Also, very much with you in terms of exposing those libs ( like PDFium and FFmpeg ) to consumers, it lowers the shipping burden by quite a bit and just enables so much.
A note to anyone going down this road, just as an aside: actually shipping w/ Chromium ( and especially w/ FFmpeg ) can get complicated. If you're shipping on Mac OS, you'll want to worry about notarizing all your helpers along with your main process ( I'm assuming Zebkit still running w/ helper processes ). Similar w/ FFmpeg if you're linking to the FFmpeg libs. And, w/ FFmpeg, just validating that your license compliance is in fact what you want to be shipping w/ - FFmpeg has GPL and LGPL modes, so making sure you have selected the right one is important. If I recall Blink is LGPL as well, so making sure you meet what's required there has some distribution effects that are important.
As an aside: decoded.legal looks like an amazing resource for folks based in UK / EU that need legal w software expertise. I haven’t easily found similar folks in the US: does anyone have good recommendations for this?
Found similar results, but the economics of the models involved have changed so quickly that it becomes so task specific on what the right thing to do was.
With an assemblage of models like Jev + Opus or Jev + Qwen, Jev would often do the right thing when deciding to escalate to another model - I was impressed that it would usually make the right call, especially if you gave it good awareness of what the other model capabilities were. "ESCALATE to opus if you need good vision awareness".
When Haiku 5.5 came out though, for my tasks, the cost differences disappeared almost immediately. Haiku was good enough / fast enough for so many things that introducing a decider layer didn't offer huge wins. Jev's "millisecond" level performance was nice but would get dwarfed by other latencies in my chain ( web browsing ), that it didn't make a measurable difference in my tasks.
This made me think... Is anyone using small models to generate candidates and let a more expensive model pick one? I.e. genetic programming on neural networks? Perhaps a decision model could be used for the picking?
Agree with this. I started down learning more about archival while I was adding it to my own thing [1]. I wish there was more of an existing toolset for diffing archivals as well - what changed is just as interesting as what is there. Equally, something I kind of think is lacking in WARC/WACZ is exploration of the archives - maybe forensics is too strong a word but it would be nice to have an easier way to explore what you’ve captured and understand the data in different dimensions.
I’ve supported some of this in my own tool, ZShot - capture in sqlite, Quicklook extensions for WARC and WACZ. Just trying to make archiving easy and interpretable.
It’s also been interesting to see people use archival but for different uses: Kiwix [2] showed up in my reading recently, “archival but for offline internet”, which I’m kind of fascinated by.
Working on ZShot [1]. I'm about a year in on a local-only browser tuned specifically for CLI and API usage - kinda filling the gap for "geez, I sure wish I had a browser, but not one to look at". You need a webpage as PNG, or an MP4, or a website as a WACZ or SQLite database. There's a Mac app to help with discoverability of options, but I wanted for my local tooling and building services on top of it.
Along the way I ended up adding an internal agent framework to satisfy stuff like "log in, click the button, and then take a screenshot", out of which fell stuff like automatically generating Playwright tests from that. It even runs the Playwright test suite and Puppeteer test suites ( a bit faster, and for a lot less memory ). I kinda love this bit because I can layer in multiple models that will bounce control of the web session between each other, to do stuff like let a Typesafe model choose to escalate to Opus when it needs higher minded behavior. It's fun to watch it happen.
I've mentioned it a few times around here, I'll probably do a Show HN soon when I feel like it's ready for that. Hopefully soon.
Sensors everywhere. If it's at all relevant to what you're working on, providing a different way of viewing something, and a skill description on how to do that, has these long-compounding effects that are magic at first, and then become "of course, it should always just work" as your expectation.
My product that I'm working on sorta acquired a bunch of that. I put a ton of work into making a browser that could just see / interact w/ the web in a ton of different contexts. Now that I have it, it's pretty amazing to just have the agent do everything I need it to ( pull a network log, debug JS, navigate a form ), and it has the appropriate context everywhere.
I'm working on documentation for zshot[1][2]. Kinda like "what if browser automation had curl-like ergonomics, but also was extremely agent friendly, but also could turn into your headless browser infrastructure". I've been working on it for a while now and have incurred a hell of a debt of "explain what it is and how to do it", so working that debt off bit by bit.
If you're up for trying something a bit fresh, I've been working on something that helps in this space. Check out: https://zshot-app.com/for/agents/
You'll have to sign up for a trial to do this, but you can just have your coding agent drive a session and just ask it to install the zshot skill, then drive a session and record a puppeteer script for you. If you try it out, love to hear what you think.
reply