Why Is a Generative System Lazy? Maybe Money and Lousy Engineering
December 13, 2023
This essay is the work of a dumb dinobaby. No smart software required.
Great post on the Xhitter. From @ChatGPT app:
we’ve heard all your feedback about GPT4 getting lazier! we haven’t updated the model since Nov 11th, and this certainly isn’t intentional. model behavior can be unpredictable, and we’re looking into fixing it
My experience with Chat GPT is that it responds like an intern working with my team between the freshman and sophomore years at college. Most of the information output is based on a “least effort” algorithm; that is, the shortest distance between A and B is vague promises.
An engineer at a “smart” software company leaps into action. Thanks, MSFT Copilot. Does this cartoon look like any of your technical team?
When I read about “unpredictable”, I wonder if people realize that probabilistic systems are wrong a certain percentage of the time or outputs. The horse loses the race. Okay, a fact. The bet on that horse is a different part of the stall.
But the “lazier” comment evokes several thoughts in my dinobaby mind:
- Allocate less time per prompt to reduce the bottlenecks in a computationally expensive system; thus, laziness is signal about crappy engineering
- Recognize that recycling results for frequent queries is a great way to give a user “something” close enough for horseshoes. If the user is clever, that user will use words like “give me more” or some similar rah rah to trigger another pass through what’s available
- The costs of system are so great, the Sam AI-Man system is starved for cash for engineers, hardware, bandwidth, and computational capacity. Until there’s more dough, the pantry will be poorly stocked.
Net net: Lazy may be a synonym for more serious issues. How does one make AI perform? Fabrication and marketing seem to be useful.
Stephen E Arnold, December 13, 2023
Comments
-
- Subscribe to Beyond Search
Feature archive
News archive
-
