Pinned Tweet
Good question @VishalMalvi_ ! Here's what's going on behind the scenes of the @gmail iOS app š
...I knew this was familiar š
xitter.cf/emergetools/status/198ā¦
Emerge Tools retweeted
At @Sentryās AI offsite this week, @indragie & I decided to kick things off with a bangāļø
Weād seen a video where friends bought each other ridiculous shirts before a flight and had to wear whatever they were given... so naturally, we did the same for the offsite (ft. @zeeg)
If you see us around SF today, make sure to say hi
Other notes:
- Flutter app
- 207 MB of dynamic frameworks (worse for size compared to static)
- lot of external dependencies
The real TL;DR
xitter.cf/emergetools/status/206ā¦
There's no excuse anymore, Sentry Size Analysis is free 100 builds / month
That's a nightly + weekly releases on Android & iOS and then some
docs.sentry.io/product/size-ā¦
We focused on the video file b/c it's so easy to optimize + having a 23 MB onboarding video, seen once by user, is ... not ideal
But all the images are also easily optimized, saving 30+ MB.
Optimizing the video & images + stripping binary symbols would save ~75 MB, a third of the total app size
Here's the metadata comparison of the new file + a side-by-side of a random frame, which is virtually indistinguishable
Note the lower bitrate (15.45 Mbps vs. 3.42 Mbps) and file size (23.2 MB vs. 5.1 MB)
TL;DR on the video file, it's encoded with way higher bitrate than needed
If you want to learn more about bitrate, check out this thread
xitter.cf/emergetools/status/198ā¦
Itās Friday, we do have a job, and we do have things to do
But weāre going to tell you about bitrate today
Why this @weedmaps video on the left is 14x bigger than the one on the right, despite looking the same šæ
It's a 12 second video of some wings flying around
What's egregious is that 1 line of ffmpeg reduces this by ~80% (and you could do even more)
There's no excuse anymore, Sentry Size Analysis is free 100 builds / month
That's a nightly + weekly releases on Android & iOS and then some
docs.sentry.io/product/size-ā¦
Some other notes:
Opus generally did a great job on all site recreations, regardless of scenario, but did see improvements w/ screenshot access
Here's an example of Opus recreating linear, 1st image is no screenshot, 2nd image has screenshot access
Might put some more rigor into this - would be cool to see how different models improve (or not) across the scenarios. For convenience this was only done on Anthropic models
Influenced by the release of Snapshots on Sentry
We're bullish on snapshot testing to both guard agent changes + as an artifact that can be used by an agent, excited to see more development here š«”
docs.sentry.io/product/snapsā¦
Take this Netflix example. Here are 5 sonnet runs of āAgent 1ā (left is actual site, right is agent output)
In every case, the agent failed to get the background right
First step: a scraper to take screenshots of a websites
We then had 3 agents try to recreate the reference image
Agent 1: just told to recreate as closely as possible
Agent 2: can screenshot its work and reference it
Agent 3: can screenshot + run a pixel diff
