Diary 2023-08-23
Lab Youth Training Camp Day 3
https://gyazo.com/4085f4382269068f3ddd3b0e8387de53
Suppressing pages that have already appeared should be limited to those appearing in vector searches.
If you happen to pull something at random, but it seems relevant, you should again do a vector search on that content and use the most relevant part, so
No vector search.
I'm running it on GPT-3.5, and it's only up to 4K.
Use the 16K version or
Then, on the other hand, it's like, "Well, we've got a lot of space, so let's put more information in.
But then the ratio of "what you read" to "what you write" would increase, and you might find yourself disappearing more and more!
Maybe it will.
In the end, what matters is the amount of your output.
Oh, I thought I was using 3.5, but it was 4.
Assuming IO to Scrapbox, there's only one input/output port and error messages come to the same place, but with that mechanism, if I loop it mechanically, I'll read the error message and write the rest.
Oh, with 3.5, I sometimes translate what I read and write it down as I read it.
Once you do that, you look at it and do the next one, so it's like a translation, not a note.
Rough estimate, it costs about $0.48 per step, and each book is going to take about 100-500 steps, so, well, that's a pretty expensive mess.
The fill_with_related_fragments system is now returning a string that fills the given token and a list of the titles of the pages used, but this should change.
Because now it's got that "stack the relevant stuff and random chunks for the rest" behavior.
In this implementation, I want to pass half of the frame and pour the text data into the remainder and the other half.
That kind of flexibility in the current design is not good.
There's a whole bunch of code that shouldn't be tightly coupled because it's tied to the interface in the beginning.
I forgot to push again today and the code for periodic execution is a bit old...
push and manual trigger, or
Let's add lots of heads with multi-heads.
Generating a story about what to read on your phone on the train ride home from camp.
Changed cron to once an hour.
Let's see if we can make it by the time we take off.
Perspective of ✅ title
✅ Pile on the NG list of vector searches.
Increase the percentage of ✅vector searches
OK for now, since "up to 3" has been changed to "3".
I would like to do some initial, balancing in the future in case of low input.
Distinguish between titles picked up at random and those picked up in a vector search.
leaving (fleeing) the venue
I made a lot of parallel notes for now.
I'll have to remember to stop this one tonight...
It runs once an hour, so it's going to be tough tomorrow morning.
The theory that we are already in a lot of trouble even now.
TODO Back to once a day when you get home.
Good to run many themes in parallel.
If you remove the 🔁 mark, the auto-renewal will stop.
But I'm thinking that explicitly cutting the mark is not a very good experience.
Wouldn't it be nice if the frequency slowly decreased?
Gradual increase in span.
If a human reacts, it goes back to day one.
Generation date and time are indicated.
If the update date is later than the generation date, it means that the edit was made, so it will be updated at the next regular timing.
If not updated, updated only if the "difference between the last generation and the previous generation" multiplied by 1.5 is exceeded
So it would be 1,2,3,5,8,13, and so on.
When an error occurs, all the fragments are output, so I have to delete them, very tedious to do from my phone.
If you didn't turn it off, you should still get an error the next time the context width is over, which is a bad implementation w
I just counted and you have 20 pages of parallel topics, you are making too much!
Some of them are keywords that I learned today from listening to the LT of the Lab Youth people.
Some of these keywords have been mentioned for years.
GPT4 can also be answered since it is a general keyword
Some are keywords such as Kozaneba so raw GPT4 is not answered
So far you've answered well, picking up information from Scrapbox.
I get a release notification from Github at 16:07, but 20 GPT4s run after this, huh? Then an overwrite runs in about 30 minutes? Hard to write...
I still think we should move during sleeping hours.
16:58 I thought maybe an update would come while I was lost at Yokohama station, but it didn't.
Is this down to some kind of error?
It hasn't even fallen off!
https://gyazo.com/ab7d0180a5019f3dce73d8847244610d
I think you may have committed with pdb.set_trace on without looking at it carefully again.
Doesn't sound like it.
I'll have to look at the logs to be sure.
Or work tethered from home on the next transfer.
I don't know if such a place exists.
Stopped the action running.
I thought I changed the schedule and forgot to push again...
I got an email notification that it was released.
Debugging, mystery solved.
Embedded API failure retries endlessly
So why do they fail?
Previous Note" is an empty string.
How did that happen?
A case where the GPT4 API failed and the error log was written as a previous note, and I didn't want it to be read as a note, so when I deleted it from my phone, it left a blank line
Hmmm, what to do.
A: Fundamentally, it is not good that the error log can be read as a note
B: When there is only a blank line above the marker, it is strange to make a blank line the previous note, which can be caused by inadvertent mistakes in the future.
C: It is strange to do a vector search when the previous note is really empty in the first place
The original design was randomly filled in, but we decided to ignore that use case.
Even if it is empty, the title exists, so use it.
D: It is strange that the embedded API retries infinitely when an empty string is really passed to it
Immediately throws an exception because it is strange that an empty string is passed in the first place.
The original plan to interact on the train on the way home was a dud.
I looked through all of them anyway.
What is turned off
What you think is still subtle may turn out to be interesting if you run it for a while.
Now, I'm doing more vector searches, so when I run out, I'll start taking what I don't have.
---
This page is auto-translated from /nishio/日記2023-08-23 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.