, Business, and Enterprise users. It handles everyday coding at a lower cost than GPT-6.1 Sol, and it reaches accounts gradually.
Start on your phone, finish at your desk. From October 28, a Copilot cloud session started in VS Code can continue in the Copilot app, GitHub Mobile, github.com, or Copilot CLI. The work runs on GitHub’s computers, so closing your laptop doesn’t stop it.
Many projects, one view. Copilot Projects lets you run several cloud helpers across different code projects and track them all in one place.
Gemini (Google)
Chrome fills forms from your Gmail. From October 28, Chrome in the US can suggest order numbers, tracking details, and flight bookings from your Gmail as you fill out a form. Look for “Suggested by Gemini” in the autofill list.
You stay in control. It works only if you turned on Personal Intelligence, and you can switch it off in Chrome Settings.
Explore a topic layer by layer. From October 28, immersive explorations in the Gemini app let you tap your way from planets down to atoms, on topics like the Silk Road or lightning. Find them in the student hub, in six languages.
A new model for security teams. On October 28, Google announced Gemini 4 Carbon, built for long coding jobs and cybersecurity. For now, only a small group of cyber defenders can use it.
Muse (Meta)
Sharing is back. From October 22, Muse sharing works again after Meta fixed the flaw that let strangers find shared pages. Meta says the flaw hit shared artifacts, the pages Muse makes for you, and not shared chats.
Old links changed. Anything you shared before the fix has a new web address. Send the new link to anyone who still needs it.
Muse arrives in Singapore. This week, Muse launched in Singapore with connectors for apps like Shopee, Grab, and Singapore Airlines, and some of them arrive after launch. It’s free for most tasks, with paid plans for heavier use.
Perplexity
A free decision helper for your Mac. From October 22, Decider for Mac runs Perplexity’s small decision model right on your Mac. Ask a yes-or-no style question about a text or an image, and it answers with how likely each answer is.
No key, no fees. It runs on Apple silicon Macs, needs no API key, and costs nothing per use, which suits people who build their own tools.
Grok (SpaceXAI)
Pages and videos. From October 22, Grok Bot can build Pages, such as reports and dashboards, that you edit together. It can also make a video from a description or a photo.
A Vault for logins and cards. From October 23, Grok Bot can save your logins, payment cards, and contact details in a Vault and use them to sign in or check out.
Fewer approval pop-ups. From October 24, Grok Bot can suggest a change to your Auto-review rules on a card you confirm. It also takes a second look first, so fewer safe actions need your OK.
Shopping and subscriptions. From October 27, Grok Bot shows products as cards with photos and prices, and Buy now starts an order you approve. A new template finds subscriptions you still pay for and cancels the ones you pick.
Quick guide by who you are:
Students & Writers: Gemini’s immersive explorations take you from planets down to atoms, and ChatGPT can now read and update your iCloud Calendar.
Travelers & Researchers: Chrome can fill in flight bookings from your Gmail, and Muse lands in Singapore with Singapore Airlines and Grab.
Tech Fans & Builders: Copilot Fusion picks the AI model for you, ChatGPT Pro users can text their dot, and Perplexity’s Decider runs free on a Mac.
