It would be interesting to hear more about how you’re actually using them. Do you have sophisticated feedback loops around the models so they can verify their work and converge on good solutions? And how do you decide what to give the 5090 vs the Spark vs a frontier model?
Correctness matters much more than speed to me, but if I can get both, that’s obviously very interesting.
Cause: Legacy Salesforce login service got into a resource-exhaustion cascade.
Fix: Rolling some unspecified fix they proved in testing out over the fleet seemingly very slowly (After their earlier attempts to roll something out faster failed).
Coincidentally, Salesforce just laid off the senior engineering lead on their shared login service team who 2 months ago was describing the practices his team follows to keep 99.95% uptime on that shared login service (this guy: https://www.linkedin.com/posts/patricktprice_engineeringmana...)
Layoffs are usually short-term money-savers at long-term cost, but hot damn this is the shortest short term I've ever seen.
And while I haven’t actually done my homework, the push to email-based login seems like it’s really pushing people to start using their own custom domains for logging in
If you can get your agent of choice to run up the app in an emulator and navigate the whole thing taking screenshots then tasks like edge-to-edge can be pretty hands off.
The way I did it was to firstly get accessibility identifiers on everything. Then have it record everything I do on an emulator to navigate the entire app. Then left it for a while to work out how to actually use the app via the emulator. Once you have this sort of baseline. A lot of changes are really quite pleasant. I often ask for an HTML gallery of screenshots covering a few device sizes as the step before I bother running it on a device myself.
We use Roborazzi[0], which is an excellent screenshot library, but:
* We target minSdk 24
* We're still using XML views
* We target a large variety of form factors
* We have a lot of screens, and these have a lot of configuration options.
I really wish it was "point an agent at it, and walk away", but many screens ended up being hours of iteration with Fable driving my phone/an emulator to produce an acceptable outcome.
Given that my Pixel 9 Pro has system screens which are still broken under edge to edge, it's not an easy change, and takes up time which could have been going towards feature development.
A tangent, but a big thank you (to you, and all other developers) for supporting older API levels from someone with devices ranging from Android 8.1 (BlackBerry Key2LE, at minSdk 27) to Android 16 (BlackBerry "Q25", at minSdk 36.1) <3
E-waste is awful, as is cutting off old devices. We still backport security fixes to minSdk 21.
To Google's credit, post-21 is a much more pleasant experience and requires very little developer time. API desugaring[0] has worked perfectly, their compatibility classes[1] handle most of the common cases, and there's often very little to actually code[2].
Some reviews are clearly automated, some get a light human review, and some involve a human really looking hard. New apps seem to get a more rigorous human review for their first 2-4 submissions. Or at least humans looking over automated test screenshots etc. One reviewer apparently only looked at screenshots because they missed something that would be blindingly obvious on the prior screen if they did it themselves.
The explosion in LLM app development has clearly created a bottleneck at the human review steps. Not only delays but much more “dumb” rejections from likely over-stressed humans.
FWIW, I find if I’m pushing the disk limit, Spotlight search becomes unusable. So I free up ~100GB (easy to do with iOS/Android dev artifacts, emulators, etc.) and then force a Spotlight re-index from the command line:
sudo mdutil -E /
You can check indexing status with:
mdutil -s /
After that it’s usually fine again until I start pushing the disk limit.
Every sibling comment's issues I've experienced. But a reindex after freeing enough disk space seems to be a semi-stable solution for me. If only I could write a tool to remind me to free disk space before it screws up spotlight. It would be nice if MacOS put the index on its own partition.
Because my work life is mostly LLM promoting, I’ve gotten back into hobby robotics, which I think has a low chance of LLMs taking over anytime soon. Designing and building a robot that can build a wooden hardwood deck using the edge-screw method.
The comments here are taking a very uncharitable definition of “power”. If I take that definition I can see how people are being so negative. But just a slightly more positive definition and this all seems like reasonable dialogue to have with an early stage startup.
reply