LLMs as we know them now are absolutely going to disappear. There won’t be nice free text interfaces to models with trillions of parameters. That will seem obscenely extravagant.
n LLM is fantastically expensive to run: OpenAI makes a loss of over 10 billion dollars per year, and loses money even on its highest-oaying customers.
There’s a reason why Anthropic and OpenAI were competing to sell their war-crimes-as-a-service to the Department of War. Private investment is starting to dry up and the Pentagon is one of the few places with the sort of money an LLM needs to keep alive.
Using generative AI is an extremely inefficient way to do most of the things the world uses it for at the moment. It only appears efficient because the AI companies are giving it away for free and eating the loss, burning through capital as fast as they can.
What will be left? Small local models that will go out of date quickly since they won’t be continuously retrained. Still ok for writing boilerplate code and such, but not the thing that seems like it can replace every job on the planet. Generative AI will be seen as a much more limited tool.
Data centers that could be useful for lots of SaaS tasks that don’t involve running supercomputer GPUs at full pace 24 hours a day so they fail and need to be replaced like vacuum tubes in a 1950s computer.
A countermovement called something like “slimline productivity” where doing stuff without AI becomes trendy. No need to spend so much on AI, just get your employees to do it by hand! Cheaper, more reliable, more human! After a few years everything is advertising itself as “slimline”.
An enormous amount of work for software engineers who know how to work without AI, debugging the terrible AI-generated code that we’ve flooded the world with… if anyone can pay for it.







Or he’s working and uses the pomodoro method