llm
I have been using Large Language Models a lot more aggressively at work. There are two scenarios that seem to work really well at the moment:
- replacing web search (Google, Stack Overflow)
- filling out function bodies from signatures (like djinn but for languages du jour)
- finding my way around repositories I am not familiar with using claude-code
Is there any positive impact on productivity? Too early to tell. Major time sinks for me are navigating legacy code and debugging complicated state machines these days.
What has really gotten a boost is prototyping. In the latest company hackathon I was able to accomplish a lot more than is reasonable by both using Claude as a coding tool and embedding Claude in the final code to perform simple code-generation tasks on demand as part of the functionality.
Simon Wilison has expressed a balanced opinion that these tools are very useful, while the hype around replacing humans at their jobs or achieving AGI any time soon is likely unwarranted. This resonates with me so far.
gptel has been very nice to bring Claude closer to my preferred computing environment in Emacs. While web search is not yet supported there is a nice workaround for Anthropic specifically that lets me use it already as of today.
Deploying and monitoring LLMs as decision makers in any sort of meaningful control plane sounds really intriguing and fraught with peril, from security to reliability pitfalls. I have not had a chance to do so yet but am watching the space closely.