The future of software is AI. The future of AI is software.

Plan
The journey towards this future is a long one and ambitious goals to nudge humanity towards more abundance can only be achieved if poolside is around for the long-term. This is why we feel very strongly about the importance of sequencing towards this future.
### Step one
Assist developers in building software by building the most capable AI for software development
### Step two
Allow anyone to build software by making AI-led, human-assisted interactions the next abstraction for building software
### Step three
Generalize these capabilities beyond software to all other fields
We believe that over time interactions with machines will go from human-led, AI-assisted to AI-led, human-assisted.
Our first major milestone in step 1 is to significantly surpass the state of the art that is currently held by general purpose models. To do this we’re focused on training a large language model that is entirely oriented towards software development and allowing it to improve by completing millions of tasks in tens of thousands of real world software projects.
We call this approach Reinforcement Learning from Code Execution Feedback and we’ll be sharing a lot more about it.
Some of our strong beliefs, weakly held (empirical results will show us) are:
-
To push beyond current capabilities you need to train your own foundation model
-
You can't fine-tune your way to success — major capabilities 'emerge' from training a base model and are made accurate and useful during fine-tuning
-
Scale matters, more compute and data solve for a large subset of problems
-
Not all tokens are equal — there is a lot of value in truly obsessing over the quality of our data
-
Building our own training stack allows us to iterate faster
-
Synthetic data generation — while seemingly counterintuitive, works and works particularly well for code. Over time we suspect that all data that models are trained on will become synthetic
-
Yes, larger models are more likely to show the strongest capabilities, but once you have them you can distill them into smaller models that are useful in production
-
For LLMs to improve at software development, they need to have a real world environment in which they can improve through self-play
Team
We believe big missions are achieved by small groups of people who are deeply passionate and committed to the problem they're tackling. They are resilient, low-ego, kind-hearted and bring lots of raw brain power to the table.
The team that is formed at the beginning of a company is the biggest determinant of its success. Which is why we're very proud of the team we've assembled and hope if you are reading this you'll consider joining the mission.