Tradeoffs

I always wonder.

Currently I study small details of nutrient contributions to daily productivity. Did I make an efficient gamble, by spending the first 20 years post-college, simply buying my way out of this concern? Hard to say. 

Similarly, did I make a good gamble by postponing maths study during this period?

How about the gamble of avoiding intimate relationships during college and the first few years of jobbing?

And how about the positive allocation of all that time to cultural studies of the Malaysian commercial environment, with hardly any cultural study of the Malays? Was it worthwhile?

Generally my approach, is that if I think X is not interesting, I will make extra time to double check it. All these gambles. 

Why (current) LLMs are fat, and dumb as shit

Consolidated notes over a few days, nothing really new, but the talking points were related, so here they are. Mainly response to he following propositions made by others :

❌ "We don't yet have the language to understand LLMs"
❌ "LLMs are bad at maths"
❌ "World models are timeless"

• 2026-06-06 0133
• "We don't yet have the language to understand LLMs"
• WTF is this nonsense - there are infinite opportunities to make language more complex. We have a perfectly precise language for describing what LLMs do ... it's the literal data, and algorithms operating on the data.
  
  If you say "it's not clear in English," well read the maths. If you say, "I don't understand the maths," then it's either a you problem, or a request for poetic expansion of English, which the LLMs are more than happy to to provide if you append "ELI5" to the name of the step in the demised computation 😂
• 2026-06-06 0809
• "LLMs are bad at maths"
• Just have ALUs do math. Use LLMs for guesswork. The dumbest shit architecture in the world making frontier models learn maths as if they are made of meat.
• 2026-06-07 1743
• "World models are timeless"
• You need to give an example of thinking that is time independent. All thinking is intrinsically time embedded
  
  Thinking= computation
  Thought objects = data
  
  Comp on data
  Operations on operands
  
  Concept of causality / implication depends on concept of time ( Good place to start )
  
  You're probably thinking about transformer archi as a way to parallelise the temporal dimension. It is not "independent".
• Let's go with a simple example.
  
  Say the input is a ball balanced on a pencil, standing upright on a flat surface, on a gravitational planet.
  
  Let's say the input is just a photo.
  
  Just to get from photo to physical model requires inference (comp) by the bot. Then more comp is needed to forecast the "next incremental state".
  
  There are two ways to do the comp.
  
  A : transpose the time dimension to space, which is what embedding layers + attention weights do : these contain a memory of past events, and associate the input data with similar past events.
  
  B : after that association is done, logic is performed, in the temporal dimension, at the FFN layer ( after the attention layer in the transformer ). Typically if you have a transformer doing logic, maths, physics, it is happening here.
  
  Because transformers are not Just A, but A+B, there is always a temporal i.e. logical computation step during inference. Only A is time agnostic. B is not.
  
  Currently LLMs are fat because they try to move more B to A.
  
  But as we all know, sometimes just a bit of B time will save a lot of A space, plus all the time needed to create A space in the training step.
  
  So it comes down to target workload. Fat A is not for everything.

political centrism in Malaysia

I think the centrist consensus is that "parties that cannot co-operate with other parties are not ideal, but still welcome to cooperate when they feel like it". 

So that includes all the people who are "never DAP", "never PAS", etc. we think these people are both nuts, and we welcome you to the happy middle of Malaysia.