😺 10 WILD Deepseek demos

Welcome, humans.

Perplexity, the AI search company, just shot its shot with TikTok, proposing a new merger that would give the U.S. government a 50% stake in the new company.

For context, Perplexity is worth ~$500M, and TikTok is worth “well north of $100B.” Talk about punching above your weight…

Offering to let the government be a majority shareholder just to get TikTok's attention is kinda like a reverse “No Hard Feelings” situation—instead of parents hiring someone to date their kid, Perplexity is bribing the parents (in this case, Uncle Sam) to date TikTok… and insisting they chaperone, too.

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

• We round-up 10 demos of Deepseek R1 in action.
• Meta planned $60B AI expansion for 2025.
• Apple named new AI chief for Siri upgrade.
• ElevenLabs, Rad AI closed major funding rounds.

Advertise in The Neuron here.

Let’s compare DeepSeek vs OpenAI across some WILD demos…

Over the last week, pretty much everyone in the AI space has been losing their minds over Deepseek R1. The open source community has been loving it, the closed source tech giants have been less than loving it, and even the mainstream media is starting to pick up on how last week’s R1 launch was a big deal

We’ve been trying to understand just how powerful R1 really is, so we rounded up everything we could find that shows off just what this little AI side project can do.

Here’s some WILD demos of what people have done with Deepseek R1 so far:

1. Make a 100% 3D game from scratch (here’s the code and the convo).
2. Use R1 with Cursor while coding for o1 (or better) level insights, for cheap.
3. Create a 3D space invaders game + a tetris clone.
4. Test its ability to solve complex reasoning problems.
5. Built a local app to chat with your PDFs.
6. Complete an entire piano app in just about one prompt.
7. Set up the small version to run on a phone (using this tutorial—not easy).
8. Directly compare it against o1 Pro on creating animations with python scripts.
9. Built a Perplexity clone, but with R1 (here’s the code).
10. Create an app to analyze Youtube videos and create trend reports.

And as long-time readers of The Neuron may remember, there’s also the Minecraft benchmark, which DeepSeek R1 is passing with flying boxes.

So what exactly makes R1 so efficient? As Armen Aghajanyan (ex-Microsoft) pointed out, while US companies threw massive compute power at the problem, DeepSeek found clever shortcuts that get similar (or better) results with less computing resources.

DeepSeek didn’t just throw a bunch of data at R1 and let it do its thing, they guided it through multiple stages and iterations to teach it to become a good reasoner. Y’know, like how we teach humans…ish.

This sparked debate about whether US companies wasted resources by not focusing enough on efficiency—or if using more compute was worth it to get to market faster and avoid complex technical challenges (some of which DeepSeek solved).

All of this seems to be inspiring / corresponding with OpenAI moving faster to keep up. For example, ChatGPT Canvas now works with o1 and can run website-style applications in the chat (like Claude Artifacts).

People are already using that feature to create full chess games (here’s the prompt), draw robots, or build their own applications directly inside ChatGPT.

Most interesting of all, Sam Altman announced this over the weekend:

So not only can you get o1-level reasoning for free (or nearly free) with R1, you’ll get 100 free o3-mini questions per day when o3 launches, which should be in a couple of weeks… or sooner, now that R1 is threatening Sam’s throne…

FROM OUR PARTNERS

Want no-code AI agents for your business?

Sana AI is an enterprise-grade AI platform grounded in all your company’s data—making it easy to deploy secure, no-code AI agents for any team or role.

Sana AI allows you to:

• Deploy custom AI agents for any team or role without writing code.
• Consolidate AI software applications and unify the UX for AI across your organization.
• Integrate AI seamlessly into your existing tech stack.

Discover more

Prompt Tip of the Day

If you've tried out OpenAI's new agent tool Operator, you might’ve found it kinda like watching a sleepwalker attempt a 3am grilled cheese: lots of fumbling, dropped cheese, and somehow butter on the ceiling (??), but hey, mission accomplished?

This happens because we’re asking the AI to do everything at once. Instead of letting Operator run wild with broad, complex tasks, here's a smarter way to structure your prompts (from MindDB's FLAT approach):

1. Use yes/no checks: “Is this [specific thing] true? Yes/No” “Does this page contain [specific data]? Yes/No”
2. Give limited options: “Is this lead: HOT (meeting requested), WARM (showed interest), or COLD (no engagement)?”
3. Request specific data points: “Find these exact fields: - Company size - Tech stack - Recent funding”

Why it works: Agents act more predictably when tasks are broken into clear decision points rather than open-ended instructions.

Try it: Next time you write a complex Operator prompt, break it into a series of simple questions or choices.

Example: Check if a company has a public pricing page before attempting full competitive analysis. We’re willing to bet your success rate will dramatically improve!

Treats To Try.

1. Gemini 2.0 Flash Thinking from Google DeepMind is Google’s fast “reasoning” model, with 73% accuracy on advanced math tests—try it here.
2. Bookread is an e-reader where you can select any text while reading to get instant explanations, and generate chapter-end flashcards to test yourself.
3. DROP helps you share files up to 10GB, create professional portfolios, and generate fresh designs through tools like DreamRender (turns ideas into HD images).
4. CopyCat is a free alternative to Operator that helps automate repetitive browser tasks by recording how to do them (mac only rn, there’s also Meha).
5. Wepost creates and schedules your social media posts across LinkedIn, TikTok, and Instagram while keeping your brand's voice consistent.
6. Telescope helps you find hyper-specific sales leads by letting you search with natural language criteria, like “5+ Product Managers with no QA team.”
7. Humanity’s Last Exam is a benchmark of the toughest questions to assess how close we are to AGI…and they’re still accepting new questions!

See our top 51 AI Tools for Business here!

Around the Horn.

• Meta will spend $60B+ in capex in 2025, and plans to have 1.3M GPUs (the chips that make AI work) by the end of the year—plus, it’ll grow its AI team, too.
• Apple brought on a new exec (Kim Vorrath) to lead its AI efforts—Vorrath is sort of like the company’s internal fixer, and now she’s going to help upgrade Siri.
• Besides Perplexity’s bid, the Trump administration is working on a deal to let Oracle manage, monitor, and “provide oversight” over TikTok. As a reminder, here’s what that might look like, according to Oracle’s CEO.
• ElevenLabs just raised $250M at a $3B valuation, while Rad AI who builds AI-powered tools for radiologists just raised $60M at a $525M valuation.

FROM OUR PARTNERS

Transform your career with a M.S. in Artificial Intelligence

Master of Science in AI at the University of St. Thomas

Shape the future of tech with St. Thomas' M.S. in AI. Gain hands-on experience with cutting-edge technologies, learn from industry experts, and tackle ethical challenges. With flexible evening classes and no prior experience required, build a career-advancing skillset tailored for today’s AI-driven world.

Lead in AI with a Masters—Learn More.

Monday Meme

A Cat's Commentary.