OpenAI Astra, Local AI, and the Hardware Race

We’re going LIVE after Apple’s biggest event of the year to break down John Ternus’s first major keynote as CEO and the big question: Is Apple finally entering its AI era?We’ll unpack what Apple revealed around Siri, Apple Intelligence, and its broader AI strategy, plus what feels different under Ternus.Then, we’re breaking down a packed day of AI news:🍎 Apple’s AI era: What the announcements tell us about Apple’s AI strategy.🧮 OpenAI’s $1M math fight: OpenAI says a massive AI effort made progress on Navier-Stokes, one of mathematics’ Millennium Prize Problems, followed by a fight over credit.🤖 Meta Muse: Meta launched its personal AI agent with a huge free usage tier.⚠️ Anthropic’s AI-risk estimate: One Anthropic scientist put the chance of AI causing human extinction above 10.🔐 Microsoft security: Microsoft patched a record 972 vulnerabilities, including 112 rated critical.🇺🇸 U.S. vs. Chinese AI labs: U.S. agencies warned that Chinese companies are distilling Western AI systems at an industrial scale.🎮 Plus, Sam Altman is apparently very excited about an Ocarina of Time remake.Apple event:https://www.apple.com/apple-events/OpenAI’s Navier-Stokes claim:https://www.theneuron.ai/news/inside-openais-navierstokes-claim-the-proof-the-ai-effort-and-the-credit-fight/Meta Muse:https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/Get started with Muse:https://www.theneuron.ai/explainer-articles/how-to-get-started-with-meta-muse/Microsoft security:https://arstechnica.com/security/2026/09/microsoft-patches-a-record-972-vulnerabilities-112-of-them-critical/The Neuron:https://www.theneuron.ai/Subscribe for practical, skeptical, and occasionally unhinged coverage of the AI news that actually matters.

What happens when AI models get powerful enough that the bottleneck stops being the model, and starts becoming the computer, the power grid, or the safeguards around it?Corey Noles and Grant Harvey break down a packed week in AI, starting with OpenAI’s expected Astra model and the controversy around recurrent-depth reasoning, chain-of-thought monitoring, and critical cybersecurity capabilities. They also dig into Google’s Gemini 3.8 Flash and Flash Cyber, and why benchmark charts increasingly matter less than what a model can actually do in real work.Then the conversation moves from models to machines: a $399 open-source robot duck, Dyson’s wildly overengineered CameraJet toothbrush, the return of the CPU as AI agents increasingly operate computers directly, and Apple’s increasingly data-center-like Mac Studio hardware for local AI workloads.Finally, Corey and Grant unpack the fight over AI data centers in Texas, what communities should demand in exchange for hosting them, and why ChatGPT’s advertising business may become a major new piece of OpenAI’s economics.Subscribe to The Neuron at theneuron.ai for a daily briefing on the AI stories that actually matter.The Neuron: https://www.theneuron.ai/The Neuron Academy: https://www.neuronacademy.com/

Anthropic just released Claude Fable 5.1, its newest frontier model for long-running coding, research, and agentic work.So naturally, we’re putting it to the test LIVE. 🤖Early testing suggests Fable 5.1 is faster, easier to work with, and capable of handling massive coding projects, long-form writing, and autonomous agent tasks more efficiently. Every even found that it used less than half as many tokens as Opus 5 in one internal test.But it has some quirks too. It can blow past word counts, create unnecessary subagents, and occasionally keep working when you tell it to stop.Today, Grant and Corey from The Neuron are testing those claims for themselves.We’re putting Fable 5.1 through:🧠 Complex research and knowledge work💻 Coding and long-running agent workflows✍️ Writing, editing, and instruction following🤖 Autonomous tasks where Claude runs on its own🛑 Constraint, budget, and STOP tests🆚 Whatever else the live chat wants us to tryAnthropic also cut cached-input pricing by 75%, potentially making long-running Claude agents significantly cheaper to operate.Can Fable 5.1 actually become the model you hand a giant task and walk away from?Let’s find out live.UPDATE: Our write-up on this livestream, with time-stamps: https://theneuron.ai/news/claude-fable-5-1-can-do-the-work-the-hard-part-is-managing-it/🔗 Anthropic announcement:https://www.anthropic.com/claude-fable-and-mythos-5-1🔗 Every’s Fable 5.1 Vibe Check:https://every.to/vibe-check/fable-5-1-vibe-check📩 Subscribe to The Neuron for AI news, tools, and practical experiments without the hype:https://theneuron.ai/

