The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Future AI models will eliminate natural language reasoning entirely
| ,.,..,.,..,.,.,.,..,.,.,,..,..,.,,..,.,,. | 09/03/26 |
| The Penis | 09/03/26 |
| phenotypic entropy | 09/03/26 |
Poast new message in this thread
Date: September 3rd, 2026 6:05 PM Author: ,.,..,.,..,.,.,.,..,.,.,,..,..,.,,..,.,,. they will be significantly faster than models which insist on processing in the older ways. AI 'reasoning' will be 100% black-box unauditable machine code. this seems to be the next step, since current models are very obviously slowed and made less efficient whenever natural language is involved. there is nothing but short-term competitive upside to dropping language out of the frame entirely. (http://www.autoadmit.com/thread.php?thread_id=5900357&forum_id=2),#50114179) |
Date: September 3rd, 2026 6:13 PM Author: The Penis The internal computation already is in learned statistical representations. There is no English monologue inside the network. Although the cot tokens can feed back into the computation.
|
|
Date: September 3rd, 2026 6:23 PM Author: phenotypic entropy They're "not supposed" to allow the models to do this (use neuralese in CoT reasoning) (http://www.autoadmit.com/thread.php?thread_id=5900357&forum_id=2),#50114246) |