@goodside

–

05 December 2022 20h36

In general, it can’t count the length of words reliably with naive prompting. This is a well known issue.

ChatGPT removes similar strings, suggesting hidden control codes

By

–

05 December 2022 16h15

The fact that ChatGPT removes not just this string, but any string formatted similarly, suggests there exist other control codes by ChatGPT, e.g. to mark the boundaries between user input and prior model output within the prompt context.

The string ” represents end of generation in GPT-3 tokenizer

By

–

05 December 2022 16h14

The string “” is handled specially by GPT-3’s tokenizer to represent the end of a generation. Whenever this token is emitted, the model stops talking immediately:

ChatGPT silently removes empty quote substrings from input

By

–

05 December 2022 16h14

A demonstration that ChatGPT silently removes from user input all substrings of form “” where “foobar” is any single word without whitespace:

Tokenization issue in counting letters of word ‘elementary’

By

–

05 December 2022 15h28

Yes. The correct response to: How many letters are in “elementary”? …really should be: Due to tokenization issues you can read about [here], I need to count these letters one at a time: e1-l2-e3-m4-e5-n6-t7-a8-r9-y10. Thus we see ‘elementary’ has 10 letters.

Remarkable harmlessness of ChatGPT overshadowed by shifted goalposts

By

–

05 December 2022 4h52

The posts of "Look, I got ChatGPT to be evil!" are missing the point. It's a remarkable achievement that it seems to never reply to an innocent question with a harmful answer. It's sad the goal posts shifted so fast that nobody cared when it happened.

Initial shock at LLM code errors fades with experience

By

–

05 December 2022 4h04

How much of that is just novelty, though? The first time I noticed an error in LLM code I trusted, it felt fundamentally unfair — as though the code were maliciously forged rather than just wrong. But you develop an intuition for it eventually.

Copilot increases code reliability by generating tedious unit tests

By

–

05 December 2022 3h25

I get the concern, but I don’t see it much in practice e.g. with Copilot. The code I most trust Copilot to generate is unit tests of code I’ve written myself. My code ends up being more reliable because I’m more inclined to write tedious verification code I would otherwise skip.

Extension code creates vast change in ChatGPT unlike other contexts

By

–

04 December 2022 17h57

I know this isn’t the point of your thread, but “Install this extension to make ChatGPT behave differently” is an extraordinary idea. Like, in no other context could such trivial code create such an expansive change in a third-party product.