Listen now | Code Augmented Inference is All You Need? How ChatGPT's new Code Interpreter model and moving towards "Code Core" architectures is enabling the next great leap. Plus: Our chat with >10k AI Engineers!
Great conversation! Code Interpreter is really exciting. I'm still able to get OS command execution via `subprocess` or other libraries very easily. Also, it seems that after a few successful command executions, it will execute commands from natural language - no need to pass "Please run import subprocess; subprocess.call()". With a prompt like "How many users are on the system?", the code interpreter will pick the shell command itself and run it through subprocess. Very interesting stuff.
The intro to this essay is not just a summary of the Twitter Spaces session (which was quite good btw). It is an argument that Code Interpreter is a fundamental leap forward, not unlike the leap forwards we’ve seen from foundational model releases. Well reasoned, and A+ on the footnotes / links.
My only add is that Plug-ins and Code Interpreter have proven, somewhat paradoxically, to be opposites.
ChatGPT Plug-ins released with major hype and fanfare but never quite found PMF, while Code Interpreter had a super quiet release and yet seems to have achieved PMF within 6 hours of people using it.
Is the market signaling something here?
Perhaps Plug-ins was merely a feature while Code Interpreter, as your essay says, is a breakthrough.
haha thank you! i put a lot of effort into the footnotes but i'm not sure people even read them.
yes exactly, the market is telling us that so far "code core" approaches are more useful than "LLM core" - which makes sense if you consider that we are still in very early days of LLMs and anything that requires the LLM to be the execution engine is limited by LLM capabilities which we don't yet have words for
Great information! I have a question: When using Code Interpreter to analyze probability and statistics data, can the final result allow Code Interpreter to directly generate a self-operating software for us (that is, implement the program: write the code for the data analysis program, including any necessary user interface and input/output mechanism, which we can then download to our PC for easy use.)? Thanks!
Great conversation! Code Interpreter is really exciting. I'm still able to get OS command execution via `subprocess` or other libraries very easily. Also, it seems that after a few successful command executions, it will execute commands from natural language - no need to pass "Please run import subprocess; subprocess.call()". With a prompt like "How many users are on the system?", the code interpreter will pick the shell command itself and run it through subprocess. Very interesting stuff.
great tip!
The intro to this essay is not just a summary of the Twitter Spaces session (which was quite good btw). It is an argument that Code Interpreter is a fundamental leap forward, not unlike the leap forwards we’ve seen from foundational model releases. Well reasoned, and A+ on the footnotes / links.
My only add is that Plug-ins and Code Interpreter have proven, somewhat paradoxically, to be opposites.
ChatGPT Plug-ins released with major hype and fanfare but never quite found PMF, while Code Interpreter had a super quiet release and yet seems to have achieved PMF within 6 hours of people using it.
Is the market signaling something here?
Perhaps Plug-ins was merely a feature while Code Interpreter, as your essay says, is a breakthrough.
haha thank you! i put a lot of effort into the footnotes but i'm not sure people even read them.
yes exactly, the market is telling us that so far "code core" approaches are more useful than "LLM core" - which makes sense if you consider that we are still in very early days of LLMs and anything that requires the LLM to be the execution engine is limited by LLM capabilities which we don't yet have words for
Thanks Swyx, I like your nomenclature and it matches experience. Everything I’ve built that doesn’t break so far is code core.
P.S. I was talking to someone at the Pinecone conference last week about this essay and they also mentioned the footnotes, haha. Keep it up!
This Substack is a great resource to keep up-to-date without wasting hours on Twitter/hacker news/etc.
sadly i have to waste my hours in order to do it 😭
So we don’t 😎
Great information! I have a question: When using Code Interpreter to analyze probability and statistics data, can the final result allow Code Interpreter to directly generate a self-operating software for us (that is, implement the program: write the code for the data analysis program, including any necessary user interface and input/output mechanism, which we can then download to our PC for easy use.)? Thanks!