CaaLM

Welcome to the official organization of CaaLM, also known as Code as a Language Model!

We build AI models that understand and work with code. Not just code generation or something.

Code execution right now and code translation as planned, we want to poke around everything that an AI might work on in code.

What Do We Care About in Our Models?

Small Sizes: Having our models small is crucial to us. This allows CaaLM to be embedded in real-world tasks and work on a device without needing the latest hardware. Sadly, because of budget reasons, it's not exactly possible for us, but we are trying with it by using workarounds.

Flexibility: One of our core advantages is that our models are designed for anything that has a vague logic syntax. Not just real-world languages. Combined with the portability mission, CaaLM could help people execute programs that resemble barely any real language.

What Do We Have Planned?

Upcoming Features!

UIML: Also known as User Interface Markup Language. UIML is designed as a universal output system that will be put onto our future models to allow them to work with code that needs user interfaces or graphical user interfaces. With the software on top rendering the output, this can expand CaaLM to a great degree.

DES: Also known as Dynamic Execution System. DES, while it sounds magical, is actually a very basic system that lets our models retain context and handle things like input and similar operations. DES works as follows: after the model ends generation, it either flags the session as DES-enabled or not. Where the software can send things like button clicks or text inputs via UIML or normal text if it's just input in a terminal program. Combined with UIML, DES could allow CaaLM to finally be useful in real-world scenarios.

Upcoming Models!

CaaLM-Translator-v1: CaaLM-Translator-v1 will be the first model of its line, called the CaaLM-Translator line. These models translate code to code, knowing only what the target language is and having no clue about the language being translated. Things CaaLM-Translator could help with are things like migrating codebases or having people with low code knowledge write vague code looking like syntax, and CaaLM-Translator handling the translation of that to a real language.

CaaLM-v2: CaaLM-2 will be the second-generation CaaLM model that will outperform its predecessor with a more sophisticated data generation system, which we haven't really considered because of budget constraints. It hasn't really been deeply planned, but we are expecting to integrate UIML and DES into CaaLM-v2.

Roadmap?

Background

CaaLM grew out of LaaLM, a project that taught language models to simulate a Linux terminal. LaaLM worked, but didn't really have that spark that got us about it being helpful, and did not have any scenarios of it being useful. So we let it rest for good.