2026-08-10 at 12:10 am
bahasa rojak
However, they tied language to race, and race to money, so too bad ... kedukaan sebangsa malaysia
2026-08-09 at 11:27 pm
I summarised AI for my dad
/ I Summarised AI for my Dad /
( An improvement on my prior understanding)
Current mainstream AI modelling :
0 input text is converted to a "token" sequence from the model's vocabulary
1 so we begin with a sequence of tokens
2 the model understands tokens as being locations, in a N-dimensional space : so each token is simply stored as an N-space coordinate "vector"
3 then the model finds relationships, between each token ( in the input sequence ), and every other token ( in the input sequence ) : it consults a database of relationships learnt during training, and also accounts for the order of tokens in the sequence
3.1 the database of relationships can be subdivided into chapters, so that each chapter focuses on different types of relationships
3.2 this step is computationally parallelisable, therefore you can crunch a token sequence of length 5 and length 5000 in nearly the same time, if you have 1000x the hardware
3.3 this step outputs a sequence of "enriched vectors" i.e. each token is now represented by a more contextualised location in N-space
4 each "enriched vector" is then subject to a select subcomittee; the whole step produces a sequence of "committee enriched vectors"
4.1 again, execution is parellisable
3 & 4 : both the database of relationships between tokens, and subcommittee processes are learnt by trial&error during the model's training
5 the entire [ 3 & 4 ] operation is called a layer, and current models are 80 to 120 layers deep ; this sequential process is what enables logical branching to occur
6 after [ 3 & 4 ], are executed 80 to 120 times, a single "highly enriched vector" representing a unique location in N-space is produced ... and this is finally used to predict ONE NEW TOKEN to be appended to the input token sequence
7 the model then uses the [ input tokens + new token ] and runs back to step 1, and it keeps doing this until step 6 says "no more token are recommended"
8 finally, [ all the new tokens ] are translated back into text
2026-08-08 at 2:44 pm
the tropical suit
There needs to be a tropical friendly structured and layered format standing in for the European suit. :p make it work!
Subscribe to:
Posts
(
Atom
)
sex, tech, and mergers
this may be unsuitable for popular consumption