Welcome, humans.

It looks like a version of SORA 2 might have been leaked:

Some fascinating insights on AI video at the end of the demo as well. The creator said they used Midjourney 6.1 to create the characters, then SORA to animate it.

The most striking improvements? Better physics (materials and objects move more naturally), longer clips (up to 15 seconds without breaking), and more realistic lighting.

Another post said this version could come with 1-min output, image-to-video (hence the use of Midjourney), and video-to-video features.

There are still quirks, tho. Characters don't stay consistent between scenes, the camera never stops moving (probably to hide consistency issues), and some eagle-eyed viewers spotted assets from Assassin's Creed Valhalla in the final output.

Reddit’s reaction? “At this rate, we'll see SORA 3 leaks before SORA 1 comes out.” LOL, so true…

As The Information wrote back in September, OpenAI has been testing a better, cheaper SORA… so this demo could have been built with that (or it could still be SORA 1.0 for all we know).

And, because pretty much everyone is assuming there will be a SORA release as part of OpenAI’s 12 Days of Shipmas, we’re guessing some sort of SORA is coming soon (could even come today).

Here’s what you need to know about AI today:

• We test your prompt ideas in o1 Pro mode (and share some o1 Pro demos, too).
• Apple’s AI features should launch this week.
• Clone robot shows human-like skeletal movement.
• Demos of Tencent's Hunyuan video model “shatter” the uncanny valley.

Advertise in The Neuron here.

o1 Pro is out of control…but it might come at the cost of o1 for the rest of us.

Obviously, OpenAI announcing Pro as a new $200 a month tier shows you where the company’s priorities are: “gotta make that bread.”

Well, from some early demos (and our own tests), we can confirm: o1 Pro is really, really good.

Check out this demo of o1 Pro easily coding 25 games, each with bots playing, in only five minutes:

Here are a few others:

1. Allie K. Miller uploaded a flowchart and asked o1 for the code to build it.
   1. McKay Wrigley did the same thing with screenshots of Coinbase, and o1 pro cloned it in one shot just with this prompt.
2. Fofr used o1 to solve the NYT Connections puzzle in 14 seconds.
3. TigranSloyan used o1 Pro to solve the hardest math question from the 2006 IMO.

For direct comparison: Here are 3 different renderings of Earth: one from 4o, one from o1, and one from o1 Pro. Only o1 Pro could figure out the mouse control feature. And speaking of video games, the same creator (Kevin Kern) also built his own mini Metroid game with o1 Pro, Midjourney, and Bolt.new (the AI full-stack app creator).

We tested o1 Pro per your suggestions, and here’s what we found:

1. “How well can it research and write this newsletter?” Hard w/o internet access, but it tried.
2. “Coding websites in WordPress theme” Also had it code a cat lovers website—anyone ambitious is welcome to publish it, we’ll feature it if you do!)
3. “What would it take to get pigs to fly?” Whoa…
4. “Analyze the most recent Supreme Court decisions that relate to immigration.” Again, hard w/o internet, so we fed it the most recent decision.
5. “Can it say David Mayer?” No spoilers…

Remember: these outputs are only as good as the data that goes in. When this thing gets internet access… * mind blown *

However, some regular o1 users are saying this new Pro mode might come at the expense of “o1 for the rest of us” being weaker than o1-preview (we really need better names here guys… you’re killing us with all these o1 variants!). Why? Here’s one theory.

Here’s the bear take: Right now, the performance of o1 regular isn’t compelling people to upgrade for even more goodness. Also, the 50 prompts per week limit is still in place (and still irritating people).

Plus, some are saying Gemini’s new Exp 1206 (also launched this weekend) can do the same things o1 pro can, but for free… (and it has the benchmarks to prove it).

Example: here’s a game Exp-1206 made in a single prompt, and here’s Exp-1206 crushing math. To try Gemini Exp-1206, go to AI Studio > Model > Experimental 1206.

FROM OUR PARTNERS

💼 Want to become an AI Consultant?

