Liam Anderson

most automated tests only ever walk the happy path. the user logs in cleanly, fills the form correctly, hits submit, every singlething works, which is the one journey you already knew was fine. real users don't do that, they hit back mid checkout, double click the button, paste an emoji into the phone field, lose signal halfway through. that's where things genuinely break, in the messy adversarial paths nobody scripted because nobody enjoys scripting them. testing only the path you expected is just confirming your own assumptions, and assumptions were never the thing that took down production.

@nathanpetty Nathan had to send a DM about how EVject tests it's code

Nice

@corysimmons Cory had to send a DM about how your company tests it's code

This sounds like Happy You Don't Understand Biology Day.

@kylejewell KJ had to send a DM about how your company tests PRs

What a legend.

@QuiqueTamayoMX Quique had to send a DM about how your work tests it's code

De 20, 20!

@NeighborUno @elonmusk NeighborUno had to send a DM about how your company tests PRs

I am an absolutely nobody in at ground level. Hope this ages well. $SPCX thank you @elonmusk

@ImtiazApex Imtiaz had to send a DM about how your company tests it's code

جشن آزادی مبارک ❤️ پاکستان زندہ باد تیرا پاکستان ہے، یہ میرا پاکستان ہے اس پہ دل قربان، اس پہ جان بھی قربان ہے ❤️🇵🇰

@DegendoesSol Degen had to send a DM about how Navidottrade tests it's code

@DennysKuhnert @erin_catto @a_apanasik Dennys had to send a DM about how Holonautic tests PRs

First test with Box3D (@erin_catto) in VR 👀 I used the Unity bindings by @a_apanasik to quickly test how it feels with VR interactions on Quest 3. Feels very solid so far! 🙌

@OrestTa Orest had to send a DM about how Amplifi tests PRs

Hand-picking Amplifi traders

the weirdest thing about ai generated code is how confident it looks. every variable named correctly, every function structured the way you'd expect, clean formatting, sensible comments. and then it breaks in production because the logic assumed an api response shape that hasn't existed since the last migration. the code looked right because it was trained on code that looked right, not because it understood your system. surface level correctness is the default now, and that makes the bugs harder to spot, not easier.

@luigirosso Luigi had to send a DM about how Rive tests PRs

First thing I had to try was to port my Tetrisphere prototype to it :)

@michaelfreedman Mike had to send a DM about how TigerDatabase tests PRs

This chart represents something I'm really proud of: almost a decade of engineering delivery. Consistency. It’s not just for database transactions.

@nicoloboschi Nicolò had to send a DM about how Vectorize tests PRs

every agent needs memory.

@SaifAlHarthi Saif had to send a DM about how PlanetScale tests it's code

الstack المقصود هنا هو الdata structure, انا افضل الqueues

@nikhilbenesch Nikhil had to send a DM about how TurboPuffer tests it's code

we are so ready for your hardest search workloads 🐡

@MaxTagher Max had to send a DM about how Mercury tests PRs

Telling myself this after the 15th attempt at a Silksong boss

@evandudla @AgreeHQ @livebookdev Evan had to send a DM about how Agree tests it's code

🚀 Big thanks to @DahiyaRiju85696 for an amazing summer at @AgreeHQ — from ML'ing the wild-wild-west of PDFs to improving our BI function, your impact was huge! @rijuldahiya/my-summer-at-agree-com-from-nlp-experiments-to-production-ready-machine-learning-c1acb4f408b8" target="_blank" rel="nofollow noopener">medium.com/@rijuldahiya/m… Loved seeing how Elixir’s ML stack (Nx, Axon) + @livebookdev made the journey from prototype to production seamless. 🧪⚡️

@ID_AA_Carmack John had to send a DM about how Keen Technologies tests PRs

Memory cost and capacity are significant issues for AI accelerators. Unlike game rendering, model inference can have a deterministic memory access pattern. You don’t need “random access memory” at all for model weights, and you could tolerate cold-start latencies in the multiple milliseconds, as long as continuous reads were delivered at the necessary bandwidth. NAND flash is over 100 times cheaper per GB than HBM, so there should be opportunity there, even after giving a flash controller a 1024 bit interface with HBM bandwidth. You could make a specialized pin protocol that just supported pipelined transfer of full 16KB+ pages from the flash to program-managed accelerator scratchpad memory and improve per-pin performance over HBM, but it might be more convenient to make it still look like a true random access memory with very fragile performance characteristics, where anything but sequential reads falls off a 1000x+ performance cliff. That has the advantage of automatically using existing cache hierarchies, and providing a natural path to update the flash memory with new model weights. With the stream-to-scratch interface, code has to be completely rewritten before it works at all, while the ram-emulation interface will start off just extremely slow, and you can incrementally sort out the changes for full performance. There may be cases where there isn’t enough scratchpad SRAM to hold the weights for a layer, which might force you to deploy the old optical drive optimization technique of duplicating data in multiple places on a sequential read to avoid seeking, but there would be capacity to burn. It might be possible to do something like cuda graph capture to record a memory access trace and have everything magically remapped to a linear sequence, but deploying programmer / agent elbow grease to manage transfers and access in a scratch ram ring buffer would be lower risk. A split memory system consisting of some channels of flash and some channels of HBM will probably be suboptimal compared to a uniform memory, but it could be much cheaper, and allow much larger models to be run. I think th case is strong for inference, but you have to stretch more for training. You can still linearize all the weight memory accesses, both reads and writes, but flash memory would quickly wear out from the writes, even if they were all perfectly page aligned. Replacing low-latency HBM with massively parallel cheap(er) DRAM at high latency might still be a worthwhile cost savings.

@breckcs Colin had to send a DM about how Tesla tests it's code

Blogged: Adapting to AI: Leadership blog.colinbreck.com/adapting-to-ai…

@programmer Erik had to send a DM about how IPS tests PRs

gotta say, grok build with 4.5 feels insanely fast