OpenAI’s next model is doing mathematics

Hola Decoder😎

If someone forwarded this to you and you want to Decode the power of AI and be limitless, then subscribe now and Join Decode alongside 30k+ code-breakers untangling AI.

🎥 How to Edit Videos With ChatGPT Using Codex and Remotion

Drag a raw clip in. Paste one prompt. Get back a captioned, tightened, animated cut ready to post. No timeline. No editor. Here's the setup.

Step 1: Open Codex in the ChatGPT desktop appClick the switcher in the top left and choose Codex. This is the mode that writes and runs code on your machine; your video never leaves your computer.

Step 2: Install the Remotion plugin Click Plugins, search Remotion, and click Install. Nothing to configure. This gives Codex the ability to transcribe, cut, caption, and render video.

Step 3: Check Node.js is installedOpen Terminal and run node -v. If you get a version number, you are set. If not, install Node 18 or newer from nodejs.org. Mac users also need Xcode Command Line Tool; tell Codex to check, and it handles the rest.

Step 4: Create a project and drop your video inCreate a new Codex project pointed at an empty folder. Drag your video file in. Type $remotion and accept the plugin suggestion.

Step 5: Paste the edit prompt

Edit [filename.mp4] for Instagram Reels. Transcribe with 
local Whisper using medium.en. Cut every silence over 0.6s 
and every standalone um or uh, leaving 120ms room tone on 
each side. Captions: 3 to 4 words per page, bold white 
sans-serif, black outline, current word highlighted in 
#F5A623. Pick 5 strong moments, add an 8% zoom and a 
headline under 24 characters on each. Output 1080x1920 
at 30fps. Before rendering, show me the transcript with 
cuts marked and the 5 headlines as a table I can edit.

Step 6: Review and approve before renderingCodex shows you the transcript with every cut marked and all five headlines. Fix any brand names or numbers Whisper got wrong. Change any headline. Then say render, nothing goes out until you approve it.

First run takes 20 to 45 minutes while Whisper downloads and builds. Every edit after that is about 10 minutes.

Together with Billo

What Meta’s Changes Mean for Your Next Campaign 

Still running the Meta setup that worked six months ago? Your next creative test could be paying for assumptions that no longer hold.

Before you put more budget behind it, understand what’s changed and where creator-led Partnership Ads fit.

On October 7 at 11:00 AM ET, join Donatas Smailys, CEO of Billo, for a free 60-minute live session to:

• See Meta’s Partnership Ads data and understand where creator-handle ads could earn a place in your account.
• Improve your creative testing with Curtis Howland, VP of Marketing at Misfit Marketing, with $100M+ in Meta spend managed.
• Hear directly from Reyna Baker, Partner Manager at Meta, about platform changes that could affect how you spend your budget.

You’ll leave knowing where your Meta setup needs to change and how to put creator-led ads to work before committing more budget.

Can't attend the session live? Register anyway, and you’ll get the recordings in 24 hours.

🧮 OpenAI’s Frontier Model Is Producing New Mathematics

OpenAI released 772 mathematical papers solved by an internal frontier model, including computer-verifiable proofs and details showing how much AI reasoning went into producing them.

AI Produces New Results - OpenAI says the model generated mathematical results intended to push the frontier of human knowledge, rather than simply solving existing benchmark problems.

Proofs Get Computer Checked - Over 162 papers are being formalized in Lean, allowing computers to independently check whether the mathematical reasoning is valid. More formalizations will be added over time.

Three Hours Per Result - The average result required compute equivalent to roughly three hours of ChatGPT Pro thinking. OpenAI is also publishing reasoning summaries, compute estimates, and attempted-problem statistics.

The Model Could Go Public - OpenAI says it is working toward responsibly releasing the frontier model behind the results, while continuing to evaluate its capabilities across mathematics and other sciences.

OpenAI is treating AI-generated mathematics more like scientific research, publishing proofs, verification tools and methodology while consulting independent mathematicians on disclosure standards. It also plans to fund workshops and conferences focused on understanding major AI-produced results.

