Diary 2023-11-11
https://gyazo.com/f076243ef41b3a971f817fa488c79e21
16704: NG
100: OK
1000: OK
8352: too much text
4176: too much text
2000: too much text
syatekiya931 An acquaintance of mine who used to be a yankee told me a saga of how he used to get chased by motorcycle policemen, but how he always gave them the slip. One time, a guy next to him at a bar was a motorcycle policeman who said, "Those guys are not good drivers, so if you chase them around too much, you might get into an accident, especially if they're not good drivers and are dangerous. He was very merciless. RecruitKotesash These days, even motorcycles are equipped with driverless recorders, so if you track a vehicle to some extent and obtain evidence of a violation, you probably don't pursue it by force and arrest it at a later date. If they have evidence of a violation, they will probably investigate and arrest if necessary. Even if the license plate number is bent, the police will be able to tell by the type of vehicle and the characteristics of the body if they are so inclined. However, the problem is that they don't have the motivation to do so. nishio OpenAI is playing the meta-game geniusly well, giving tools and romance and leveraging all the humans in the world to maximize the number of attempts. ono_shunsuke It's a leverage that's only possible today because of the development of social media and the ease of viral incitement. The invention of the strategy of focusing on tools and incitement and using the resources of all mankind to make a number of attempts. Especially the genius of realizing that the agitation is super important.
OTA57 You are a role model for advanced technology players... We need technical skills that people all over the world find romantic at the same time.
It's really amazing that one or the other tends to be lacking.
mr_bay_area In this age when OpenAI regularly throws in some great flying pieces of technology, I think the important thing is "to be prepared to abandon what you have been doing". I think the important thing to do when you feel that what you've been doing is about to be replaced by a stronger technology is not to throw up a poem in X, but to look for the best way to utilize that technology. ono_shunsuke There is an argument that LLM utilization services should be a chat UX, Just as there are times when chat is useful and times when non-chat UX is useful in human collaboration, both can be the right answer for LLM services.
Sharp thinking about what issues need to be solved, and whether chat flexibility in particular can be poison or medicine for them.
・Start with a soft request in chat (e.g., a good human being)
・Based on a slapstick, fine-tuning is a collaborative editorial UX (e.g. Google Docs, GitHub)
・Test when changes are made (e.g., automated testing, QA)
・Version control (e.g. Git)
The chat UX is the most powerful tool to create a good beat.
nishio This is really true, and "work that could be accomplished with chat at the moment but is not using chat" is likely to have some "benefit of not being in chat". There is a high possibility that there is some "advantage of not using chat". ono_shunsuke Exactly! It's good to see how the problem you want to solve is being solved in human society now! Tasks that are solved by more standardized communication (non-chat) even between humans may be better solved that way when using agents. The reverse is also true, there are many tasks where chat is by far the most convenient
nishio ChatGPT was able to broaden its audience by drawing on the "existing UI that users are familiar with" called Chat, but it cannot be used beyond a certain point if the mental model of chat remains the same. However, the mental model of chat cannot be utilized beyond a certain level if it remains the same. For example, it is not possible to edit the history to change the conditions and retry, which is not done in human chat. ChatGPT was able to expand the range of users who are not good at learning a new tool by drawing on the "existing UI that users are familiar with" called Chat.
but on the other hand, if the mental model of the chat remains the same, it cannot be utilized beyond a certain
For example, editing the history to change the conditions and retry is not done in a human chat.
For those who use the latter, the GPTs include support by GPTs for the "retry with different conditions" itself.
GPT will update the prompt itself.
Then the next layer is "the ability to see the prompts generated by GPT and debug them if necessary".
---
This page is auto-translated from /nishio/日記2023-11-11 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.