relyonai 0.1.2

Python Runtime Inline AI Calls

relyonai

Python Runtime Inline AI Calls

• 🤖 AI language extension.
• 📉 Reduces code size.
• 🐍 Can use python runtime objects.
• 💿 Caches calls, so it's fast on re-runing.
• ✅ Fixes mistakes internallly.
• 🎁 Saves runtime context for reusing.
• ⭐️ Uses Few-Shot and COT under the hood.

Warning
See Reflections.

Usage

pip install relyonai
export OPENAI_API_KEY=...

from relyonai import ai
f = ai('get print hello world func')
f()

from relyonai import config
config.model = 'gpt-4'
config.clear_cache()

See tests for more.

Reflections

Although the idea is working to some extent, it remains largely impractical for the time being. It falls short in a few crucial areas:

• Reliability and Consistency: The code that's being generated is not consistently accurate, with minor prompt modifications often leading to significant changes.
• Controllability: GPT-3.5, which was the primary model for the development of 0.1, practically ignores system messages and often fails to follow few-shot examples, even for simple calls.
• Privacy and Security: This conserns could be largely mitigated by clever prompting and potentially runtime sandboxing (like patching Python's open call or something similar).

GPT-4 is expected to be a major leap over all axes with usable clever system messages and better Few-Shot/COT, but it needs to become publicly accessible, approximately 5x cheaper and roughly 2x faster to be practical for this task.

0.2 Plans

Note
this is a rough draft

• improve quality
  • gpt-4: new system prompt and better cot
  • more tests + add reliability tests
  • include invocation context (ai, ±5 above/below exprs)
  • include project info / files / pip / hardware / os / time / etc.
  • final result check
  • different prompt for different py versions
  • other runtimes: bash, remote, c++, browser, etc.
  • bootstrap the package itlsef (just for the fun of it)
• better introspection
  • rich dialogue printing (+spinner)
  • much better error handling
  • view source code/doc
  • more granular sessions (trees)
• advanced caching
  • redis
  • limit size (up to ~20mb)
  • save pre-compiled code
• async
  • aai
  • agpt
  • aemb

License

MIT