Jhet Chan

[Jhet Chan @jhetchan]

taste is not one word.

it's a function of two words, ie:
what is good and what are the constraints to get that

ai can see, can read, can write, can speak …
but it literally can't taste 👅

and until then
humans shall remain the long pole in the tent

[Jhet Chan @jhetchan]

“what should be done? “ is too sloppy

try “what must become true, remain true, or be proven false? ”

[Jhet Chan @jhetchan]

the "almost right" is almost never right.

because completely wrong is easier to reject.

Almost right is sticky because it has enough truth to feel legitimate, but also enough distortion to mislead the next step.

it feels close enough that the mind stops searching. That is why it traps.

the work is to keep asking…
what is the missing variable?
what does this phrase hide?
where is the compression loss?
what are the edges?

because the last 5% of precision often changes the entire path.

[Jhet Chan @jhetchan]

magic of foundational research is half exhilaration, half vertigo. one moment you’re terrified to discover “oh, they did it already” the next moment you realize nobody has done quite this.

[Jhet Chan @jhetchan]

A sword master is not simple because he lacks complexity. A sword master is simple because the complexity has been trained into reflex, posture, timing, breath, perception.

[Jhet Chan @jhetchan]

and see ourselves clearly so we can see the world clearly.

not good not bad.
just for what it is.

world-self correspondence

[Jhet Chan @jhetchan]

how we see the world is how we see ourselves

[Jhet Chan @jhetchan]

introducing a new lens always produces contributions that look modest under the old rubric

because the contribution is the new rubric itself, not the magnitude of the first measurements made through it.

[Jhet Chan @jhetchan]

a tale of illusion.

in physics, mathematics, and cognitive science, we don't usually call it an illusion, we call it an abstraction, an effective theory, or a representation.

because the cost of total truth is infinite.
to model trajectory of a thrown baseball "perfectly", we'd need to simulate every particle and their quantum interactions.

there is not enough computing power in the universe to do that. so we use the 'illusion' of classical mechanics pretending the baseball is a single object with a center of mass to give us clean trajectory.

neural systems do the same.
reality is a continuous, high-dimensional chaotic mess.
so the system compresses it onto a low-dimensional manifold and forces stable representation.

even computers, rely on illusion with perfect 1s and 0s with noisy continuous voltages underneath

without these collapses there would be
No categories
No logic
No beauty
just an infinite stream of white noise.

a bad illusion is one that fails to predict reality.
a good illusion is not truth but a structure that preserves what matters

any system that acts must hallucinate a simpler world and then live inside it.

it is the only way an observer can zoom out far enough to actually interact with the universe without being paralyzed by its infinite complexity.

[Jhet Chan @jhetchan]

instead of …
“Is what we are seeing the reality? ”

try the alternative:
“what invariants of the world survive this projection pipeline? ”

[Jhet Chan @jhetchan]

Don’t try to learn the whole terrain. Learn how to move safely on the parts you’ve seen and recognize when you’re at somewhere new.

[Jhet Chan @jhetchan]

take one weird thing and make it undeniable.

[Jhet Chan @jhetchan]

it's either stabilises into something real or it doesn't

it's either anchors to a reality where something observable, testable, repeatable, and holds when pressure is applied… Or it's vapor.

No poetic license.
No "it depends on perspective"
No "my truth vs your truth"

if the claim, model, belief, or system cannot be stress-tested against what actually exists (independent of wishes, narratives, or social consensus), then it fails the test.

it doesn't stabilize anything real then it's merely a decorative noise that collapses under its own weight the moment reality pushes back.

not everything has to survive the test. but it has to survive where it matters

[Jhet Chan @jhetchan]

we thought we were pursuing something
but in parallel, in the process, we were being shaped into someone who optimizes that system

[Jhet Chan @jhetchan]

progress is mostly about eliminating illusions

[Jhet Chan @jhetchan]

if failure is specific rather than random, then it means the failure has a structure. And structure is exploitable.

[Jhet Chan @jhetchan]

some truths can't be reached from where we currently stand, no matter how valid they are. The path has to exist, not just the destination.

[Jhet Chan @jhetchan]

2 dispatch bugs found by dogfooding exoprotocol on a real project:

- priority was inverted (5 dispatched before 1)
- epics got dispatched as work items

the agent figured it out by reading its own reasoning trace. fixed in v0.1.27.
https://t.co/Tn3cRXfhb3

[Jhet Chan @jhetchan]

agentic coding has made code generation trivial. the catch:
if a requirement can't produce executable proof, then it's a liability, a slop

ideally:
spec → .md rules → agent codes → test → ci gates
misaligned code = build fails.

here's a take to turn spec into infra, not to slow things down but to keep things aligned.
https://t.co/uv3B2T5z4v