DeepSeek returns with an IMO-crushing AI

Good morning, AI enthusiasts. The large whale is back — this time with an open-source reasoner that hits gold-medal performance on certainly one of the world’s hardest math competitions.

But the true story is what it unlocks: frontier reasoning, once gated by proprietary labs, is now free for all. For the primary time, anyone can own the brain of a world-class mathematician.

In today’s AI rundown:

• DeepSeek’s recent reasoner crushes IMO 2025

• OpenAI’s API user data leaked in third-party breach

• Create Instagram product shots with Nano Banana Pro

• NVIDIA’s case for scale isn’t every little thing in AI

• 4 recent AI tools, community workflows, and more

LATEST DEVELOPMENTS

DEEPSEEK

🤖 DeepSeek’s recent reasoner crushes IMO 2025

The Rundown: DeepSeek just released DeepSeek-Math-V2, an open-source MoE model that achieves gold-medal performance at IMO 2025, democratizing “research-level” mathematical reasoning that was previously locked behind proprietary partitions.

The small print:

• The model scored 118/120 on the 2024 Putnam competition (beating the highest human rating) and solved 5 of 6 IMO 2025 problems, hitting the gold standard.

• On IMO ProofBench, it hit 61.9%, nearly matching Google’s specialized Gemini Deep Think that won IMO gold and crushing GPT-5, which scored only 20%.

• Math-V2 uses a generator-verifier system where one model proposes a proof and one other critiques it — as a substitute of rewarding final answers only.

• The verifier assigns confidence scores to steps, forcing the generator to refine weak logic and ensuring step-by-step self-debugging of reasoning.

Why it matters: By open-sourcing a model that rivals Google’s internal heavyweight, DeepSeek has broken the monopoly on frontier mathematical reasoning, providing the community with a blueprint to construct agents that debug their very own thought processes. This is usually a game-changer in domains like engineering, where mistakes are costly.

TOGETHER WITH BLAND

📞 There’s an AI company without churn?

The Rundown: In San Francisco, there’s an AI company who’s making inbound voice agents that sound human, run 24/7, and are so good that they deliver 127% net revenue retention.

Should you call it yourself, you’ll have the opportunity to:

• Have Bland role-play any use case

• Experience the bottom latency on the planet (<500ms)

• Have a phone call that you just actually enjoy

For Black Friday, Bland is offering to create your enterprise a custom agent without spending a dime so you may validate the standard before you commit.

OPENAI

‼️ OpenAI’s API user data leaked in third-party breach

The Rundown: OpenAI just revealed that its analytics vendor Mixpanel suffered a security incident, with an attacker exporting a few of its API users’ profile information — although no chat data, API keys, payment details, or credentials were compromised.

The small print:

• The breach occurred on November 9, covering Mixpanel’s systems that provided web analytics on the frontend interface of OpenAI’s API product.

• The information the attacker exported included profile information related to the API product, akin to names, emails, locations (city/state), and device details.

• OpenAI confirmed that users of ChatGPT and other products weren’t impacted, and no chat, API data, credentials, or payment details were leaked.

• It removed Mixpanel and is notifying affected users directly, while urging vigilance against potential phishing attempts that might use the leaked data.

Why it matters: While OpenAI’s defenses held, this incident serves as a stark reminder of the safety risks third-party partners can introduce. For affected API users, the immediate danger isn’t account compromise but relatively social engineering, where attackers may use the leaked emails to create much more trouble.

AI TRAINING

🖼️ Create Instagram product shots with Nano Banana Pro

The Rundown: On this tutorial, you’ll learn the right way to use Nano Banana Pro to generate a full 9-image Instagram feed from only one inspiration photo, turning your product shots into cohesive, high-quality visuals for social media campaigns.

Step-by-step:

1. Go to Gemini → Tools → Create Images, ensure Pro mode is enabled, and upload an inspiration image that reflects your required style or aesthetic

2. Upload your product image, describe it, then prompt with: “Create a 9-image Instagram feed for this product with varied angles, people, and environments”

3. Click Undergo generate your 9-image grid. Review results and, if needed, ask Nano Banana to regenerate or isolate specific shots

4. Download your favorite visuals and post them on to Instagram, TikTok, or your brand’s storefront for an quick, consistent feed

Pro tip: The more specific and visually aligned your examples are, the higher the AI matches your required aesthetic.

