@enginerrprompt Interesting review of Gemini API's code execution. I see this as integrating algorithmic thinking into LLMs. Unlike function calls, this feature allows the model to execute code at an algorithmic level as part of its reasoning process - essentially bringing a mini Turing machine into play. It's fascinating because it addresses a fundamental limitation in LLMs. Instead of relying on increased model size or fine-tuning, it provides a mechanism for true computational reasoning. For instance, counting 'r's in "strawberry" becomes an algorithmic task rather than a pattern recognition one. I hope it remains focused on this level of algorithmic execution. Tasks like database updates or API calls should stay in the realm of function calls. This separation maintains the purity of the "algo-execution" concept, enhancing the model's core reasoning capabilities without overreaching into application-specific processes. This approach effectively bridges the gap between language processing and computational thinking in LLMs, opening up new possibilities for more accurate and logically sound responses. Btw, I posted a detailed review of this on my X account and shared your video over there.
I agree, it could really address some of the issues we had with LLMs and can really improve their reasoning capabilities. I like the algorithm-execution term that you came up with :)
@@engineerprompt We need to make this “algorithm-execution” distinction clear. It serves its purpose well and defines function-call better. Really enjoyed this, it’s the kind of thing I missed seeing more from Google. Excited to experiment with it. Looking forward to your context caching video!
Should be easy for openAI to implement this. I used their custom GPTs a while back when it came out and it would run the user’s python code/scripts in an isolated environment then analyse the result and follow any instructions you specifically set. Sounds like they already have all the ingredients to do this on the fly.
Gemini as chat sucks. I dont think it was ever meant for zero shot. I tried with code execution and it did great. At least I dont have to hand roll my own agent