Some Predictions about LLMs

1 min read

Some predictions about how LLMs will affect business processes over the next couple years:

  1. We will hit a point of diminishing returns w.r.t. the amount of hardware used to run an LLM + the data being fed to it v/s business value derived. When we will reach this point is not clear and can vary by task.

  2. Monstrously large LLMS with 10s of billions of params tuned on 100s of GBs of text data are probably unnecessary for domain specific tasks. An LLM that is trained to serve as a receptionist at the dentist could be run on much cheaper hardware and be trained on data specific to its domain.

  3. Businesses will seek to automate white collar tasks in order to cut down costs. Blue collar tasks will become increasingly automated within a few years.

  4. Vision, Audio, and Spatial recognition technologies will continue to improve. Research output in almost all scientific domains is expected to rise due to enhanced associative abilities of LLMs.

  5. There is an entire range of options for automating decision making - from massive general purpose LLMs - to local, domain specific LLMs - all the way to simple heuristics implemented with minimal hardware. Figuring out which solution to apply for what business problem will still require human expertise.

  6. People that own businesses that leverage AI, own the hardware that runs the AI, and build the AI, are the ones set to benefit most. It’s not evident to what extent will wealth percolate to the bottom.