Creating an industrial GPT from proprietary sources
At OSS Ventures, the leading startup studio builder specialized in manufacturing, we try to be on the edge.
So when Sam Altman released GPTs, assistants focused on one task, and leveraging unique database, we had to do it.
Being a full-time nerd and part-time CEO, I decided to pull the proverbial all-nighter and release it for the world to see.
Part 1 : A one-night project, 4 years in the making
At OSS Ventures, we are the leading Studio in the world for operations. Operations is everyone wearing a hard hat, gloves, safety shoes. As a Studio, we create startups alongside amazing founders, providing the idea, coding capabilities, growth professionnals, and letting the company work alone when the startup does 500K ARR.
Our team of 20 people has visited 860 factories, deployed 1200 software in factories, started 20 projects, launched 15, of which 8 performed a series A. Enough to say, we’ve been seeing our fair share of factories.
But there’s a catch.
I’m a nerd, former exited founder of an AI company. And former factory director. As such, there are few things in life that I like more than data and processes. So, from first day on, we put in place strong processes, leading to a unique dataset.
User interview database
Each time any of the 20 OSS Ventures employees spent time with a user, stakeholder, it got recorded. And generated a transcript. Tagged with a set of tagging words that are relevant to us.
This database has around 11K hours of interview total and 2K other insights (written insights by the team).
We zipped the file and moved on.
All startups ever created at OSS Ventures
Every single startup launched has a standard template with all relevant informations in it. It’s in the form of a notion template, duplicated.
The result of that database was a 500 mega file.
Part 2 : actually doing it
Honestly, the experience was deceiptevely simple.
Step 1 : uploading the content and stating the purpose of the bot
By just uploading the different files (bulk) , GPT was able to digest all informations.
The purpose of the bot was made using Custom instructions best practices.
Step 2 : fine-tuning
I started playing with the sandbox. Some responses were off, some were not satisfactory.
I spent around 3 hours to fine-tune the different answers. It grew the custom instructions and uses cases.
Step 3 : the results
The results are mind-blowing, and significantly better than the regular GPT-4 engine we use every day, even with carefully crafter custom instructions (I have personally never seen a warehouse director with an MBA).
Step 4 : early communication
A simple linkedin post gave us the way to stress test the model by having roughly 100 conversations.
Early use met positive feedback.
Reflecting and next steps
The whole experience was otherworldly. It is very clear to me that this shift in computer science is here to stay, and will meet with unprecedented demand.
The whole ecosystem, willingness to share proprietary data, is still unclear and unfolding.
We cannot wait to be part of that.
Try out industrial innovator here : https://chat.openai.com/g/g-99cmbWKeR-industrial-innovator
If you are interested in the world of technology for manufacturing, drop us a line at maroussia@oss.ventures , or visit us at www.oss.ventures / 4 rue du Vertbois, 75002, Paris !
