The Neuron
Posts
😺 Sora lasted 6 months. Disney's $1B deal lasted 3.

😺 Sora lasted 6 months. Disney's $1B deal lasted 3.

PLUS: Arm made its first chip in 35 years

Grant Harvey
March 25, 2026

Welcome, humans.

"Is anybody else bored of talking about AI?" That question hit the top of Hacker News yesterday with 358 points and 277 comments. The top response? A senior engineer arguing the people getting the most value from AI are the ones who were already great at their jobs. "Give a great craftsman a new tool and he will find a way to apply it."

Anthropic's own data backs this up. Their 5th Economic Index report found that people who've used Claude for 6+ months have a 10% higher success rate than new users, even controlling for task complexity. The experienced users iterate more, hand over less autonomy, and tackle harder work. So the question isn't whether AI is useful. It's whether you've put in the reps.

Meanwhile, OpenAI decided to announce approximately fourteen things in one Tuesday. So yeah. We're not bored yet.

Here’s what happened in AI today:

🙀 OpenAI killed Sora, prepped a new model called "Spud," and Sam Altman handed off safety oversight to focus on data centers and fundraising
📰 Arm unveiled its first-ever chip after 35 years of only licensing designs, with Meta as launch customer
📰 A federal judge called the Pentagon's actions against Anthropic "troubling," and Meta was ordered to pay $375M in a child safety case
🍪 Anthropic's Claude can now control your Mac while you're away and run tasks from your phone
📖 AI designed new polymers that turned out to be 4x tougher than expected, surprising even the scientists who built them

… and a whole lot more that you can read about here.

P.S: Want to reach 675,000 AI-hungry readers? Click here to advertise with us.

🙀 OpenAI Had the Busiest Tuesday in AI History

You know those days where you clean the house, do laundry, fix the leaky faucet, meal prep for the week, AND reorganize the garage? OpenAI just did that, except the house is a $300B company and the garage is a $1B Disney deal they're lighting on fire.

Here's what happened:

Sora is dead. The standalone AI video app launched six months ago, briefly topped the App Store, and is now being shut down entirely, app and API both. Employees said it was devouring GPU resources during a period of intense competition with Anthropic and Google.
The Disney deal is dead too. Disney's $1B investment and 200-character licensing agreement? Over. Disney said it "respects OpenAI's decision to exit the video generation business." Ouch.
Sam Altman handed off safety and security oversight to focus on raising capital, supply chains, and "building data centers at unprecedented scale."
OpenAI completed initial development of its next major AI model, codenamed "Spud."
ChatGPT's shopping feature flopped. Instant Checkout is being scaled back after users largely ignored it.
OpenAI is raising another $10B, bringing its latest round to ~$120B total.
The OpenAI Foundation, now sitting on ~$130B in equity, committed $1B+ this year to AI-driven scientific discovery and societal resilience.

Why this matters:

This looks like seven separate stories. It's actually one: OpenAI is pruning everything that isn't ChatGPT ahead of its IPO.

Video generation? Gone (too expensive, too many copyright headaches). Shopping? Scaled back (users didn't want it). Safety oversight by the CEO? Delegated (he's building data centers now). What's left is a super-app strategy: ChatGPT + Codex + the Atlas browser, powered by whatever Spud turns out to be.

The Sora shutdown validates what Anthropic bet on all along: focus compute on text and code, not flashy demos. While OpenAI was licensing Disney characters, Anthropic was shipping Claude Cowork and watching software stocks tank.

Our take:

The Disney deal dying is the detail that stings most. Three months ago it was OpenAI's biggest mainstream partnership. Now it's a press statement with the word "respect" doing heavy lifting.

The real question isn't about Sora. It's about Spud. If the next model is good enough, none of today's pruning matters. If it isn't, OpenAI just killed a product, lost a partner, and handed off safety for nothing.

FROM OUR PARTNERS

61% of listeners chose Rime’s AI voices over Google and ElevenLabs.

Rime's AI voices are built for enterprise:

Low latency
On-prem or cloud
Preferred by real listeners in independent testing

If you're building agents that talk to customers, the voice matters.

→ Try Rime free + get $100 in free credits

🎓 AI Skill of the Day: How to Actually Run Open Source AI Models (It's Easier Than You Think)

If you've been meaning to try open source AI models but assumed you needed fancy hardware, Tina Huang just published the best explainer we've seen. Tina is fantastic at making AI easy to understand, and this is simply the best "how to use open models" guide out there.

