Sam Altman: maybe selling AI by tokens was a mistake (link)
| sienna razzle resort tattoo | 06/04/26 | | comical orchid senate macaca | 06/04/26 | | Trip self-centered fortuitous meteor meetinghouse | 06/04/26 | | ebony stain church | 06/04/26 | | Lake Frum State | 06/04/26 | | Balding legal warrant | 06/04/26 | | Trip self-centered fortuitous meteor meetinghouse | 06/04/26 | | Balding legal warrant | 06/04/26 | | Scarlet cruel-hearted international law enforcement agency | 06/04/26 | | ebony stain church | 06/04/26 | | Trip self-centered fortuitous meteor meetinghouse | 06/04/26 | | Scarlet cruel-hearted international law enforcement agency | 06/04/26 | | Trip self-centered fortuitous meteor meetinghouse | 06/04/26 | | ebony stain church | 06/04/26 | | comical orchid senate macaca | 06/04/26 | | Scarlet cruel-hearted international law enforcement agency | 06/04/26 | | Trip self-centered fortuitous meteor meetinghouse | 06/04/26 | | Balding legal warrant | 06/04/26 | | comical orchid senate macaca | 06/04/26 | | Balding legal warrant | 06/04/26 | | Purple sickened sex offender | 06/04/26 | | ebony stain church | 06/04/26 | | Purple sickened sex offender | 06/04/26 | | ebony stain church | 06/04/26 | | Vigorous walnut theater stage pocket flask | 06/04/26 | | Balding legal warrant | 06/04/26 | | hairraiser institution elastic band | 06/04/26 |
Poast new message in this thread
 |
Date: June 4th, 2026 5:53 PM Author: hairraiser institution elastic band
I don't have coding needs (until I try to make a video game). But I've used deepseek for a bunch of basic, non-legal shit, and it generally sucked
The only good point I can think of is that it wouldn't bullshit when it was caught not following the prompt or not pulling data. ChatGPT and most American models dig their heels in and seem to intentionally fuck up requests once they've been caught
An example is that I was using AI for some dividend screening help. Despite explicit instructions to conduct research by calculating yields based on the prior day's closing price, it would pull 3 and then make up the rest (about 7 in total). When called on it, it would admit to the problem of not pulling data. I would explicitly tell it to pull the data for all tickers to generate the yield data. It would regenerate the data but refuse to pull the actual data for multiple tickers each time.
After four rounds of this, it finally suggested starting all over with a clean prompt tomorrow
If the Chinese models can get better quality-wise, they will fucking eat these lazy American LLMs for lunch. But that day is not yet here
(http://www.autoadmit.com/thread.php?thread_id=5871022&forum_id=2],#49915832) |
|
|