🦣 Mistral Goes Big With a 1-Trillion-Parameter Open-Weight Model

Mistral has launched the public preview of Mistral Large 4, nicknamed “Le Chonk,” its largest model yet, targeting coding, agents, multimodal work, and specialized enterprise tasks.

One Trillion Parameters - ML4 is a natively multimodal 1T-parameter model with 49B active parameters. Mistral says it leads open models on several enterprise workloads and even beats frontier closed models in some visual-grounding tests.

Coding, Agents and Cyber - ML4 scores 61.7% on DeepSWE and 59.9% on AutomationBench. In cybersecurity, it scores 82% on a vulnerability reproduction-and-patching test and 93% on Cybench.

Built Entirely in Europe - Mistral trained ML4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centers, using multilingual data spanning 160+ languages. Organizations will also be able to self-deploy it.

Weights Are Coming Soon - The preview API is available through Mistral Studio, with open weights scheduled by the end of October. Mistral says the ongoing RL run uses around 3,000 GPUs and produces 33B tokens per day.

ML4 is Mistral’s biggest attempt yet to combine frontier-level performance with open weights and European-controlled infrastructure. The model will also become the foundation for a new generation of specialized Mistral models, with further improvements expected as its reinforcement-learning run continues.

Together with Playbook Pro

Your Next 6 Months of Ad Spend Could Disappear in 6 Weeks

CAC is up 60% this year, and 70% of first-time buyers never return. Keep leaning on ads, and the budget you’re planning for the next half year could be gone before spring.

Rewriting the DTC Rulebook is the shortcut top operators are using to escape the trap, a 24-page report distilled from 200+ hours of real brand research and field tests.

Inside, you’ll learn how to:

🔥 Build emotional campaigns that double repeat-purchase rates in 90 days.
💰 Raise prices without losing volume using the Brand Value Canvas.
🧠 Cut 30% of paid-ad dependency with the Owned Media Checklist.

Download today and get 2,000+ viral video hooks already tested to grab attention in under 3 seconds, ready to plug into your next launch.

Stop feeding the ad machine. 

Grab the free report now and lock in sustainable growth before your next dollar disappears!

🏆 Tools you Cannot Miss:

🎬 Spira Maxima – Turns a script into a finished social video with a presenter, B-roll, captions, music, and optional AI clones of your face and voice. 

🎞️ LaunchReel – Reads a product website or brief, builds a storyboard, and turns it into a polished launch video with screenshots, motion design, captions, and 4K exports. 

✂️ WeftCut – Open-source desktop video editor where an AI agent can trim footage, remove silences, add captions, create transitions, apply effects, and edit directly on the timeline. 

🎥 HyperFrames Studio – Desktop video editor designed for agent-driven editing, letting creators work with local media while AI handles parts of the editing workflow. 

🧑‍🎤 Syllaby AI Avatar 2.0 – Turns scripts or ideas into presenter-led videos with realistic AI avatars, generated voices, B-roll, subtitles, and customizable expressions.

🚀 Quick Hits

📝 Google Docs will soon let users open, edit and collaborate on Markdown files directly, while Google Drive adds native Markdown previews, eliminating the need to convert files and potentially lose formatting or comments.

🖥️ Japanese real estate giant Mitsui Fudosan plans to build a data center inside an underground tunnel in Kobe, housing servers across more than 10 shipping containers, with additional underground facilities reportedly planned.

🛍️ TikTok is expanding its shopping features with external retailer checkout and an AI shopping chatbot, letting users discover, compare and purchase more products without leaving the app.

🚨 Anthropic’s Claude reportedly flagged a conversation containing alleged violent threats against a Florida sheriff’s office, prompting human reviewers to alert police. A 30-year-old woman was subsequently arrested.

⚡ Anthropic signed expanded compute deals with Google and Broadcom, including 3.5GW of additional capacity coming online in 2027, as Claude demand grows and Anthropic’s annualized revenue reaches $30 billion.

🤳AI Nugget of the Day

Thanks for Decoding with us🥳

Your feedback is the key to our code! Help us elevate your Decode experience by hitting reply and sharing your input on our content and style.

Keep deciphering the AI enigma, and we'll be back with more coded mysteries unraveled just for you!