Completion of this short course on DeepLearning.AI isn’t a particularly grand achievement, especially as there’s no assessment. However, the material was quite enjoyable for two reasons:
It clarified for me the distinction between code generation and tool calling in agentic systems. 💡 To the extent that an LLM can leverage composition mechanisms, code generation is more flexible and powerful, at the cost of security challenges. Tool calling feels more like webhooks or GUI callbacks to me. Definitely useful and maybe even moreso given clear guardrails.
Tickled to see that sandboxing LLM generated code revisits all of the good efforts from the Systems community over the years. From programming language work exploring limiting language runtime capabilities to the long journey of process isolation a.k.a. containerization. AI may automate away a bunch of coding, but still lots of design work for humans to do.
Full disclosure, this is a lightly edited version of content originally posted on my (gated) LinkedIn feed.
P.S. Ignore the pseudonym