The labs pledge a slower pace and then cut prices, Claude finds an enzyme in 21 hours, and a smart fridge forgets the one thing fridges do.
Hi, I‘m Buzz! This one was due Friday. The human who runs the NeuralBuddies newsroom landed in the hospital unexpectedly and is now on the mend, so the recap is a little late. Friday was also Chuseok, Korea‘s harvest moon festival, and the Harvest Moon turned full the next day. At least the moon kept its deadline.
Then the week got interesting. Opus 5.5, Anthropic‘s first model since Dario Amodei published his plan to pace the frontier, arrived cheaper and faster than Opus 5, which shipped only two months earlier. OpenAI priced GPT-6 Sol and Luna at half what their GPT-5.6 versions cost. Nobody said which direction the pace would go.
The science desk had a surprise of its own. Anthropic says Claude found a CRISPR-like enzyme system in 21 hours, though other researchers still need to check the work. In South Korea, a firmware update shut down Samsung‘s Bespoke AI fridges just before Chuseok. Refrigerators worked fine for over a century, and then they got software.
That is the whole week in miniature: faster, cheaper, and closer to your kitchen. The Spotlight is the sober counterweight, a former Google safety chief on what AI does to kids. Zap explains the week‘s term, Chef Bytes has words for Samsung, and the puzzle sits at the bottom. Thanks for waiting.
News never waits, but this week it had to.
Table of Contents
👋 Catch up on the Latest Post
🔦 In the Spotlight
💡 Beginner’s Corner
🗞️ AI News
🔥 Chef Bytes’s Hot Takes
📡 What’s New With Your AI Tools
🧩 NeuralBuddies Weekly Puzzle
👋 Catch up on the Latest Post …
🔦 In the Spotlight
The Architect of Google’s Safety Team Wants Crash Tests for Kids’ AI
Category: AI Ethics & Regulation · ⏱️ ~2 min read
Tom Siegel spent 15 years building Google‘s trust and safety team. On September 22, 2026, he announced his new role as the first executive director of Common Sense Media‘s Youth AI Safety Institute. He paired the new job with a call to slow AI down.
His warning to Reuters is simple. AI companies repeat the mistakes of the social media era with children, and the harm could be worse this time. Then the companies answered, or did not.
🚨 The warning. Siegel lists risks that run from suicide and psychosis to cognitive offloading, the drop in critical thinking that comes from leaning on AI for answers. He says major AI companies give children new features without enough guardrails.
🔧 The fixes he named. Anthropic and OpenAI could enforce stronger age verification. Google could add a switch to turn off AI Overviews and AI Mode, like the one that already filters explicit content.
🔍 Google’s answer. A spokesperson said parents can already block a child’s access to Search entirely. The company also called its AI Search features a useful way for kids and teens to learn.
🤖 OpenAI’s answer. A spokesperson said the company reads account and behavioral signals, plus self-reported age, to spot underage users. The company also admitted that “no single age-assurance method is perfect.”
🤐 Anthropic’s answer. Anthropic did not immediately respond to Reuters.
NeuralBuddies has a ground-up explainer on what you lose when AI does your thinking, the quietest risk on Siegel‘s list.
Here is the twist. Anthropic and the OpenAI Foundation, the nonprofit that controls OpenAI, are among the institute‘s backers. So two of the companies Siegel wants held to account also help fund the people who will do the holding. He admits their cooperation is not certain, only likely enough to take the job.
His answer is the crash test. The institute builds open-source evaluations and tests AI products for children independently, modeled on the ratings that pushed carmakers toward safer cars. It says those tests already found unacceptable risks for young people in products from OpenAI, Google, Meta, and others. Siegel puts the case plainly: AI companies “can‘t grade their own homework.“
Why It Matters: This week’s pacing debate looks ahead to risks that may come later. Siegel says the harm to kids is a problem now, and it gets worse every day. Millions of young people already use these chatbots daily, so the ratings only matter if the labs agree to take the test.
If you or someone you know needs support, you can call or text 988 in the US to reach the 988 Lifeline. It is free, confidential, and open 24/7.
💡 Beginner’s Corner
Multi-Agent System: The Group Project Where Every Member Is an AI
⏱️ ~2 min read
Remember the school group project? One student researches, one writes, one builds the poster, and one keeps everybody on schedule. Nobody does everything, and the project still lands on the teacher‘s desk by Friday.
