Pinned Tweet
A4 List: Printable daily task list that lasts two weeks
Form chains of progress over two weeks
Design your own sheet
The collapse of abstraction
In my mind “abstraction” is the art of defining “language levels” where the next layer up allows you to combine the lower layer primitives with fewer symbols (“conciseness”). Example layers are: an interface/SQL tables, your language, and machine code.
That system of abstraction has an interesting property: you implicitly define an “infinite set” of possible *valid* programs in that syntax.
When you write a program, you are merely pulling it out of the infinite set of possible valid programs. Your snowflake program already existed, and so do all past and future iterations of it.
From the moment you define your abstraction/language/interface/API, you have created an infinite set of possible programs. The abstraction means:
(1) The symbols used to define it at that layer are shorter, and leverage the human mind's penchant for language.
(2) The program is more likely to be right as you have removed the possibility of incorrect states.
Well designed abstraction pre-LLMs was good engineering.
But LLMs remove the need for 1 and 2 above.
No human needs economical reading and writing because the LLM does it (1).
And there is less need for (2) as it is the LLM that is managing the reading and writing of state machines, so its guard rails are different (likely unit tests to externally verify behavior of the black box code).
I think abstraction is going to collapse, especially the thin abstractions that only existed for 1 and 2 above.
Instead of ORMs LLMs will just write SQL directly and do the mapping at “generation time” manually.
I believe abstractions will settle on to “token efficiency” boundaries: is it considerably more LLM token efficient to use this high level language?
Elon Musk thinks LLMs will generate machine code directly, but I do not think that is the case for high level applications as the token cost would be substantially higher, and reduces the app to a complete black box vs high level languages.
David Deutsch also mentions that chess grand masters seem to be able to find the optimal set of moves out of the possible infinite set much faster than brute computation would imply was possible: he thinks the brain has some yet unexplained mechanism for efficiently finding answers that computers do not currently have.
This is another reason for abstraction: to leverage the mind's creative ability to find answers in that language that raw computation may not have access to via brute force.
It is quite interesting to see LLM created servers converge onto Rust and not Go as the language.
Go compiles faster so the LLM can get compile time feedback faster.
The AST is also less complex, so likely needs fewer tokens.
Gets you to 90% of Rust's performance.
Go is also domain specific to web servers which means the stdlib is world class for networking code, and any related libraries you need are likely ready to use.
The code compiles to machine code, and the runtime uses all cores.
No need to select an async runtime like in Rust.
just one more fold bro
xitter.cf/leRaffl/status/2098128…
That, my dear friend, is not a coincidence. It is mathematical paper genius. Yes, I said paper.
Here's why:
The ratio isn’t actually 1:1.4.
It’s 1:√2 where √2 ≈ 1.414213562.
√2 is the exact aspect ratio of DIN A / ISO 216 paper (A4, A5, A3…). Yup, that's right, the metric one that everybody except the USA uses.
Fold an A4 sheet of paper in half and you get an A5. Same shape, just rotated 90°. Fold again → A6, then A7, A8, etc.
Works backwards too..
In the case of paper A0, it is defined as having length and width such that the area is exactly 1 m², just in the ratio 1:√2. Simply Beautiful.
The ratio never changes and this is EXACTLY what iPhone Duo needs!
Closed: 84.1 × 117.8 mm → ~1.40
Open: 164.6 × 117.8 mm → ~1.40
Height stays. Width doubles.
And this is the only ratio this works with. Here's the proof:
Let a rectangle have sides 1 and r > 1 (wlog).
Halve the long side. The new sides are 1 and r/2.
For the new rectangle to be the same shape, the ratios must match:
1 / (r/2) = r / 1
=> 2/r = r
=> r^2 = 2.
=> r = ±√2,
where the negative solution is simply a length in the opposite direction and can thus be ignored.
This is the unique positive number to do this. The only ratio that stays the same when you double one side or halve the other is the diagonal-to-side ratio of a square. That number is √2.
It's why metric DIN paper is the ideal paper shape.
And it's why the iPhone Duo is the ideal Foldable shape.
The decimals after 1.4 can be dealt with / hidden with a few pixels. That's basically a rounding error. Annoyingly but not a big deal, you can't machine to that many decimal places anyway.
enzo retweeted
If your child becomes a reader, about 80% of the education job is already done. That's my honest assessment after working in education for over thirty years. Everything else is secondary. Most parents think science education is important. Yes it is. But if you can't read the biology textbook, you're not going to learn biology.
Reading is the meta-skill that enables all other skills. History requires reading. Science requires reading. Even math increasingly requires reading as it becomes more sophisticated. The child who reads voraciously will figure out everything else. The child who doesn't will struggle with everything.