Welcome, humans.

Supposedly, and this is a big “we have no idea if this is true”, some early users got access to Gemini 3.0 through what’s called “Gemini 3.0 Pro - ecpt checkpoint” that has been showing up in A/B tests in AI Studio.

We've seen dozens of posts claiming early access to Gemini 3.0, but this was the first one that made us actually believe it:

Code, Prompt, and Windows variant Code and Prompt.

An AI red teamer (Chetaslua) asked the supposed Gemini 3.0 Pro to build a fully functional web-based operating system like macOS, complete with text editor, file manager, paint app, video editor, and all the bundled software. The prompt? Make it work in a single HTML file and make it beautiful.

The result was a pixel-perfect macOS replica with working apps, smooth animations, and even a functional browser (they created a Windows variant, too). All generated in one shot.

Also, Chetaslua had to preface their post with “this is not cherry-picking” and “Google didn't pay me.” When your AI demo needs a disclaimer because it seems too good to be true, that’s a good sign.

We even asked Logan Kilpatrick of Google DeepMind to verify if this is true or not, but he was sleeping, which TBH, we should probably be doing ourselves as we write this… good morning, tomorrow!

Here’s what happened in AI today:

1. OpenAI revealed its partnership with Broadcom to build custom AI chips.
2. Andrej Karpathy enabled anyone to train a ChatGPT clone for $100.
3. xAI hired two former NVIDIA researchers to develop game AI.
4. NVIDIA began shipping its DGX Spark desktop systems to key customers.

Advertise in The Neuron here

OpenAI bets big on building its own chips…

DEEP DIVE: Get all the details on OpenAI’s Broadcom deal here.

It’s becoming a bit of a tradition in the AI industry to start yet another week with another 10 gigawatt announcement from OpenAI. Shall we start calling it Ten Gigawatt Tuesday?

This week, it’s OpenAI’s long-rumored deal with Broadcom getting publicly announced:

• The partnership will deploy 10 gigawatts of custom OpenAI-designed AI accelerators and networking systems, with deployment starting in late 2026 and completing by 2029 (they’re expecting initial silicon back “very soon”).
• The two companies have been quietly collaborating for the past 18 months, designing not just a custom chip but an entire system optimized for OpenAI's specific workloads.
• Plot twist: OpenAI used its own AI models to help design the chip, achieving “massive area reductions” by having AI optimize components that would've taken human engineers another month to get to.
• And because the infinite money glitch go brr, Broadcom’s stock closed up 9.88% by the end of the day.

OpenAI x Broadcom — The OpenAI Podcast Ep. 8

Now technically, Broadcom’s President said its secret $10B customer announced back in September is not OpenAI. But could it be Arm and/or SoftBank (who owns 90% of Arm), since OpenAI is apparently working with them to develop its new CPU for inference?

If you haven’t watched our recent podcast episode on AI inference, inference is basically what an AI model does when you ask it a question and it decides how to respond.

The reason OpenAI is signing all these deals is to establish its own hyperscale level cloud, and as a byproduct, reduce reliance on NVIDIA, the AI industry's most singular point of failure (controlling 90%+ of the AI chip market), by making AMD and Broadcom legit competitors, and by extension, grow the resilience of the whole industry.

And according to Sam, this deal is in addition to everything else it announced. To put that in perspective:

• OpenAI started with a 2-megawatt cluster (one large Costco)…
• Scaled to 2 gigawatts this year (enough to serve 800M weekly users)…
• …And will hit close to 30 gigawatts (literally all of Australia) with these recent partnerships.

But here’s the problem: because America’s economy is basically just one big bet on AI this point, all this attention on AI could inflate the AI bubble even bigger (because, TBH, what else are you going to invest in atm?). This is only a problem if, as Noah Smith writes, any new AI model even just “slightly disappoints” on launch.

The next turn in the launch cycle will be the release of Gemini 3.0, which if this leaked screenshot is to be trusted, will be officially released on Oct 22 (you never know for sure, though). Let’s hope it does at least sliiightly better than disappoint!

FROM OUR PARTNERS

Innovate Faster with Clutch

Got a project in mind? Finding the right expert doesn’t have to be complicated. On Clutch, all it takes is sharing your project goals — and they’ll handle the rest.

Their matching process is tailored to your specific project, connecting you with vetted AI professionals who have the skills and experience to deliver measurable results.