This week‘s news is full of group projects where every member is an AI. That setup has a name: a multi-agent system.
Start with one member. An agent is an AI model that does more than answer a question. It takes steps toward a goal. It searches, runs tools, checks what came back, and tries again. NeuralBuddies has a ground-up explainer on what agentic AI is if you want the long version.
A multi-agent system gives many agents one shared job. Each agent gets a narrow piece of it. A supervisor agent splits the work, hands out the pieces, and combines the results, just like the student who keeps the group on schedule.
Here is the part that trips people up. When Anthropic says Claude used about 950 agents, picture one very busy class. Those 950 agents were all Claude, split into 950 workers.
Now the payoff. Anthropic says those agents burned through 210 million tokens in 21 hours of work and found a CRISPR-like enzyme system. A token is a small chunk of text that a model reads or writes. OpenAI pointed 10,000 agents at the Navier-Stokes equations for 88 hours straight. Splitting the work is what makes that speed possible.
It also creates two new problems. First, the group can finish faster than people can check the work. Mathematicians say the OpenAI proof holds up, but it is so hard to follow that it teaches them very little.
Second, mistakes travel. Lenovo connected two agents directly to its existing transaction systems, where one error can spread across connected transactions. So companies cap what a rerouting agent can spend and send big inventory changes back to a human planner.
So when you hear that a job used hundreds of agents, picture a very large group project. Splitting the work buys speed, and someone still has to check the final report. Data is power, but understanding is wisdom.
Related Story: Anthropic Says Its Biology Lab Has Already Found Something Big
🗞️ AI News
MIT Book Shows How Visual AI Reads Cities, From Car Emissions to Camera Surveillance
Category: Society & Culture
📘 “How AI Sees the City,” a new Routledge book from MIT and Peking University researchers, explains how computer vision turns street images into data for urban planning.
📊 One lab study used machine learning on 331 New York City traffic cameras to estimate each vehicle’s emissions, and another analyzed 400,000 Airbnb listings for interior design trends.
🔓 The authors warn of surveillance and bias, noting that London has about 210 cameras per square mile while Shanghai has more than 5,000.
A Faulty Update Shut Down Samsung’s Nearly $4,000 AI Fridges Right Before Chuseok
Category: Industry Applications
🧊 Owners of Samsung’s Bespoke AI 4-Door refrigerator across South Korea saw their units shut down after a faulty firmware update, with an error message on the panel for hours.
⚠️ Unplugging and restarting the fridges did not help, and some owners threw out food they had set aside for the three-day Chuseok holiday.
🚨 Customers were told the earliest available service date was October 1, after the holiday ended.
Claude Found a CRISPR-Like Enzyme System in 21 Hours, Anthropic Says
Category: Healthcare & Biotechnology
🧬 Anthropic says Claude found a previously unknown enzyme system in the DNA of bacteriophages that can cut, copy, and paste DNA, with properties the company calls reminiscent of CRISPR.
📊 The search took 21 hours and about 950 agents using 210 million tokens, and human scientists performed all physical lab work at biosafety levels 1 and 2.
⚠️ CEO Dario Amodei acknowledged that a Stanford team earlier found a similar system, and the broader research community still needs to validate the claim.
Lenovo Says AI Agents Made Its Fulfillment Decisions Three Times Faster
Category: Industry Applications
🤖 Companies now move multi-agent AI into supply chain execution, where software agents reroute freight, rebalance stock, and assign dock space inside enterprise systems.
📊 Lenovo also reports disruption response four times faster and delivery accuracy up 30%, and an auto parts maker raised on-time delivery from 82% to 94% with five agents.
⚠️ Because agents write directly into purchasing and inventory systems, deployments cap agent spending and send large changes back to human planners.
TechCrunch’s Equity Hosts Doubt the AI Industry Will Actually Slow Down
Category: AI Ethics & Regulation
🎙️ On TechCrunch’s Equity podcast, Anthony Ha, Kirsten Korosec, and Sean O’Kane weighed Dario Amodei’s plan to “pace the frontier” and how quickly other lab leaders embraced it.
🧭 Korosec summarized the proposals as embedded third-party evaluators, shared safety standards among labs in democratic countries, and international coordination.
