😸Everything that happened in AI this weekend

OpenAI's Sora 2 updates, new agent builder, and the mysterious Cheetah model...

Welcome, humans.

This is pretty much the best video you can watch to understand the risks and opportunities of the AI bubble (and yes, it’s definitely a bubble; even Jeff Bezos thinks so, but he argued society will still reap “gigantic” benefits from the tech).

Anyway, if you want the clearest eyed view of where we are and the players involved, there’s really no better source than Dylan Patel:

Inside the Trillion-Dollar AI Buildout | Dylan Patel Interview

We’re simple folk here at The Neuron: we see Dylan Patel, and we hit share.

P.S: This is not financial advice (obvi), but Dylan DID say multiple times that him doing this interview was going to pop the bubble… we’ll see if he’s right!!

Here’s what happened in AI today:

A new model called “Cheetah” has entered the chat.
OpenAI is about to deploy an agent builder like n8n at today’s Dev Day.
Sam Altman revealed plans for “revenue-sharing” on Sora 2.
Researchers create an AI powered by actual human brain cells.

Advertise in The Neuron here

Here’s everything that happened in AI this weekend

Over the weekend, OpenAI's invite-only Sora app shot up to #1 on the App Store, and that’s despite the rising backlash over the content filters kicking in. While rights-holders had to “opt out” to stop people from making Grand Theft Auto videos featuring Sponge Bob and Mario, the filters seem to be blocking everything atm.

That said, Sora users were still able to get a few of these gems before the great outage took hold…

In an update on his blog, Sam Altman said Sora will soon add granular, opt-in controls for rights-holders (more about that progress here) and is exploring revenue sharing so both creators and rights-holders alike can eventually get paid for creating videos, which incentivizes creators to keep creating, and rights-holders to lend out their IP for a cut of the memes. I mean really, if there’s fair monetization involved, why not let people have fun, right??

Well, Casey Neistat's latest video offers one stark reason why not: we could create a future where effortless “AI slop” drowns out genuine human creativity.

Next up: OpenAI's DevDay kicks off today, and leaks from this weekend suggest they're unveiling “Agent Builder”, a drag-and-drop canvas for building AI agent workflows without code:

In the video shared above, you can wire together customer service bots, Q&A agents, and data pipelines using modular blocks (like if-else logic, MCP connectors, file search). Basically n8n, but with GPT-5 baked in. This could let non-coders build production-ready agents in minutes instead of days of API integration.

And then a Mystery “Cheetah” model appeared in the wild (a.k.a Cursor). Cursor users discovered the mystery model in their IDE, but apparently no one knows who made it.

At $1.25/1M input and reportedly ~400 tokens/second, theories are all over the map:

Ian Nuttall thinks it might be Gemini 3.0 based on the pricing and blazing speed.
GosuCoder suspects it's a new Grok variant (better than Grok Code Fast).
Dan Mac predicted it's OpenAI's GPT-5-Codex or even possibly a Gemini diffusion model.
Sajid Mehmood thinks it's Cursor's first in-house codegen model.

The undisclosed provider has privacy-conscious devs worried about sending code to an unknown backend, and also, some users report extremely high token consumption, suggesting something's off. Guess we’ll see later today / this week!

FROM OUR PARTNERS

Airia: Enterprise AI Without Compromise

There's a reason most enterprise AI initiatives never make it to production: Companies can't find a platform that's both powerful and secure enough. But not anymore.

Airia is the Enterprise AI platform built to deliver speed without compromise. Unlike other platforms that force companies to choose between fast deployment and secure operations, Airia engineers speed and security together. Launch AI quickly without cutting corners on compliance. Move at the speed of business without moving past security requirements.

Ready for AI at full speed with zero compromise? Start your free trial today. No credit card required.

Prompt Tip of the Day

Stop writing prompts—let AI write them for you. Seriously, it’s at the point where we might rename this section…

Simon Willison says this 30-min talk talk from Drew Breunig is the best explanation of DSPy, an open-source framework that treats prompting as programming, not string-crafting. He used DSPy to automatically generate optimized prompts instead of manually crafting them. His accuracy jumped from 60% to 82% with 14 lines of code.

How it works: Define your task (input → output), provide eval data, and DSPy uses a large model to write and test prompt variations. When you switch models, it auto-optimizes new prompts for each one.

Tomorrow there's always a better model. Don't waste hours perfecting prompts that'll be obsolete next week: write tasks, not prompts. You can read Simon’s notes here.

Treats to Try

Stuff for everyone
1. Flai for dealerships answers your store's calls, texts, and emails with voice agents that book service, route leads, and follow up (raised $4.5M seed).
2. Super Intern group chat adds a teammate to your Discord/Telegram that answers FAQs, flags scam links, summarizes threads, posts news, and schedules reminders—free plan; $20/month starter, $200/month Pro.
3. Kerns turns books/papers/links into interactive knowledge maps, reads sources with summaries, and keeps researching in the background so you can synthesize faster.
4. AI Spine Animation Generator lets you upload a 2D character and auto-rig bones, apply 2,000+ motions, then export Spine JSON/Atlas/PNG for Unity or Godot—so you can go from concept art to a walking, jumping, punching character in minutes.
Technical finds
1. Liquid AI released LFM2-Audio (1.5B), an end-to-end audio to text model claiming <100 ms first-audio latency and strong VoiceBench scores.
2. ProofOfThought repo adds a logic check to your model's answers—you ask a question, it auto-writes a Z3 proof and only returns the result if the proof passes (open source).
3. run universal runner lets you execute snippets in 25+ languages from one CLI and drop into a persistent REPL (e.g., echo '{"name":"Ada"}' | run js --code ...).
4. Renderarxiv searches arXiv (pre-print paper platform) from your terminal and opens clean HTML (human view + LLM-ready text) with direct PDF links—great for citations.
5. FLE v0.3.0 release lets you spin up headless Factorio servers and run long-horizon agent evals via a Gym-compatible SDK and simple CLI—perfect for testing planning, recovery, and world-modeling.

Around the Horn

Or as Dylan says, 7M sixteen year olds. Here’s the full report.

Groq (AI-inference chip maker) planned 12+ new data centers for next year after opening 12 this year, with first Asia sites on deck.
MIT and McMaster University said a new antibiotic, enterololin, targets IBD-linked bacteria and that AI predicted its mechanism (LolCDE) before lab proof.
OpenAI acquired Roi (personal finance app) as an acqui-hire, with only the CEO joining, to deepen its push into personalized consumer AI.
Google is testing a Gemini app makeover that swaps the chat screen for a scrollable feed with image-led prompts and quick actions (Create Image, Deep Research, Live).
Swiss lab FinalSpark grew mini human brains from stem cells to power biocomputers for energy-efficient AI.
Here’s Rohan Paul’s picks for the top AI papers of last week, and here’s Elvis Saravia’s top picks.