- Language models are multiverse generators
::Moire|:ϝ∷¦ϝ’s blog post on exploring language models.
- Datomic with Rich Hickey - YouTube
- Writing Datomic in Clojure - Rich Hickey - YouTube
- richhickey.md GitHub
- Rich Hickey on becoming a better developer
- Notes: https://github.com/semiosis/code-org-tidbits/blob/master/datomic/basic-query.
—
- Project code
- https://github.com/semiosis/pen.el/
After reading |:ϝ∷¦ϝ’s blog article
LMs are multiverse generators :: Moire, I
decided to take a closer look into Datomic as
a possible store for LM generations.
Something which caught my attention which Rich Hickey mentioned was the need for immutability and to not throw away data, in order to make progress.
|:ϝ∷¦ϝ also mentioned this when we talked
about keeping our notes around and one day
fine-tuning LMs on them.
Since we are now using LM’s to facilitate our creativity, we need a rock solid and distributed foundation for retaining memories. Datomic is a great candidate for such a database.
- It’s a time-travelling database
- An entire database is considered a value.
- Transactions are broken down into functions.
- A local copy is kept inside the application.
- Datomic has a lisp domain specific logic-oriented language embedded for doing querying.
- Generate graphviz and prolog from org-brain // Bodacious Blog
- Datomic is a semantic subset of prolog.
- It’s a lisp, so is naturally multidimensional in structure.
- Generate graphviz and prolog from org-brain // Bodacious Blog
- Datomic can harness storage from many different storage and database services
- Even relational databases
- Has pluggable comparators
- Two sorts that are always maintained
- Primary by entity
- Primary by attribute
- Two sorts that are always maintained
- Has both the characteristics of a document store and a column store
- assert and retract are insufficient
- So we have tx-fn (transaction functions)
- stable basis
- same query (including time), same result
- transactions well defined
- A transaction is a function to a database
- functional accretion
- Embedded Datalog
- Datomic with Rich Hickey - YouTube
- Subset of Prolog.
- guaranteed termination.
- Set oriented. Has set orientation, unlike prolog which is tuple at a time.
- A datalog program contains
- db(s) + rules + queries
- one or more databases,
- a set of rules which you can map to views
- a set of queries
- db(s) + rules + queries
- Completely data-driven interface to datalog
- has the power of relational algebra
- It’s embedded in clojure
- syntactically subset of Clojure
- semantically subset of prolog
Can read this as a map where each key maps to a list.
{:find [?customer ?product]
:where [[?customer :shipAddress ?addr]
[?addr :zip ?zip]
[?product :product/weight ?weight]
[?product :product/price ?price]
[(Shipping/estimate ?zip ?weight) ?shipCost]
[(<= ?price ?shipCost)]]}No joins – all joins are implicit.
Unification unifies all the values of customer to be the same customer as the product ~~…
unification
[#prolog]
[operator]
Essentially a combination of assignment
and equality.
vimlinks +/"?- 2 = 3." "https://learnxinyminutes.com/docs/prolog/"
If both sides are bound (ie, defined),
check equality.
2 = 3
IMPORTANT (the essence of unification):
If one side is free (ie, undefined),
assign to match the other side.
X = 3
If both sides are free, the assignment is
remembered. With some luck, one of the two
sides will eventually be bound, but this
isn't necessary.
X = Y are both free, so Prolog remembers
it. Therefore assigning X will also assign
Y. IS is another assigment operator in prolog.
X = Y, X = 2, Z is Y + 3.
See:
vim +/"modus ponens" "$NOTES/ws/logic/glossary.txt"
Database is distributed.
Brings power from database (where traditionally it would be found), to the application logic.
- The application gets its own brain (datomic database copy)
- The app becomes a peer
- The db is effectively local
- The big advantage over
core.logicorprolog- The semantics of those are tuple at a time
- The semantics of datalog is set at a time
- This means that underneath the hood, entire sets are being merge-joined
Writing Datomic in Clojure - Rich Hickey - YouTube
- Functional tests
- Simulation-based testing
- I like.
In the markdown file linked to above and in his videos about Datomic Rich Hickey tells us that tells us that mastery comes from always working slightly beyond your comfort/ability zone, pushing it ever forward.
He also designed Datomic to efficiently accrete and distribute knowledge, whilst remaining queryable. It allows you to both offload logic to the database and run logic locally which traditionally would’ve been conducted remotely on the database.
- Complete notes
- https://github.com/semiosis/code-org-tidbits/blob/master/datomic/domain-modeling-with-datalog.org
Example: github.
New users.
There are 3 users here in our system.
3 ids means they are different things.
Each row is a datum.
| entity | attribute | value |
|---|---|---|
| 11 | :user/name | richhickey |
| 22 | :user/name | tonsky |
| 33 | :user/name | pithyless |
We use pattern-matching for querying (Datomic).
A pattern in this case will be a vector of 3 elements that represents the EAV structure in our database.
Underscore means I don’t care (wildcard).
[11 _ _]
Variables start with a question mark and some
name i.e. ?entity.
I want to mention this excerpt about the origins of surreal numbers because I think there is some allegory here.
https://ianopolous.peergos.me/maths/surreal
In 1972, Conway described his system of
numbers to computer scientist Donald Knuth at
Stanford.
Knuth (creator of the $\TeX$ typesetting
system) then went away and wrote a short
novelette introducing these numbers.
It was the first time for a major mathematical
discovery to be published in a work of fiction
first.
Knuth coined the term surreal numbers; taking
“Sur” from the French for “above”.
The surreal numbers satisfy the axioms for a
field (but the question of whether or not they
constitute a field is complicated by the fact
that, collectively, they are too large to form
a set).