I don't know if I would call it an insight, but in the process of listening to some thought leaders on the future of AI/LLM's, I'm having some realizations as to maybe what PHD level LLM, and "solve any problem" mean. Like what gets the big companies excited about AI, so excited they put hundreds of billions of dollars into it, and why it doesn't seem to align with lay-engineers experience of it.
I worked at a FAANG company, and one thing that surprised me was how many manual tasks there were. How often the solution was "ssh into this server and run a shell script". And I suspect, maybe rather obviously that the problem is just that the surface area grows that much faster than the number of ops folks you have, and so having an LLM that can maintain all those scripts. with interfaces to 100's of applications, maybe building libraries translated into 10 different languages. Those kind of problems will benefit hugely from AI, because it's code, that may already be clearly defined in an API or existing script, and needs to expand it's surface area to all sorts of different use cases.
But I think above and beyond that. There's a certain kind of problem that needs raw intelligence thrown at it to solve. I think some of the excitement at Bigco, is just that they have problems that can be solved just by throwing millions of gpu's and an LLM at a giant matrix of problem spaces. No small company would spend millions training an LLM to solve advanced algorithms to improve efficiency by 1%. But 1% at a FAANG? And growing the ability to make those improvements over and over again. Superpower.
Does this mean that software developers will be out of a job in the short term? I don't think so. I think all the things sort of lay-software engineers talk about, in terms of the problem being able to translate the customers definition of requirements into working software (including the cross-org and functional conversations, making the right tradeoffs in terms of cost and efficiency, endless meetings etc.) is not what LLM's will be good at, and this is why people will always sort of need to be in the path, because defining agents and workflows will still be a relatively human-centered exercise, need thing specific business context a certain human developer has.
Top comments (0)