Welcome, humans.

Did you know you can actually watch Claude 3.7 play Pokémon live on Twitch?

Apparently, Anthropic researchers spent a year seeing if Claude could master the classic Game Boy game Pokémon Red, with early models failing spectacularly.

Claude 3.5 could barely leave Pallet Town, often getting stuck in corners. In one adorable moment, Claude got so confused it typed out a formal request to reset the game when it couldn't figure out how to move.

Claude 3.7, however, has defeated THREE gym leaders and even named its rival “WACLAUD” (like Wario, but for Claude).

The improvement comes from 3.7’s extended thinking capability, which allows it to plan ahead, remember objectives, and adapt when strategies fail.

As Anthropic notes, these are “critical skills for battling pixelated gym leaders—and, we posit, in solving real-world problems too.”

In practice, Poke-Master Claude does get stuck A LOT whenever it has to update its project knowledge mid-game. Then again, there's a certain comfort in watching advanced AI stumble through a 90s video game I beat when I was 10…

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

• Claude 3.7 has fundamentally rocked the developer world.

• OpenAI and Microsoft expanded access to their AI tools.

• Amazon prepared to launch paid Alexa upgrade and new hardware lineup.

• New research found majority (52%) of workers fear AI will reduce jobs.

Claude 3.7 Sonnet is flexing its “hybrid reasoning” muscles—and devs are going wild…

Yesterday’s launch of Claude 3.7 was huge—Anthropic basically delivered what might be the year’s most significant AI leap.

After playing with it more and seeing what everybody else is building, we're straight-up stunned by what this thing can do. 

The catalog of what Claude 3.7 creates with a single prompt is becoming legendary. One-shot prompts now yield functioning games, 3D models, and interactive web apps.

Here are 15 WILD examples:

1. Websites:

   1. AI agent startup (view here).

   2. SaaS landing page (another example, prompt). 

   3. Luxury Estate finder.

   4. Animated weather cards with wind, rain, sun and snow effects.

2. Games:

   1. Three Snake games—one with a fourth-wall breaking self aware snake (try it), one Apple Watch game that ratchets up the snake’s speed based on the player's heart rate (try it), and one autonomous Snake battle.

   2. “ClaudeKart” 3D racing game, from a simple prompt. 

   3. Simple side-scrollers and full games based on short stories.

   4. “Anthropic Researcher Simulator” (prompt).

   5. Basic (but functional) clone of Pokémon Red—topical!

3. 3D Modeling:

   1. City simulation with skyscrapers, moving cars, and day/night cycles (view here). 

   2. Models for unicorns, Voxel dragons, katanas, and a palm tree built entirely from rectangles.

   3. Solar system sim with proper orbital mechanics.

   4. Fluid simulator (technically took three prompts).

   5. Interactive starship control panel that outperformed o1 and Grok 3.

   6. Song to rollercoaster converter (code).

One developer called Claude 3.7 “the closest thing to AGI I've seen,” after it built a complete Connect 4 app from scratch, complete with:

1. Mobile support.

2. A minimax-tree AI opponent.

3. Frontend, backend, databases, git integration, and deployment.

All of that (~5K lines of code) took 30 minutes and cost $10.

To us, this all means the era of “vibe coding” has officially begun. Vibe coding is what Andrej Karpathy calls it when you essentially stop writing code yourself and just... vibe with the AI.

The workflow is radically different from traditional coding: You're just seeing stuff, saying stuff, running stuff, and copy-pasting error messages until things work.

The code grows beyond what you'd normally write yourself, but for weekend projects? It works shockingly well. 

And if you want a list of tools to help you Vibe Code, someone made one here (with Vercel, who makes the vibe coding tool V0).

FROM OUR PARTNERS

Here’s one AI tool we use every week at The Neuron.

A little well-kept secret for companies crushing it with AI is that only a handful of AI tools are actually worth using.

That’s why we—Noah & Grant—use ChatGPT and Claude, along with a killer product called Attention.

Here’s how it works:

1. We tell Attention exactly what to pay attention to in our meeting (goals, budget, etc).

2. Attention listens in to our call.

3. After, Attention outputs important insights, action items, and follow-up emails.

There’s plenty more features, too. So if you sell anything, literally drop what you’re doing and go book a demo with Attention ASAP. 

Agent Tip of the Day

This is hands-down the best ~20 min overview of AI agents for beginners we've ever seen.

Tina Huang covers everything from core concepts (what makes something an agent vs just an AI) to actual implementation strategies—she even shows a completely no-code approach to build agents with n8n via a Telegram bot that manages your calendar.

TBH, we’ll probably do a full explainer on this video at some point—it’s that useful!

Treats To Try.

1. Wan 2.1 is a new open-source video model from Alibaba that turns your text into videos (that many early users say are better than OpenAI’s Sora) using just your regular computer and 8GB of memory—try it here.

2. Gemini Code Assist is Google’s free AI coder that helps you write with intelligent completions and review suggestions in your IDE.

3. Poe Apps lets you build custom AI tools without coding or paying for API calls (powered by Claude 3.7 Sonnet). 

4. Captiwate enables you to conduct instant video calls with high-value web visitors.

5. Fab’s Analyst Agent creates customized data analysts that work with your approved datasets to deliver reliable insights.

6. TestAI automates over 1K real-world test scenarios to stress-tests your AI agents.

7. Permit.io protects your AI applications by controlling what data they can access and what actions they can perform.

See our top 51 AI Tools for Business here!

Around the Horn.

• OpenAI rolled out 10 Deep Research queries per month to all Plus, Team, Enterprise, and Edu users and released the System card for Deep Research.

  • Related: Microsoft made Copilot Voice and Think Deeper free with unlimited use.

• Lovable, another “Vibe-coding” app, just raised $15M on $17M in annual recurring revenue and says it has 500K users building 25K new products a day.

• Amazon’s AI event today should reveal “Remarkable Alexa,” a $5-10/month upgrade to Alexa (will go live late March), alongside likely hardware announcements like a new flagship Echo speaker, Echo Frames smart glasses, and new Fire TV features with deeper smart home integration.

• Pew Research found 52% of workers are worried about the impact AI will have on the workplace—32% think it will mean less jobs, and only 6% think it will lead to more jobs.

FROM OUR PARTNERS

🙅Ditch your old data connectors

Data connectors are super important—and a complete waste of time for your team to build and maintain. 

Join Unstructured for a deep dive into the best practices for data connectors for genAI applications, and learn why traditional ETL requires lots of engineering overhead and struggles at maintaining context.

Register to join or get a recording.

A Cat's Commentary.