😺 OpenAI's perfect chess move

PLUS: The ultimate prompt whisperer tool...

Welcome, humans.

Wow, DeepSeek’s success has REALLY got people excited in China. It inspired a whole tech stock rally, and this robot dance party—it’s on and poppin’ over there rn!

Where. Can. I. Buy. That. Sweater?!

Okay, so that video is actually the Unitree H1 at the Spring Festival Gala; completely unrelated to DeepSeek. But it’s such a vibe, isn’t it?

Speaking of the H1, this humanoid robot isn't just the Fred Astaire of the robot world—it's also the robot Usain Bolt: It can sprint at 3.3 m/s (that's about 7.4 mph for us Americans), making it the fastest running humanoid in China.

At 180cm tall and just 47kg (way lighter than Tesla's Optimus!), it's got 360° vision and enough joint power to... well, actually dance.

H1’s battery is also hot-swappable, so when it runs out of juice mid-performance, you can just “pop and lock” in a new one. Don’t worry, I’ll see myself out…

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

OpenAI launched ChatGPT Gov at an interesting time.
Hugging Face planned DeepSeek model clone.
KELA found serious DeepSeek R1 security flaws.
Another OpenAI researcher quit over safety concerns.

Advertise in The Neuron here.

Well played, GovGPT, well played…

Timing is everything in AI, and after Mondays's DeepSeek-induced chaos, OpenAI may have just pulled a perfect en passant.

For everybody with a sub-1000 elo in chess, en passant is that rare move where you can capture an opponent's pawn only if you act at exactly the right moment.

So what did OpenAI do? It announced ChatGPT Gov, a special version of ChatGPT designed specifically for U.S. government agencies to deploy in their own secure cloud environments.

Here's what makes ChatGPT Gov special:

Supports serious security frameworks (IL5, CJIS, ITAR).
Includes all ChatGPT Enterprise features.
Agencies control their own security and compliance.

The most interesting part of the announcement: OpenAI says government employees are already using ChatGPT like crazy—90,000+ users across 3,500+ agencies have sent over 18 million messages. For example:

Air Force Research Lab uses GPT for admin and education.
Los Alamos National Laboratory uses it for bioscientific research.
Minnesota's Translation Office uses it for, well, translation.

And gov workers in Pennsylvania even found GPT reduced time spent on repetitive tasks by 105 minutes per day.

Now you might be wondering, what makes yesterday a good timing for GovGPT?

OpenAI is trying to ride a wave of pushback after DeepSeek’s privacy policy started raising concerns—in particular, because DeepSeek collects user data like device info, keystroke patterns, IP addresses, and chat history + stores it on servers in China.

Now, this doesn’t apply to local models or apps built w/ DeepSeek, but it was certainly enough concern to get the attention of big brother:

The White House announced it is looking into the “national security implications” of DeepSeek.
The U.S. Navy has outright banned DeepSeek use “in any capacity.”
Australia and The UK have urged caution as well.

To be fair, after Kevin Weil announced “GovGPT” on X, users were quick to question ChatGPT Gov's own data practices: “It's concerning that official public data is being given to AI without explicit permission."

Weil's response was equally quick: “The product is self-hosted in the customer's already accredited cloud environment... The whole point is data protection!”

Our take: This isn't just about launching a government product—it's OpenAI drawing a line in the sand about AI security and sovereignty, positioning itself as the secure, American-made alternative. The trusted choice for all your sensitive operations.

Subtext: OpenAI might not have a new model that can crush R1 ready to deploy right now, so playing the safety card is the next best thing.

FROM OUR PARTNERS

Mastering AI Agents…

Learn how to create powerful, reliable AI agents with Galileo’s in-depth eBook! Enjoy 100 pages of expert content on:

Choosing the right agentic framework for your use case.
Evaluating and improving AI agent performance.
Tips for identifying failure points and production issues.

Get your copy now

Prompt Tip of the Day

Ever wished for an AI “prompt whisperer”? That's basically what Maskara.ai is—a tool trained on thousands of high-quality prompts to help you write better ones. Think of it like Grammarly, but for talking to AI. You can even test your prompts across different models to see how they'll perform before you use them.

Fun fact: Maskara is specially tuned for reasoning models, too. Here’s some demos of Ashutosh Shrivastava generating prompts for ChatGPT o1 (which he noted also works well with DeepSeek R1), and validating the prompts w/ Claude 3.5 Sonnet.

Treats To Try.

Goose is a free open-source agent-building framework to make building agents easier (from Jack Dorsey of Block/Twitter fame).
YuE is an open-source music tool like Udio or Suno that’s actually good—listen to these demos, then watch this guide to set it up yourself.
Agentfund invests in startups building autonomous software tools, offering $50k-$100k checks to help you grow your company.
Pika 2.1 is a new, extremely realistic AI video generator that exports videos at 1080p—here’s 10 wild examples of it in action.
Kling has a new tool called elements that let you maintain character consistency in your AI videos (similar to Pika’s Ingredients)—here’s a demo.
Tripo 2.5 can generate incredibly detailed 3D meshes with high quality textures—there’s also Spell by Spline, which lets you create and publish your interactive 3D designs for the web + Apple platforms.
Nowadays handles all your corporate event planning tasks, from venue negotiations to group travel bookings—dog themed app, but we’ll allow it.

See our top 51 AI Tools for Business here!

Around the Horn.

ICYMI: OpenAI apparently created this 20 minute video of Operator, it’s new agent tool that’s been almost completely buried by DeepSeek news

Meta set up four “war rooms” dedicated to DeepSeek—two to analyze how DeepSeek was able to reduce training costs and one for what data it used.
Dylan Patel of Semi Analysis says DeepSeek spent “well over $500M” on GPUs over the course of the company’s existence, so NVIDIA need not worry.
Hugging Face announced plans to create Open-R1, an open-source reproduction of DeepSeek's reasoning model, aiming to fill in missing pieces like training code and datasets that weren't publicly released.

FROM OUR PARTNERS

Want AI assistants that actually know your business?

Learn how to build GenAI systems that can access all your company data (Drive, Salesforce, Confluence, etc.) without breaking the bank or burning out your engineering team. This live tutorial from Unstructured will show you how it's done, step-by-step.

Bots Behaving Badly

Interestingly, DeepSeek’s biggest security flaw might actually be how easy it is to jailbreak R1, as red teams from KELA found the model is highly vulnerable.
Character AI claimed in court that restricting its chatbots would violate its users’ constitutional rights to free speech.
Another OpenAI safety researcher quit due to the “terrifying” pace of AI development and “very risky gamble” of rushing products to market.
France’s government-backed chatbot Lucie was taken offline after claiming cows lay eggs and insisting “the square root of a goat is one.”