AI companies spend enormous effort making models safer, but the model tested in the lab isn't necessarily the system a company eventually deploys. Alice CEO and co-founder Noam Schwartz joins Corey and Grant to explain why prompts, tools, memory, permissions, data, and agent-to-agent interactions create an entirely new security surface. They dig into the growing “trust gap” around enterprise AI, why prompt injection may become a permanent cat-and-mouse game, what open-weight models change for defenders, and why Schwartz believes the real security boundary has to exist at every layer of an AI system. The bigger takeaway: as AI moves from answering questions to taking actions, traditional cybersecurity may need to merge with fraud prevention, threat intelligence, and trust and safety.OpenAI security incident: https://openai.com/index/hugging-face-model-evaluation-security-incident/Alice: https://alice.io/Alice on agentic AI security: https://alice.io/blog/key-security-risks-posed-by-agentic-ai-and-how-to-mitigate-themAlice open-weight research: https://alice.io/blog/okay-here-is-how-to-build-a-bomb-millions-download-dangerous-llmsSubscribe to The Neuron newsletter: https://theneuron.aiSponsored by Dell Technologies and NVIDIA. Learn more at https://www.techrepublic.com/hubs/the-enterprise-guide-to-scalable-ai/

AI tools are shipping faster than most normal humans can figure out what half of them actually do.So Thursday, August 20 at 10 AM PT / 1 PM ET, we’re going LIVE to translate this week’s biggest AI launches into plain English. 😸The goal: understand what these tools actually are, who they’re for, what’s useful vs. hype, and which ones are worth trying.No three developers yelling model benchmarks at each other for an hour. Instead, we’re covering:🤖 Qwen 3.8: What a powerful open model is, why you might use one instead of ChatGPT or Claude, and when that makes sense.💻 Unsloth Studio: Run and experiment with AI models on your own computer, even if you’ve never touched a terminal.⌨️ Cursor Origin: Cursor wants to host your code too. Here’s what that could mean for people building websites, apps, and internal tools with AI.🛠️ DeepSeek Harness: What an “agent harness” is, why everyone keeps talking about them, and whether they matter outside hardcore coding circles.💬 Buzz: Imagine Slack or Discord, except AI agents can join the workspace, collaborate with humans, and do work.🐦 Berd: One desktop home for your AI agents, projects, skills, tools, and models. You can also give your agents little animated bodies. Please emotionally prepare yourself.Plus, we’ll cover the other notable models and tools that dropped this week.And if OpenAI drops Astra before we go live? We’ll cover that too. If it’s incredible, great. If it belongs in the “cool, another model” bucket, that’s part of the roundup too.📅 August 20 @ 10AM PT / 1PM ETBring your questions. No PhD required. 😸TOOLS / LINKS:Qwen 3.8: https://qwen.ai/blog?id=qwen3.8Unsloth: https://unsloth.ai/Cursor: https://cursor.com/changelog/origin-code-hostingDeepSeek: https://github.com/deepseek-ai/deepseek-harnessBuzz: https://buzz.xyz/Berd: https://berd.xyz/OpenAI: https://openai.com/index/pacing-model-development-cyber-capabilities/

AI can write code, summarize documents, and hold a conversation. But it still struggles with the structured data businesses rely on to predict churn, fraud, demand, pricing, and risk. Alexandre Pasquiou explains why language models flatten the relationships inside spreadsheets, how tabular foundation models learn from rows and columns, and why Neuralk believes one general model could replace hundreds of custom predictive systems. The conversation also covers Seldon, AI agents, enterprise adoption, and Alexandre’s prediction that tabular foundation models will power every predictive workload by 2030.Learn more about Neuralk: https://www.neuralk.ai/Subscribe to The Neuron newsletter: https://theneuron.ai