She ranks four approaches easiest to hardest: Local (download Ollama, pick a model, chat in 2 minutes), Browser/Hosted (arena.ai or Hugging Face Spaces, zero setup), Managed API (get an API key from GroqCloud, call it in 5 lines of code), and VPS ($5-10/month virtual server, full control).

The key insight most people miss: any usable computer can run smaller models. A MacBook Air M4 with 16GB handles 4B and 8B models fine. No GPU required unless you're fine-tuning.

Step 1: Download Ollama from ollama.com
Step 2: Open terminal and run: ollama pull qwen3.5:4b
Step 3: Run: ollama run qwen3.5:4b
Step 4: Start chatting. Everything stays on your machine.

Our favorite insight: lots of people are buying Mac Minis to run models 24/7 without tying up their laptop. Same workflow, dedicated hardware.

Want more tips like this? Check out our AI Skill of the Day Digest for this month.

Have a specific skill you want to learn? Request it here.

New episodes air every week on: Spotify | Apple Podcasts | YouTube

🍪 Treats to Try

Claude Cowork by Anthropic gives Claude access to your local files and apps to complete tasks autonomously; it can now control your Mac while you're away via your phone —included in Pro ($20/mo) and Max plans.
Cloudflare Dynamic Workers sandbox AI-generated code in lightweight isolates that are 100x faster than containers (milliseconds vs. hundreds of milliseconds to start), with no limits on concurrent sandboxes —free for paid Workers users during beta.
ElevenLabs Music Finetunes lets you upload your tracks to fine-tune ElevenLabs' Music model, then generate vocals, instruments, or full tracks that stay true to your sound —no pricing details.
Talat transcribes and summarizes your meetings locally on your machine (nothing goes to the cloud), a subscription-free alternative to notetaking tools like Granola —free.
OpenResearcher takes a plain-text question, searches the web, reads full pages, and builds a cited answer; think deep research but fully open-source (GitHub) —free.

📰 Around the Horn

Arm unveiled the AGI CPU, its first-ever in-house chip after 35 years of only licensing designs. The 136-core processor is designed for AI inference (the part where models answer your questions), with Meta as launch customer and OpenAI, Cerebras, and Cloudflare also signed up.
A federal judge called the Pentagon's actions against Anthropic "troubling," saying "it looks like an attempt to cripple Anthropic." Meanwhile, Apple is testing a standalone Siri app with an "Ask Siri" button for iOS 27.
A supply chain attack was discovered in LiteLLM (a popular AI library with 97M monthly downloads): a malicious file harvested SSH keys and cloud credentials on every Python startup. Karpathy signal-boosted the warning.
Meta was ordered to pay $375M for violating New Mexico law after a jury found it failed to safeguard its apps from child predators. Separately, Baltimore became the first U.S. city to sue xAI over Grok deepfake porn.
Amazon acquired Fauna Robotics, entering the consumer humanoid market. And Figma launched an MCP tool letting AI agents design directly on the live Figma canvas.
Halter, a startup making AI-powered collars for cows, raised $220M. Yes, cows. The future is unevenly distributed and apparently grazing.

Want absolutely EVERYTHING that happened in AI this week? Click here!

FROM OUR PARTNERS

The production-ready agent infra blueprint

AI agents don't fail because the model is wrong, they fail because the infrastructure isn't built for the real world.

Join Redis and Arcade to see how to build enterprise-grade agents using a real loan servicing app as your blueprint. They'll cover auth, policy enforcement, shared memory, and intelligent caching. Everything you need to ship with confidence in regulated environments.

Hear from Arcade's co-founder and Redis' AI field leaders on April 16. Grab your spot.

📖 Midweek Wisdom

This week's must-listen: Latent Space sat down with MIT professor Heather Kulik, one of the first materials scientists to combine AI with computational tools before it was cool.

The highlight: her team used AI to design new polymers that turned out to be four times tougher than expected. The AI discovered a purely quantum mechanical effect that surprised even the lab scientists who synthesized them. But she also drops a great reality check: LLMs still can't design a 22-atom ligand (a molecular structure) that any chemistry expert could sketch in seconds. The wins are real, but only if you deeply integrate domain expertise. Nature doesn't care about how hyped a model is.

Full episode: YouTube | Latent Space