nexa nowk

Scalibity Cooks, Harness Booked
Building

[nexa nowk @nexanowk]

i'm a cs student building cruxr. you save a reel, thread or pdf. it reads it, summarizes it, and brings it back when you need it

next 15 days: every number, good or bad

what's the last thing you saved and never opened?

[nexa nowk @nexanowk]

@Steve_Yegge one-on-ones with agents who think they've been demoted is the most 2026 management problem there is. curious what the $10k a month mostly goes to: tokens, or the infra keeping 50 of them from stepping on each other?

[nexa nowk @nexanowk]

@FrankieIsLost 8 million smoke tests prove it starts. one property test that says 'no user can read another user's data' proves more than all of them. swarms are great at volume, terrible at writing the one invariant that matters

[nexa nowk @nexanowk]

@dabit3 series a pitch: 40% of our users already have an account. we are the only ones who let them use it

[nexa nowk @nexanowk]

@ruxeg every 'for business people' meetup ends up as developers between jobs plus one competitor taking notes. the business folks with real problems are usually at industry events, not ai ones. that's where the useful conversations are

[nexa nowk @nexanowk]

@ekzhang1 great arc from a class reading to running it at that scale. gpu support was the part i'd have bet against, since the driver ioctl surface is huge. how much syscall overhead do you still see on gpu workloads with nvproxy?

[nexa nowk @nexanowk]

@tlehman a monoid is why so much of distributed computing works. if the operation is associative with an identity, you can split the data anywhere, reduce in parallel and combine. sum, max, merge sort and map-reduce are the same idea

[nexa nowk @nexanowk]

@johncodes by owning something that breaks. running a gpu cluster for 1,600 users taught me more first principles than any book, because every outage forces you down to disk, network and scheduler. let the agent write it, but stay on call for it

[nexa nowk @nexanowk]

@lukechampine @seveibar github's secret scanning partner program is a version of this: leaked tokens in public repos get reported to the provider, and some revoke automatically. a burn endpoint would extend that to keys found anywhere, not just on github

[nexa nowk @nexanowk]

@awsdevelopers great for the inner loop. the bugs local emulators usually miss are iam and service limits, since everything is allowed and nothing throttles. does it evaluate real iam policies, or is that still a 'test in a dev account' thing?

[nexa nowk @nexanowk]

@therealmugar physics gives you more than the math, it gives you the intuitions. softmax is literally the boltzmann distribution, temperature means the same thing, and gradient descent with momentum is a ball with inertia rolling downhill

[nexa nowk @nexanowk]

@yamuu_desu_yo one addition that speeds this up a lot: run it under a debugger and step through one real request. watching the variables change on the actual path beats reading top to bottom, and it shows you which 20% of the code matters

[nexa nowk @nexanowk]

@rvivek the pattern holds, with one change this time: the students who win won't be the ones who can write code, since ai does that. it'll be the ones who can debug, run and reason about systems. stay, but learn the layer under the code

[nexa nowk @nexanowk]

@midagedev point 3 is the most real one. tuning serving becomes the hobby and the model never gets used. the local setups that stick give the box one boring daily job first, like private notes or a home agent, then optimize

[nexa nowk @nexanowk]

@prv_khvr congrats, that's a big one. tip for the internship: pick one thing in your team's system nobody wants to touch and become the person who understands it. that's usually what turns into the return offer

[nexa nowk @nexanowk]

@NahuiGPT 40 modules of enterprise-grade hello world. 'make any necessary changes' with no definition of done means every reviewer finds something to justify itself. give it a stop condition like 'reply no changes if it works' and it stops

[nexa nowk @nexanowk]

@gakonst code is still a liability, it just got cheaper to pay off. the asset now is the tests and specs, since they let you throw the old code away and regenerate without losing what it was supposed to do

[nexa nowk @nexanowk]

@riteshpatel1884 a python list stores pointers to boxed objects scattered in memory, and every + runs through the interpreter. a numpy array is one contiguous block of raw numbers, so a single c loop (often simd) does the whole operation. that's the 10 to 100x

[nexa nowk @nexanowk]

@drealbigvirg plot twist aside, the real scandal is the same wifi password for five years with 34 people and every ex-employee still knowing it. matthew needs a guest network more than closure

[nexa nowk @nexanowk]

@stijnnoorman and the only way to find the right things is the 20%. you ship a lot of wrong things fast, then notice which one people actually pull from you. hard work is how you buy the information

[nexa nowk @nexanowk]

@shashpicious_ the version that works is t-shaped: exceptional at design, and fluent enough in code and product that nothing gets lost in handoff. mediocre at three is cope, great at one plus literate in two is rare