ADHDNotes
Five highlight meanings
Colour is not decoration here. Each prefix carries a reason, while the simplest form remains a Claim.
Attention lets every token ==look directly at== every other token, not just its neighbours.Attention lets every token look directly at every other token, not just its neighbours.
==o:Rewrite the introduction before Thursday.==Rewrite the introduction before Thursday.
Is this ==p:actually true for sparse attention==, or only for the dense case?Is this actually true for sparse attention, or only for the dense case?
==g:Softmax turns the raw scores into a probability distribution.==Softmax turns the raw scores into a probability distribution.
A ==b:key== is the vector a query is compared against.A key is the vector a query is compared against.
| Syntax | Meaning | Colour |
|---|---|---|
| ==text== | Claim | #FFD84D |
| ==o:text== | Action | #F48D40 |
| ==p:text== | Doubt | #E25BA0 |
| ==g:text== | Settled | #22B979 |
| ==b:text== | Define | #12BFEF |
==A claim worth checking==
==o:Rewrite this section==
==p:Is this assumption justified?==
==g:The result is settled==
==b:Term to define==Typing == closes the markers and opens the five-choice palette. Existing Markdown without a prefix remains valid.
Twelve callouts
Every callout is an ordinary blockquote. Hue confirms the type; structure keeps the cards distinguishable even without colour.
Freeze the shuffle before the sweep, not after the first surprising result.
The evaluation drop was a tokenizer mismatch, not the model. Two days lost to a hash nobody logged.
Heads are concatenated, then projected once. The projection is not optional — it is what lets heads interact.
Come back to the interaction between BPE dropout and low-resource languages. Parked mid-paragraph; the argument did not close.
Divide the dot products by the square root of the key dimension, or softmax saturates as that dimension grows.
tokenize("unbelievable") -> ["un", "belie", "vable"] # 3 pieces, not 1
Log the tokenizer hash alongside the model hash. A mismatch is invisible in the metrics until it is very expensive.
Write the next concrete step before you stop. A useful return point should survive the distance between two work sessions.
Does the same argument hold for the multi-query variants, where keys are shared across heads?
Be able to derive the scaling factor from the variance of a dot product of two independent unit-variance vectors.
Byte-level and word-piece vocabularies are not interchangeable, even at the same vocabulary size.
Transformers qualifying exam — dry run
- Re-derive the scaling factor from scratch
- Re-run the tokenizer ablation on the held-out split
> [!NOTE]
Note
An ordinary observation worth keeping.
> [!AHA]
Aha moment
The moment it clicked, marked so you can find it in six weeks.
> [!REMEMBER]
Remember
The thing you will otherwise forget.
> [!REF]
Come back to
An unfinished thought, deliberately parked.
> [!DEF]
Definition
A term pinned down once, in your own words.
> [!EXAMPLE]
Example
The concrete case that makes the abstract one land.
> [!TIP]
Tip
Something that will save you time next time.
> [!IMPORTANT]
Important
Do not skim past this one.
> [!QUESTION]
Open question
Something you do not understand yet, stated plainly.
> [!EXAM]
Testable
Likely to be assessed. Findable the night before.
> [!WARNING]
Warning
The trap you already fell into once.
> [!PREP]
Preparing for
Seeded automatically across the days before a date.
Source forms
The GitHub-style form is the normal source:
> [!AHA] The result clicked
> The body continues inside the same card.The heading-style compatibility form is also accepted:
> ### Aha
> The same idea, rendered with the same semantic treatment.Press Enter inside the card to continue with > . Lists, ordered items, and tasks continue naturally. An empty quoted line exits.
Insert without memorising tokens
Type / to open the unified palette; callouts come first. Keyboard shortcuts Alt/Option+1…0 wrap the current paragraph or selection with the most common types. The menu also includes tables, code, rules, calendar mentions, and Mermaid starters.
Callouts on PDFs
PDF notes use the same cognitive vocabulary except PREP. Preparing for belongs to the calendar runway system, not to an immutable page. The original PDF is never rewritten; the app stores the local annotation and returns it to the exact page when the document fingerprint still matches.
Portable creation time
When ADHDNotes creates a Markdown annotation, it may add an adjacent, hidden creation comment:
==g:settled claim== <!-- adhdnotes:annotation-created=2026-08-23T18:30:00.000Z -->The comment is optional. Existing or manually written syntax is never rewritten to invent history; the catalog falls back to the source file's modification time.