PRESENTED BY YOU.COM

🧠 Tinkering with prompts can only get you up to now

The Rundown: Most firms get stuck tinkering with prompts and wonder why their agents fail to deliver dependable results. This guide from You.com breaks down the evolution of agent management, revealing the five stages for constructing a successful AI agent and why most organizations haven’t gotten there yet.

On this guide, you’ll learn:

• Why prompts alone aren’t enough and the way context and metadata unlock reliable agent automation

• 4 essential ways to calculate ROI, plus when and the right way to use each metric

• Real-world challenges at each stage of agent management and the right way to avoid them

Should you’re able to transcend the prompt, that is the playbook for you.

AI RESEARCH

📈 NVIDIA’s case for scale isn’t every little thing in AI

The Rundown: NVIDIA and the University of Hong Kong published a paper suggesting that the longer term of AI won’t come from scaling but from smarter orchestration, with their recent tool training small models that may surpass frontier AI at a fraction of the fee.

The small print:

• ToolOrchestra trains an “orchestrator” model that decides when to reason internally and when to call specialized tools and models, based on the duty.

• An 8B model trained with the system surpassed GPT-5 and Claude Opus 4.1 on Humanity’s Last Exam, scoring 37.1% while being 2.5x more efficient and faster.

• Even when tested with unseen tools, the orchestrator adapted well — showing its ability to work with changing toolsets and pricing structures.

• Prior agents overused the strongest (and most costly) tools and models, but ToolOrchestra avoided this by orchestrating targeted model and gear usage.

Why it matters: In step with Ilya Sutskever’s recent comments, ToolOrchestra challenges the “greater is best” ideology. As an alternative of 1 giant system, NVIDIA shows how small models coordinating tools stands out as the path forward. If orchestration beats scaling, the neatest model/tool conductor will probably be the following big AI breakthrough.

QUICK HITS

🛠️ Trending AI Tools

• 🤳 Perplexity – AI answer engine, now with virtual try-on for shopping

• 🧠 Math V2 – DeepSeek’s open-source mathematical reasoning model

• 🤖 Stories – Character AI’s interactive experience for teenagers

• 🎆 FLUX.2 – Black Forest Labs’ recent visual intelligence model

📰 Every little thing else in AI today

Jeff Bezos’ recent stealth AI enterprise, “Project Prometheus,” quietly acquired General Agents, an agentic computing startup, Wired reports.

OpenAI lost a key discovery ruling, forcing it at hand over internal communications about why it deleted two datasets of allegedly pirated books, boosting authors’ possibilities of proving willful copyright infringement.

Perplexity launched persistent memory, enabling the assistant to recollect user preferences, interests, and conversations for beneficial context on relevant tasks.

Perplexity also updated its email assistant to work across multiple calendars without delay, currently available for Gmail and Outlook.

Cohere expanded its partnership with SAP, taking its agentic AI platform, North, to SAP’s Cloud infrastructure and Business Technology Platform.

Alibaba released Quark AI Glasses, a sensible eyewear line powered by its in‑house Qwen LLMs and Quark assistant, in China — with prices starting at 1,899 yuan ($268).

COMMUNITY

🤝 Community AI workflows

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader Allen W. in Santa Barbara, CA:

“I’m a songwriter trained in traditional folk music. I write songs, after which ChatGPT helps me translate them into indie folk. I give it the lyrics and ask how an indie folk songwriter would write these lyrics. ChatGPT desires to do every little thing, like determine chord structure, tempo, etc., but I keep it only on the lyrics.”

How do you employ AI? Tell us here.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Karpathy’s advice for the AI classroom

• Read our last Tech newsletter: This device wiretaps your ‘second brain’

• Read our last Robotics newsletter: Figure sued over ‘skull-crushing’ force

• Today’s AI tool guide: Use Nano Banana Pro to create product shots for Instagram

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer—the humans behind The Rundown