1
Would love to hear your thoughts on this: repo: [https://github.com/mindsdb/mindsdb](https://github.com/mindsdb/mindsdb)
blog about knowledge bases: [https://mindsdb.com/blog/mindsdb-knowledge-bases-revolutionizing-ai-powered-data-queries-with-sql-algebra](https://mindsdb.com/blog/mindsdb-knowledge-bases-revolutionizing-ai-powered-data-queries-with-sql-algebra)
So, casually trying to make LLaMA achieve consciousness (as one does on a Tuesday), when I stumbled upon something hilarious. Turns out, you can make these language models "reason" with about as much code as it takes to write a "Hello World" program. No, really!
https://colab.research.google.com/drive/1jfsG0_XP8a5mME76F5a6xeP9uu-tvDZ3#scrollTo=fNg798sHpuqk
Here's the philosophical crisis I'm having now: When you ask an LLM to code something vs. asking it to reason about something... are we basically watching the same neural spaghetti being twirled around?
The real question is: If one can make an AI model "think" with 5 lines of code, does this mean: a) An LLM should be able to write its own reasoning code b) We've been overthinking AI? c) The simulation is running low on RAM d) All of the above
Would love to hear your thoughts, preferably in the form of recursive functions or philosophical paradoxes.
P.S. No LLAMAs were harmed in the making of this experiment, though several did ask for a raise.
However. Instead of asking an LLM to "do the whole thing" (which is indeed prone to inconsistency) and thus letting LLMs run wild like unsupervised toddlers maintaining the "do not press" buttons at a nuclear facility. A FLAT approach puts control and predictability to LLM interactions by treating them more like traditional programming constructs but enhanced with LLM's natural language understanding. like:
- Binary decisions (gates) - Limited classification (using match/case) - Structured data extraction (using Pydantic models) - well typed function calling
Anywho, Would love to hear your thoughts on an experiment F.L.A.T (Frameworkless LLM Agent... Thing) https://github.com/mindsdb/flat-ai
Showcasing that it is possible to leverage the power of LLMs in Agents though absolute simplicity:
So, we are running an experiment, a Poly-LLM Service
(you can play with it and read more at https://mdb.ai/llm-serve).
What's exciting? - Unlimited tokens!!!
To make LLM-based applications more viable, we're exploring the possibilities of a world where we developers don't have to worry about price per token. Instead, as developers we want to focus on identifying the LLM that best meets our needs, considering various trade-offs such as throughput, context window size, and domain knowledge (e.g., coding, logic, etc.). For this small release, we are starting with GPT-3.5-Turbo and comparable models, mostly because they have a great balance between quality, throughput and context window, making them viable for many future production applications!
- A single universal API for the most popular LLMs: The ability to query all models, including anthropic and gemini models, through the same openAI completions API standard, this helps maintain clean code. We handle the translation and brokering for you. In this release, we support models, like: GPT-3.5, LLama2-70b, CodeLlama-70b, Mixtral, gemini-pro, dbrx, etc (a complete list https://docs.mdb.ai/docs/api/models), although the idea is not new, we felt it was important for there to be a service that was not opinionated between open source and closed source models
Once again, we look forward to hearing your ideas and feedback.
You can create a Text-to-SQL skill, connect it to an agent, and publish a chatbot to Slack.
Here's the demo.
https://www.loom.com/share/7c206876892e4ae688d9e6d3ad94b011?sid=75d521ec-15c0-448a-bba2-cbeea6ee7bd3