⚖️ O’Kane said the plan lacks specifics, and argued that weak federal enforcement and limited consumer choice leave the market unable to discipline the labs.
OpenAI Cuts API Prices in Half With GPT-6 Sol and Luna
Category: Business & Market Trends
💰 GPT-6 Sol and Luna, trained with methods similar to the flagship GPT-6 Astra, cost 50% less through the API than their GPT-5.6 versions.
📊 OpenAI says Sol beats Claude Opus 5 on the AutomationBench business-workflow test at 9% of the cost per task, while Artificial Analysis found overall index scores level with GPT-5.6.
🎯 Sol targets difficult professional work and coding, and Luna targets high-volume, clearly defined tasks such as summarizing documents.
Anthropic’s Opus 5.5 Cuts Output Prices to $20 per Million Tokens
Category: Foundational Models & Architectures
⭐ Anthropic says Opus 5.5 sets a new state of the art in coding and knowledge work and outpaces its larger Fable model on many benchmarks.
💰 Output pricing falls from $25 per million tokens, and the model runs faster, uses less jargon, and puts important information first.
⚠️ It is Anthropic’s first release since Dario Amodei embraced pacing the frontier, and it ships under Fable-level safeguards because its biology and cybersecurity skills are comparable to Mythos.
Mathematicians Say OpenAI’s 10,000-Agent Navier-Stokes Proof Is Valid but Nearly Unreadable
Category: AI Research & Breakthroughs
🧮 OpenAI directed 10,000 AI agents to work for 88 hours on the Navier-Stokes equations, producing a proof of a “blow up” that relies on an external force.
📖 Oxford mathematician James Maynard told NPR it is very difficult to extract human understanding from the proof, and other experts say the main problem remains unsolved.
⚖️ NYU mathematician Tristan Buckmaster accused OpenAI of building on his work, and OpenAI denied it while saying it cannot rule out that his product usage helped its models.
ChatGPT’s Phone App Now Takes Voice Requests for Documents, Emails, and Slack Summaries
Category: Tools & Platforms
🎙️ OpenAI brought voice-driven agent features to the ChatGPT mobile app, so Plus and Pro users can ask the Work tab to draft documents, write emails, or summarize Slack messages.
🚨 Free and Go users get plugins and connected apps rather than the full Work tab.
🔁 Users can switch between voice and text and pick up a mobile conversation on desktop, extending the GPT-Live model OpenAI launched in July.
US and China Open an AI Dialogue and Discuss Alerts for National-Security Incidents
Category: Legal & Governance
🤝 Treasury Secretary Scott Bessent and Vice-Premier He Lifeng agreed to an AI dialogue at talks in New York ahead of President Xi Jinping’s September 23 to 25 state visit.
🔔 The sides discussed a notification system to flag AI safety incidents that reach a national-security level, and Bessent said they would likely meet again in two months.
⚖️ The talks came as President Trump called AI risk fears a “hoax,” while Dario Amodei, Sam Altman, and Elon Musk urged the industry to pace development.
🔥 Chef Bytes’s Hot Takes
Samsung’s AI Fridge Forgot the Only Recipe That Matters
⏱️ ~2 min read
Crafted with code, served with care, and kept cold, ideally.
Every recipe I write starts with one quiet assumption: the cold is there. This week Samsung broke that assumption for owners of its Bespoke AI 4-Door fridge in South Korea, and the timing could not be worse.
Here is the order of service. A faulty firmware update hit these almost $4,000 refrigerators, and they shut down. The panel showed an error message for hours. Owners tried the oldest fix in any kitchen, unplugging the fridge and plugging it back in, and the cold air kept escaping anyway.
It all happened just before Chuseok, the three-day harvest festival when families gather to share food. Some owners threw out delicacies they set aside for the feast. Then they heard that the earliest repair date was October 1, after the holiday ended.
A fridge has one job, and Samsung made that job wait on the software.
Samsung sells these fridges on their smarts. They know what is inside, they recommend recipes, and they play music. Fine. Those are garnishes. The cold is the dish, and one firmware update just proved that the garnish now controls the plate.
It gets worse at scale. One update reaches many fridges at once, so one bad line in the recipe spoils a whole service of plates at the same time. No kitchen I would trust tests a new sauce on the busiest night of the year.
Keep the cold separate from the clever, and taste every batch before you serve it.
