Coding just changed forever

Plus: Robots can teach themselves?

Happy Friday! Have you ever wanted to strike up a convo with a robot? Well, now you can: This week, Figure AI implemented OpenAI’s tech into its Figure 01 robot. In the same way that ChatGPT can now speak aloud, this humanoid robot can too…while performing physical actions like picking up trash and putting away dishes (no, seriously). While it’s not yet commercially available, I’m predicting a multitasking, chatty robot on every holiday gift guide in 2025. 🙃 

Cognition’s Devin Redefines Coding

Venturebeat via Ideogram

There’s a new software engineer on the job market…who doesn’t need coffee breaks, works 24/7, and has the coding knowledge of a thousand Stack Overflow forums combined: Meet Devin.

Some context: Cognition, a small, recently-formed AI startup, just dropped Devin: a 100% autonomous AI agent that functions like a software engineer.

What’s the buzz? Where previous AI assistants could only write naked code or suggest snippets, Devin can write entire software projects (from scratch!) based on simple text prompts—and iterates upon its output until it nails the project.

The details: A user prompts Devin with a description of the project they want to develop. Devin then crafts a step-by-step plan to tackle the project and uses common developer tools—including the shell, code editor, and a sandboxed browser—to complete it.

  • Devin can craft code, debug like a pro, and perform complex coding tasks with hundreds of steps (like fine-tuning an LLM). 

  • In demos, Devin was able to build entire websites and apps in less than 10 minutes—it could even complete actual coding jobs listed on Upwork.

  • Though Devin runs autonomously, if something doesn’t look right, users can still interfere with a command in the chat interface.

Devin isn’t publicly available yet. Only a handful of select customers have early access for testing purposes.

Why it matters: Devin could speed up development cycles, reduce costs, and free up human creativity for the more complex, nuanced aspects of tech development. It's not all about replacing human developers—it's about supercharging their capabilities. Soon, every software engineer might be promoted to manager (of an AI worker, that is).

Midjourney Debuts Consistent Characters

Rory Flynn / @Ror_Fly on X

Midjourney’s most requested feature is finally here—and it’s a game changer for creatives.

Why is this a big deal? Most image generators run on diffusion models, which generate each image as an entirely new creation. That makes it difficult to maintain specific traits or characteristics consistently across images…no matter how great you are at prompting

Enter: Character Reference, Midjourney’s new feature that allows users to maintain the appearance and essence of characters across multiple image generations. In other words, you take the same character and put it in different scenarios, clothes, image styles…you name it.

How to use it: Simply provide Midjourney with the URL to a reference image of an AI-generated character (it seems to work best with previously generated Midjourney images as opposed to photos). Then type in a prompt describing what you want your character to do/wear/etc. and add the tag “--cref” after your prompt. 

  • You can control how closely the new images reproduce your original character by adding the tag “-cw” and a number between one and 100 at the end of the prompt.

  • The higher the number, the closer the new images will follow your reference character’s appearance. 

Here are some cool examples of Character Reference in action.

Why it matters: Character Reference just made AI way more useful to creative professionals: Artists can now create coherent storylines, comics, and portfolios with characters that retain their unique identities throughout. Designers in gaming and animation can explore different environments and actions for their characters, ensuring brand consistency and character integrity. I can’t wait to see what types of content this new feature inspires. 

Free Webinar: Optimize Your Team with No Code AI

Integrate AI into your workflow in this free MindStudio webinar on Monday, March 18th at 2:00 PM PST. Learn how to:

  • Build out complete apps with AI-generated prompts and custom automations

  • Add and query your own data sources

  • Share AI workflows and invite your teammates to an organization

  • Integrate your AI application workflows with your current tools, like Slack and Mailchimp

Whether you're a seasoned AI enthusiast or a novice, this webinar will give you the skills to launch your own AI applications quickly.

Covariant’s AI Robot Breakthrough

Imagine a world where robots can learn on their own, much like how a child learns from experience. AI robotics startup Covariant is making that a reality.

What happened: Covariant just launched RFM-1, a foundation model designed to bring AI’s learning capabilities directly into the physical realm of robotics. This isn't just about programming a robot to do a job; it's about teaching a robot how to learn to do any job.

The details: RFM-1 (short for “robotics foundation model”) is basically an LLM for robot language. It’s trained on internet data as well as massive datasets of robot camera feeds, sensor data, and language.

  • RFM-1’s advanced AI algorithms enable robots to understand, interact with, and learn the physics of their environment by themselves

  • For example, users can tell a robot running RFM-1 to “pick up an apple”. After identifying the apple by relying on learned characteristics (like shape and color), RFM-1 simulates the best action through video predictions based on its training. 

  • This process is like the human method of planning actions mentally before executing them.

Use cases: RFM-1 is already being deployed in the logistics sector, where it's proving to be a game-changer in warehousing and order fulfillment processes. Covariant believes RFM-1 addresses the growing shortage of workers willing to perform highly repetitive and dangerous tasks (particularly at assembly lines). 

Why it matters: RFM-1 hints at a future where machines can learn, adapt, and evolve without needing a programmer. This opens up a new world of possibilities: from manufacturing lines that adjust in real-time to customer demands, to service robots in healthcare that improve their assistance strategies as they interact with people.

  • Limited time offer: Secure your business effortlessly with a 3-month trial to NordPass—the all-in-one password manager for team. Use "futuretools" activation code here.*

  • Elon Musk plans to open-source his Grok chatbot. 

  • OpenAI announced the addition of several board members and the reinstatement of Sam Altman to his board seat.

  • Google restricted Gemini from talking about elections.

  • President Biden wants to ban AI voice impersonations.

  • Inflection rolled out a major update to its Pi chatbot.

*This is sponsored advertising content.

More important AI news: Dive deeper into this week’s hottest AI news stories (because yes, there are even more) in my latest YouTube video:

AI hack: Here’s a cool prompt trick that gets you better results from Claude 3:

Speaking of Claude 3: Check out my full review of Anthropic’s new model (spoiler: I love it):

And there you have it! If you try out Midjourney’s new Character Reference feature, I’d love to see the results. Can’t wait to see all of your character examples!

—Matt (

P.S. This newsletter is 100% written by a human. Okay, maybe 96%.