Half way building a new company with AI in 30 days! My key lessons so far

14 days and Future Works reached 5 figures revenue, several clients in the pipeline, it is now employing its first humans and a few instances of AI!

What does all the new AI stuff have to offer? What about people? Is singularity here?

I’m nearly at a half way mark of my 30 challenge to build a new company from the ground up, solo with AI as my partner.

You can track the journey in detail on Twitter.

Here’s what I learned so far 👇

Is Singularity here?

The wild thing is that the best answer right now is… we don’t know.

How crazy is that we’ve arrived to a place where this is not a definite “no”?

The future is already becoming quite hard to predict. There is a lot of fear among most knowledge workers. Will most programmers be obsolete in 18 months? What about lawyers, illustrators and copywriters? We don’t know.

Is AI scary?

In past weeks and months before I have not had a single scary moment. Most examples online are overblown and what you get these days is just help — or attempts to help. Even free agent experiments like BabyAGI haven’t produced scary outcomes.

In the future? We don’t know. But definitely more exciting path to an abundant future than a scary dictatorship right now.

Most will tell you AI at current stage is not a silver bullet.

Not so sure. In the short span of last 2 weeks I learned to use Python scripts, multiple web libraries and frameworks, created a website, automated funnels, designed a brand, setup social media, produced 10s of pieces of content (coming!), etc.

And I only ever had about 1/2 day free for this and took weekends off. Absolutely not possible before.

Does it solve everything? Nope, and definitely not by itself, but it helps productivity like nothing I’ve seen before, even at its current still early stage.

Silver bullet for me.

So you can create whole things with a prompt now?

Also not, prompt engineering itself is fun to learn and enables really quick outcomes, but in all cases those still need to be manually adjusted, extended or completed. I don’t see this changing all that soon. Even with gigantic context ability code writing can be tricky, design needs tweaks, hallucinations come at random etc. It’s a human + machine kind of deal. Self-served-everything might be coming, but it doesn’t seem to be soon.

What works then?

It’s all about the human, the machine and underlying process. The connections are not there yet.

We need to build processes that connect the different pieces. Sometimes wild 90% optimizations come from AI, but sometimes it’s just 5% or nothing. Sometimes a human is better.

It’s patchwork for now. But one that can be iterated on. We’re seeing a lot of improvement week to week.

What is most exciting?

Recursive exploration capabilities, self-healing code, web access and plugins overall. These enable incredible amounts of possibilities. We’re yet to see a truly successful use of AutoGPT or another agent AI, but the experiments are exciting!

What is an AI agent?

AI that acts on its own, it doesn’t just respond to prompts. There are several experiments running already, with very exciting results [eg. this Stanford study] . In not too distant future bots and robots will have their own agency. It’s not as scary as it sounds.

What about people?

People are slow to trust it. One one hand we’re hearing that 100M+ people use OpenAI’s ChatGPT, on the other, most people I talk to either plateaued at very simple use of the tool or thought it’s not good enough. Prompting takes time to learn. We’ve seen interesting push back from people who said they’re faster without it, while it’s technically impossible.

Time will tell if the push back will turn into a wider movement, but so far I see many people only adopting these tools when forced by the job market etc. This is perhaps my most important observation: it’s easy to get into a bubble, but people are not fast at adopting this tech yet.

More opportunity for my AI partner and I!

Solopreneurship?

Not fun, but also very rewarding. We’ve ran fairly big teams solely remotely for nearly 8 years now, so I’m used to interacting with people only via a screen, but going solo is quite lonely.

It’s only been 2 weeks and I started added some people to the team, mostly to handle the work, but in truth, also because it’s so much better to be able to bounce ideas off humans. It is very rewarding though and fast. They say if you want to go fast, go alone, if you want to go far, go together.

Standing up a business in 30 days requires a lot of fast, but once we want to go far — we’ll definitely keep building the team.

Moments of magic

• We’ve tested 100+ AI tools. Any app that says “X in minutes” is lying, except for a few, including storywizard.com — you really can get a pretty decent custom children’s book in 2 minutes! For me a great simple example of where much more of creative industry will go.
• GPT-4 integrated in Siri provides “Siri 2.0” but right now, before Apple gets there. You can customize it to save transcripts to Notion etc. I use often on my runs.
• Self-healing code — AI notices test doesn’t pass, and proceeds to improve result, arriving to a solution without corrections needed from a human
• The right prompt in Midjourney V5 can create incredible results (But also a slightly wrong one can produce something completely undesired)
• Single-use apps are amazing and will be huge. You need something done, AI creates the app, app performs its role and can be discarded. Definitely seeing a lot of these in the future.

Some nice tricks that work immediately

• Giving your ChatGPT instances different roles is the obvious first step of prompting, but preserving them as “team mates” is the real magic. Without external storage you will sometimes need to remind it, but it’s a quick prompt.
• You can get ChatGPT to create a single use app for many operations using a Python script — manage and organize files etc. Just ask and if you get an error, simply report the error and repeat.
• ChatGPT, Claude, Bing, Bard etc. all have a have a lot to offer for a particular task, but I noticed it’s best to go deep on one and learn how to speak to it.
• You can ask ChatGPT to teach you to do all sorts of things. So far best hack ever for me.
• To feed Midjourney an image to work from, upload that image to a Discord channel, then copy link to it and enter into a prompt + write prompt after it. Saves a lot of time.
• You can use AI to create scripts to run in Figma, all the way through to manipulating layouts and content.
• Notion AI can do more than you think! (But also less than we hope)

What doesn’t work

• A lot of material online is clickbait garbage written by generative AI itself. Takes some time to sift through guides full of barely useful tips that just provide volume.
• A lot of tools and apps slapped “AI” on them to attract users or investors. They often are incredibly limited and just do templates + some shallow text from GPT-3.5 integration

The process

It really is all about the piping in-between. What is the flow between the tools? Do they communicate? What is automated, what is not? I’ll write on this in the future!

Our current stack

• Our current stack includes Midjourney, DALL-E, GPT-4, Claude, Replit, Figma, Make, Copilot X and some unique custom magic ✨

What’s next?

• 8/10 initial projects are in, time to book the rest! Typical sales cycle in this business is 2–3 months, I have 2 weeks. Fingers crossed (but also: offering some amazing pricing for the first 10)
• 17 days to go 🎬!

If you’re interested in working with us or just chatting, see future.works/#contact or DM me!