This stat is WILD: The AI consulting market is projected to explode from $6.9 billion to $54.7 billion by 2032. Want to capture this opportunity?

So how do you, an AI enthusiast, become a handsomely compensated AI consultant?

Well, The AI Consultancy Project from Innovating with AI has already helped 300+ students build successful consulting businesses.

Current students love:

• Tools and frameworks to find and serve clients.
• 6-month roadmap to building a 6-figure business.
• Methods to land your first client in as little as 3 days.

As a Neuron reader, you get early access to the next enrollment cycle!

Take advantage of the AI boom—request early access to The AI Consultancy Project ASAP!

Treats To Try.

1. *Attention eliminates your biggest sales roadblocks by breaking through language barriers, turning calls into case studies, and making every rep your best rep yet. See how → here.
2. Invideo creates a final video output (visuals, score, narration) from a single prompt.
3. Martin is a personal assistant that automates and streamlines everyday tasks like managing your calendar, email, to-do lists, and Slack (launch video).
4. Plus AI for PowerPoint makes PowerPoint presentations for you from a prompt, a doc, or a PDF.
5. Autogmail automatically writes responses as a draft every time you get emailed that you can quickly approve or tweak and then send.
6. n8n automates complex workflows and technical tasks with over 400 integrations and full source code available for self-hosting.
7. Llama 3.3 is out, and apparently ~25x cheaper to use than GPT 4o on OpenRouter, which provides a single interface to chat with any AI model.
8. Icebreakers.wiki generates fun ice-breaker questions for teams and group discussions—or if you want to ask literally anything other than “so what do you do for work” at parties…

See our top 51 AI Tools for Business here!

*This is sponsored content. Advertise in The Neuron here.

Around the Horn.

SORA? Computer use? Advanced voice mode with vision? Check here at 10AM PST to find out…

• Here’s more footage of Clone, the humanoid robot with a skeleton of all 206 human bones made out of myofiber, complete with minor fusions, artificial ligaments, fully articulated joints, and connective tissues. Amazing + terrifying!
• Apple will likely release iOS 18.2 this week, which will include Visual Intelligence, natural language search, image generation through Image Playground, ChatGPT integration via Siri, and customizable emoji creation.
• Here are 14 WILD examples of Tencent’s new Hunyuan AI making realistic AI videos that “shatter the uncanny valley” (shared by filmmaker @EHuanglu!).
• Here is a recap of all the major AI papers from last week, complete with explanations from the authors for each in the thread. Our fave is this one on “reverse thinking” as a way to make LLMs stronger reasonsers.
• Someone used ChatGPT to help them win more money in a legal challenge against an airline, and shared their process for using GPT to do this.

FROM OUR PARTNERS

Live Webinar Next Week:

Learn to generate chunked, vectorized JSON from any document type and build accurate RAG pipelines using Unstructured + Ragas in minutes.

Register to join or get the recording.

Prompt Tip: Simple format changes can make AI 40% smarter.

Want better results from cheaper AI models? Microsoft researchers found a simple format trick that makes a big difference:

For GPT-3.5 ($0.50/1M tokens): Use JSON format

{

"Instructions": "Your task",

"Examples": "Example output",

"Task": "Your request"

}

For GPT-4 ($3.00/1M tokens): Use Markdown format

## Instructions

Your task

## Examples

Example output

## Task

Your request

These formatting preferences might change as models evolve, but for now, this is an easy way to squeeze more performance out of your AI budget.

Quick hack: Don't want to format things yourself? Just ask any AI to “reformat my prompt into JSON format” or “convert this to markdown format” and it'll do the heavy lifting for you.

Dev Note: FYI, GPT-4 is “overtrained” to output in markdown format (because the OpenAI chat UI uses it). If you're using the API, for most use cases, using a markdown-to-html library is the simplest solution—but if you absolutely need to prevent markdown in your output, use the logit_bias param to discourage markdown tokens.

Monday Meme.

When you read how o1 tries to escape (and then lies about it) 5% of the time…

A Cat's Commentary.