Whether you’re building, optimizing, or innovating, skip the endless searching and guesswork. With Clutch, your perfect match is just minutes away.

Meet AI experts

This is honestly so freaking cool…

AI legend Andrej Karpathy just dropped nanochat, a repo (code repository) that lets you train your own ChatGPT clone from scratch for about $100 (on a platform like Lambda, which Karpathy used as an example). That's right, the price of a decent dinner for two (or, let's be honest, one dinner if you're in San Francisco).

Read the full post.

Here's the wild part: you boot up a cloud GPU, run a single script, and four hours later you're chatting with your own AI through a web interface. The whole thing (tokenizer training, pretraining on web text, finetuning on conversations, and serving it up ChatGPT-style) lives in just 8,000 lines of readable code.

Now, before you get too excited, at the $100 tier you're basically talking to a very confident kindergartener. It can write poems and answer simple questions, but don't expect it to ace your physics exam.

Scale it up to $1,000 though? You get something that starts solving math problems and passing multiple choice tests. The repo will become the capstone project for Karpathy's LLM101n course, and it's designed to be minimal, hackable, and maximally forkable… basically the opposite of every other AI framework that requires a PhD to navigate.

Here’s the technical breakdown, and if you want to chat with the repo, click here. And if all this fascinates you, but it’s too technical, watch this deep dive from Karpathy to get a nice overview and introduction to language models like nanoGPT.

Treats to Try

*Asterisk = from our partners (only the first one!). Advertise in The Neuron here.

1. Exa adds near instant web search to your AI app, so your chatbot can instantly answer “what's the latest on Tesla?” with current information in under 350ms (you can also use it as your own AI search engine; we use it here in our apps!).
2. Microsoft just released MAI-Image-1, which generates photorealistic images from your text descriptions, with particularly strong lighting effects and realistic, which is now testing in the LM Arena.
3. Nano Banana is now available in Search, NotebookLM, and Photos; just type “make this a watercolor” on your selfie in Lens, turn your notes into anime-style video overviews, or generate new images from text in AI Mode.
4. GraphBit helps you build agents that connect your databases, APIs, and services, running 14x faster and using 13x less memory than LangChain.
5. Extrovert finds your prospects' LinkedIn posts and drafts comments that sound like you, turning engagement into warmer sales conversations.
6. Fruitful alerts you in Slack when competitors change pricing, hire, or launch products—free to try.
7. Tight Studio records your screen and automatically adds smart zoom-ins, AI voiceover (to fix your accent), captions, and music, turning raw recordings into polished product demos in minutes.

Around the Horn

• NVIDIA began shipping its first DGX Spark systems (petaflop-performance platforms with GB10 Grace Blackwell Superchip, ConnectX-7 networking, and NVIDIA AI Software stack), with orders starting Wednesday, Oct. 15.
• Elon Musk's xAI hired two former NVIDIA researchers specializing in “world models” (AI simulating environments that understand physics) to develop game AI and compete with Meta and Google.
• As The Information writes, a toxic brew of leverage (debt) in crypto trading led to the $700B sell-off over the weekend, where positions got programmatically liquidated. Yueqi Yang suggests that this can be seen as a dress rehearsal for what happens if AI’s “multilayered leverage” gets similarly margin-called (where, us talking now, companies like Oracle borrow billions to build data centers based on OpenAI's future revenue commitments, but AI adoption disappoints and OpenAI can't pay its $60B/year bills, the debt can't be serviced, triggering a cascade through the entire circular financing loops).
• We just watched a great (but nerdy) interview with Columbia professor Vishal Misra on whether or not large language models will lead us to AGI; Vishal created TokenProbe, which shows you how language models generate text token-by-token, visualizing the probability and entropy (how confident vs. uncertain the model is) of each word as the model decides what to write next (paper). He thinks AGI will need some form of “simulation-based reasoning"…

FROM OUR PARTNERS

Recraft's New Chat Mode Makes AI Design Feel Like Having a Creative Partner

Stuck explaining what you want to an AI image generator? Recraft just launched Chat Mode—think ChatGPT meets Photoshop canvas.

It remembers your entire workflow, suggests smart refinements, and helps you iterate without starting over. Used by Netflix, HubSpot, and Asana to create production-ready designs.

[Join the Chat Mode waitlist—available in beta now]

A Cat’s Commentary