5 ms·

It's amazing how ChatGPT makes code the reader may not be familiar with faster to grok. The examples in this article assume familiarity with SIMD. In the past

by ijidak 2y ago

It's amazing how ChatGPT makes code the reader may not be familiar with faster to grok. The examples in this article assume familiarity with SIMD. In the past, I would have tabled an article like this for deciphering later. But, with ChatGPT, I was able to ask about just a few bits of code and immediately grok the implementation. From a technology standpoint, the future is looking scary productive. Can't wait to see what new vistas developers open up with these new super powers.

secondcoming 2y ago
you’re assuming the output from ChatGPT is correct, and it’s not bluffing.
- spartanatreyu 2y ago
  Bluffing is the wrong word. It assumes that the LLM is capable of knowingly telling falsehoods. Instead it always has full confidence in whatever it's last "thought" must be correct, even when asked to double check it's own output, even in a loop. It'll either doubledown on a falsehood or generate a new result and have complete confidence that it's last result was wrong and that it's new result must be correct, even though the initial confidence was the same. "Bluffing" certainly isn't the right word. "Hallucinating" fits much better.
  - kelnos 2y ago
    > By comparison, modern C# is memory safe all the way down, possible to implement even very complicated data structures without a single line of unsafe code "Confidence" is the wrong word. Confidence is a capability of thinking, sentient beings, and LLMs are not that. I say this to point out the futility in your post; "bluffing" is a perfectly fine thing for the GP to say, because it gets the point across. All of us here should know that an LLM can't actually bluff. (And if you'd prefer a different term, remember that not everyone on HN speaks English as their first language, and may not immediately come up with a more appropriate way of phrasing it.)
anonymoushn 2y ago
LLMs are completely useless for explaining code that uses shuffles, let alone code that uses lookup tables of shuffles or uses shuffles in unconventional ways.

by ijidak 2y ago

secondcoming 2y ago

you’re assuming the output from ChatGPT is correct, and it’s not bluffing.

spartanatreyu 2y ago
Bluffing is the wrong word. It assumes that the LLM is capable of knowingly telling falsehoods. Instead it always has full confidence in whatever it's last "thought" must be correct, even when asked to double check it's own output, even in a loop. It'll either doubledown on a falsehood or generate a new result and have complete confidence that it's last result was wrong and that it's new result must be correct, even though the initial confidence was the same. "Bluffing" certainly isn't the right word. "Hallucinating" fits much better.
- kelnos 2y ago
  > By comparison, modern C# is memory safe all the way down, possible to implement even very complicated data structures without a single line of unsafe code "Confidence" is the wrong word. Confidence is a capability of thinking, sentient beings, and LLMs are not that. I say this to point out the futility in your post; "bluffing" is a perfectly fine thing for the GP to say, because it gets the point across. All of us here should know that an LLM can't actually bluff. (And if you'd prefer a different term, remember that not everyone on HN speaks English as their first language, and may not immediately come up with a more appropriate way of phrasing it.)

anonymoushn 2y ago

LLMs are completely useless for explaining code that uses shuffles, let alone code that uses lookup tables of shuffles or uses shuffles in unconventional ways.