https://futureofcoding.org/ logo
Title
g

Gabriel Grinberg

05/01/2023, 7:54 PM
Something I've been wanting to play with for a long white! A naive integration of OpenAI into Flyde It generates a file that implements a Flyde "Code Part" based on a prompt. Too slow and far from perfect, but hey.. it works 😄
Next steps will be using the fine tuning API to train a custom model using the Flyde standard library, hopefully will be faster, more accurate and even cheaper
j

Jarno Montonen

05/02/2023, 6:13 AM
I think this is the way for LLMs + no/low-code tools. Using LLMs to generate code to bring in data and to output to other systems. Been planning on doing something similar and I'll be interested to hear how training the model to use your standard library goes.
g

Gabriel Grinberg

05/02/2023, 6:17 AM
Thanks for the interest @Jarno Montonen I also think (hope?) it fits the current zeitgeist - the user take care of the high-level workings, let's LLM take care of the details. Did you plan on building a flow-based visual tool as well? or mostly focus on the LLM part? And sure, I'll definitely share the progress 🙃
j

Jarno Montonen

05/02/2023, 6:21 AM
I'm working on a structured natural language computational notebook thingy: https://futureofcoding.slack.com/archives/C0120A3L30R/p1682426173149839
g

Gabriel Grinberg

05/02/2023, 6:56 AM
oh right! saw that 😄 looks super cool
j

Jarno Montonen

05/02/2023, 7:16 AM
thanks! 🙂
n

Nilesh Trivedi

05/02/2023, 3:10 PM
An interesting direction to explore would be these blocks to be deployed as durable functions (functions that retain their state even past restart of VMs).So, something like
sleep(30.days)
would not consume any computational resources. Many microservices patterns are architected like these. I came across this idea here: https://learn.microsoft.com/en-us/azure/azure-functions/durable/durable-functions-overview?tabs=csharp-inproc
g

Gabriel Grinberg

05/02/2023, 3:17 PM
Very interesting indeed. Like a visual https://temporal.io/
@Jarno Montonen I tried fine-tuning and it failed miserably comparing to gpt-3.5 based chat completion! I just published a blogpost covering the process here https://medium.com/@gabrielgrinberg/openai-api-fine-tuned-models-vs-chat-completion-a-case-study-e3774fadc8c7 Feedback is welcome!
j

Jarno Montonen

06/02/2023, 5:29 AM
Interesting! I'll read through your post at some point