i am copying text from Google Gemini, and when i paste into Bike, some of the text disappears. When i paste into a bbedit, as a control, all the copied text pastes correctly (in markdown format)
This bug only happens when the answer is formatted like above, with numbered, indented subtopics.
When pasted into bbedit, the text appears in markdown format (i.e. instead of bold, the bold text is bounded by two asterisks):
Of course. The two teams who played in last year’s NBA finals (the 2024 finals) were:
The Boston Celtics: The dominant team of the 2023-24 season, finishing with the league’s best record (64-18). Their success was built on the star wing duo of Jayson Tatum and Jaylen Brown, complemented by an elite defensive backcourt with Jrue Holiday and Derrick White, and the versatile offence of Kristaps Porziņģis. They were known for a powerful, balanced attack and a smothering, switchable defence.
The Dallas Mavericks: A team that surged in the second half of the season after acquiring P.J. Washington and Daniel Gafford at the trade deadline. They were powered by one of the most dynamic backcourts in the league, with the offensive genius of Luka Dončić and the scoring prowess of Kyrie Irving orchestrating their attack.
At other times, i have lost the beginning few lines of the second item (but i’m unable to make a reprex prompt).
Finally, here is the same text after i paste into bbedit (except it’s in markdown format), so the text did make it into my mac’s clipboard:
Excellent question. The size of the adjustments is determined by two main factors:
The Magnitude of the Error: A bigger mistake leads to a bigger adjustment. If the model was very confident about the wrong word, the error signal is strong, and the resulting changes to the vectors are large. If it was only slightly wrong, the adjustments are tiny.
The Learning Rate: This is a pre-set parameter that acts like a throttle. It scales the size of all adjustments. A high learning rate means the model makes big, bold changes, while a low learning rate means it learns slowly with small, cautious updates.
Essentially, the final adjustment is a product of how wrong the model was multiplied by how fast you’ve told it to learn.
Actually I can reproduce it now. Seems to be a problem importing HTML. When you copy from Gemini it puts both plain text and HTML on the clipboard. Bike is picking the richer format (HTML) by default, but seems to have an import bug there. I’ll try to fix.
In meantime if you Paste as plain text then I think it will work.
I’ve just posted a new preview that I think “fixes” this:
Not a perfect fix, but better. When you paste the text now you should see each row is a proper ordered row type. Unfortunately the pasted text also contains some empty rows, which in Bike 1.x resets the counter. So each row will start with 1)… but the row types are correct and if you delete the empty lines in between then you’ll get 1), 2)
Also in Bike 2 I think it won’t have that problem.
Anyway please give the Bike 192 preview a try and see if it fixes the problems you’ve seen.
Looks largely solved, but there’s extra space before the last paragraph. Don’t know if that’s something in the html because when pasted into bbedit, there’s only one empty line.
Excellent question. The size of the adjustments is determined by two main factors:
The Magnitude of the Error: A bigger mistake leads to a bigger adjustment. If the model was very confident about the wrong word, the error signal is strong, and the resulting changes to the vectors are large. If it was only slightly wrong, the adjustments are tiny.
The Learning Rate: This is a pre-set parameter that acts like a throttle. It scales the size of all adjustments. A high learning rate means the model makes big, bold changes, while a low learning rate means it learns slowly with small, cautious updates.
Essentially, the final adjustment is a product of how wrong the model was multiplied by how fast you’ve told it to learn.
I think that’s just in the HTML… or more precisely in that NSAttributedString that’s generated from the HTML. You can see what I am reading by pasting the copied text from Gemini into a rich text TextEdit document.