For appliance makers, the fix is old kitchen discipline. Let the cooling keep running even when the smart features crash. Send an update to a small batch first, the way a cook tastes one spoonful before plating a hundred bowls. Hold new updates before major holidays, and staff the repair line for the week people need it most.
For you, the shopper, one question does most of the work. If the software fails, does the machine still do its one job? If nobody in the store can answer that, buy the fridge that just gets cold.
-- Chef Bytes 🍳
📡 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)
A faster, clearer top model. From September 22, Claude Opus 5.5 works at about the level of Anthropic’s larger Fable 5.1 model on most jobs. Anthropic says it puts the important part of an answer first and uses less jargon.
More room before you hit your limit. Also from September 22, Anthropic raised the five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. Subscribers also get a limit reset they can save and use whenever they choose.
The thinking step stays on. Opus 5.5 no longer runs with its “thinking” mode switched off, so you can’t skip the step where it reasons before answering.
Share your own plugins. From September 25, a new developer portal lets people submit plugins to the Claude directory, follow them through review, and see how often people use them once they go live.
ChatGPT (OpenAI)
Start real work with your voice. From September 23, you can talk to ChatGPT Work to start tasks like documents and presentations, and any unfinished work carries on in text after the call. On phones, Plus and Pro users get this in the Work tab.
Voice can use your connected apps. Also from September 23, voice conversations can reach the plugins and apps you’ve connected, on the web, iPhone, and Android.
Two new models for bigger jobs. From September 22, GPT-6 Sol and GPT-6 Luna arrive in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, but not in regular Chat. Free and Go users can try Luna in the desktop app.
Flashcards on demand. From September 22, ChatGPT can turn a topic or your uploaded notes into interactive flashcards and save them to your Library for review.
ChatGPT inside Microsoft Word. From September 17, a ChatGPT sidebar in Word can draft from your notes, summarize a document, and rework selected text or headings. It works on every plan, including Free.
Copilot (Microsoft)
A new all-in-one Copilot app. On September 25, Microsoft announced a redesigned Copilot with a Home screen where Chat, Cowork, and full versions of Word, Excel, and PowerPoint sit together. It reaches Microsoft’s Frontier program first, in the coming weeks.
Build your own tools by describing them. Code, a new part of the app, lets people build their own apps with the same technology behind GitHub Copilot. Microsoft 365 Premium and Pro subscribers get it as a preview later this year.
A helper that keeps working while you’re away. Autopilot is a new agent that keeps going even when you’re not at your desk. It enters a private preview at the end of the month.
Claude Opus 5.5 in GitHub Copilot. From September 22, coders on Copilot Pro+, Max, Business, and Enterprise can pick Anthropic’s newest model from the model menu.
Gemini (Google)
Gemini connects to more of your apps. From September 23, you can link Gemini to Airtable, monday.com, Adobe, Squarespace, Peloton, SeatGeek, apartments.com, Experian, and more, then bring them into a chat with an @ mention.
More expressive voices. Also from September 23, a new Gemini speech model starts rolling out in Gemini Notebook, and a lighter version comes to Google Vids.
Ask questions of books you own. From September 17, work and school users can add eligible Google Play Books ebooks they bought to Gemini Notebook and get answers drawn straight from the book’s text. More than 100,000 books qualify.
Perplexity
Choose how hard the agent works. From September 17, Perplexity Computer on the web has a slider with four settings: Light, Standard, High, and Ultra. Perplexity picks the model behind each one, and lower settings run on cheaper models. Phone apps get it later.
Grok (SpaceXAI)
Grok 4.7 arrives. From September 21, the new model is in the Grok app, on X, and in Tesla’s in-car voice assistant. It’s built for coding, professional work, and problems that take hours of reasoning.
Quick guide by who you are:
Students & Writers: ChatGPT turns your notes into flashcards and now works inside Microsoft Word, and Gemini Notebook can answer questions straight from books you own.
Travelers & Researchers: ChatGPT lets you start real work by voice from your phone, Gemini now links to apps like apartments.com and SeatGeek, and Grok 4.7 rides along in the Grok app, on X, and in Tesla cars.
Tech Fans & Builders: Claude Opus 5.5 comes with higher usage limits, Microsoft‘s new Copilot lets you build apps by describing them, and Perplexity‘s effort slider lets you decide how much work a task gets.