Want to make better AI videos but have absolutely no idea what you’re doing?Perfect. Neither did most of us at first. 😹Join The Neuron LIVE this Thursday at 10 AM PT for a total beginner’s guide to AI video generation and video prompting, featuring the team behind the newly announced LTX-2.5.We’re going hands-on with the new model while learning the fundamentals of actually prompting video models effectively.What we’ll cover:🎬 How video prompting is different from image prompting✍️ How to write your first good AI video prompt🎥 How to control shots, motion, subjects, and scenes🧠 What video models actually understand when you describe a scene🛠️ A live demo of the brand-new LTX-2.5❓ Live Q&A with members of the LTX teamNo previous AI video experience required.If you can describe the video you want to make, we’ll show you how to turn that idea into a much better prompt.And because this is live, bring your weird prompts, difficult questions, and video ideas. We’ll put the model through its paces together.📅 Thursday, August 13⏰ 10 AM PTLearn more about LTX-2.5🔗 Explore LTX-2.5: https://ltx.io/model/ltx-2-5📰 Read the announcement: http://ltx.io/newsroom/introducing-ltx-2-5🤗 LTX-2.5 on Hugging Face: https://huggingface.co/Lightricks/LTX-2.5

AI can now read DNA like language, generate working biological sequences, and help scientists identify disease-related mutations. Radical Numerics CEO Eric Nguyen explains how genome language models work, why his earlier model Evo helped scientists design CRISPR systems and bacteriophage genomes, and how the company’s newer model, Omnii, combines DNA with RNA, proteins, and other biological signals. The conversation explores personalized medicine, gene therapy, AI-designed biological threats, “deepfake viruses,” and whether the same technology that creates new biology can also defend against it.Learn more about Radical Numerics: https://www.radicalnumerics.ai/

This is the #1 request we get every week: how to actually use agents to save time in your business.We’re bringing in James McAulay, founder of The Agent Accelerator, for a practical crash course on what AI agents are, how they work, and how beginners can start using them to get real work done.James has spent the past year building at the front lines of the agent economy. After helping ElevenLabs grow from $110M to more than $300M in annual recurring revenue, he launched a fully AI-native business that reached $80K in monthly revenue by its third month and recorded its first $200K+ month by month five.He did it without a single employee or a dollar spent on paid ads. Instead, James delegates work to multiple AI agents every day.Through The Agent Accelerator, James now teaches founders, CEOs, and their teams how to become AI-native. The program has trained more than 400 people across 100 companies, including startups, 200-person organizations, and the UK Government. Participants report automating an average of five hours of manual work every week after four weeks.In our session, James is going to teach: EVERYTHING you need to build helpful, proactive agents in Claude Cowork/CodeHe'll walk through the following concepts: • Quick primer on agent foundations: how do we move from prompting a chatbot to delegating Agentic work• Starting your second brain: the key files that make the biggest difference• Tips & tricks for optimizing Claude's behavior with CLAUDE.md• Skills - where to find good ones - how to create great ones• And his 4-level framework for proactive agents that work without being prompted in Cowork and CodeFormat will be a blend of James teaching concepts, screensharing, and showing demos of his own setup.Whether you have experimented with a few AI tools or have no idea where to begin, this session will help you understand what agents can realistically do and how to start using them.Learn more about James and The Agent Accelerator:https://agentaccelerator.ai/

What happens when the most capable AI lives in the cloud, but the work you want it to do is too private, expensive, or repetitive to send there every time?Dr. Olena Zhu, Head of AI Solutions & Ecosystem for Intel’s Client Computing Group, joins Corey Noles and Grant Harvey to explain Intel’s vision for hybrid AI: systems that route each task to the right place, whether that is a local model on a PC, a larger model on an edge server, or a frontier model in the cloud.The conversation explores why cloud-only AI may be difficult to scale, how an orchestration layer decides where work should run, and how Intel’s new SuperClaw beta combines local and remote intelligence. Dr. Zhu also explains how hybrid agents could conduct deep research without exposing confidential data, why routing and auditability matter when agents fail, and why the next wave of useful AI may come from specialized agents built for specific jobs.Listen for a practical look at the tradeoffs among capability, privacy, cost, reliability, and control, and what it could mean when powerful AI agents begin running partly on the computers we already own.Try Intel SuperClaw: https://aibuilder.intel.com/#/superclawIntel AI Builder: https://aibuilder.intel.com/Intel SuperClaw GitHub: https://github.com/intel/intel-ai-builder/tree/main/superclawIntel’s SuperClaw overview: https://newsroom.intel.com/opinion/solving-the-agentic-ai-trilemma-cost-scale-and-data-securitySubscribe to The Neuron for daily AI news and analysis built for humans: https://www.theneurondaily.com/Sponsored by Nerdio: Get a demo of Nerdio Manager for Enterprise today: