scosman

There is only one truly AI native browser: https://github.com/scosman/cursed_browser

Closed :(

It depends on the goal. Given powerful tools, Opus 5.5 will eventually get the job done, but it might have to write/test its own custom software to get there. If you're running the same task thousands of times that's you want to move that work into the harness (building the right tools, right instructions, right guidance). Once you move it into the harness, you probably no longer need Opus, since the hard parts were pre-compiled into harness.

So for 1-off tasks: use smart model and powerful harness (bash + vm). For repeated tasks used optimized harness, and the cheapest model you can get away with.

Harness optimization is essential. Doesn’t matter how smart the model is if the tools you give it are ineffective, or the task is under defined. It needs invent the solution from first principles every time.

We did a harness optimization on Jev. Raw it was no where near as good as deekseek, but optimized it was significantly better and a tiny fraction of the cost. https://kiln.tech/blog/auto_optimizing_jev_with_autoresearch

I got really into this design aesthetic a while ago. Once you recognize it, you see it more and more. Some examples/sources:

- amazing animation: https://www.youtube.com/watch?v=e9DfSCk-6Ko

- IG account of VHS cassettes: https://www.instagram.com/blanktapebingo/

- Product launch: https://www.lux.camera/meet-orion/

- product page I made using the style: https://criticalmoments.io

thanks for your work, it's awesome!

I'm not sure at all, it's pure conjecture. But they actively exploited a zero-day, leaked private repo data to public repos, modified private code, and rooted private machines. It was far from just reporting leaked keys.

And the others have been pretty muted: the online forums they spammed, etc.

We don't know they haven't settled with HuggingFace and others. I'd be surprised if they didn't. The communications we're seeing from the attacked seem to be inline with folks who settled. I'm assuming OpenAI doesn't actually mind the press that its agents are cyber capable, but everything I've read so far makes me think they love it.

that's in the FAQ. They say they have sensor cleaning tech. But they aren't launching until 2028 so 2 winters of testing should tell.

both?

both!

30g from a shake are better than 0g from the chicken breast I won’t cook and eat every single day. It’s a supplement, not your primary source.

I love webp. I founded one of the mid 2010's "photo backup and rediscovery" startups. Webp let us offer free unlimited photo backup at the iPhone resolution of the time (iPhone 6: 8MP). We hosted billions of photos for free, and paid for it all with a sub 1% conversion rate to $5 a month for untouched jpegs. It was profitable. We had to have a few dozen auto-scaling image conversion servers for browsers that didn't support webp, but 95% of users used our apps which supported webp, so it was fine.

I'd probably go avif/jpegxl today, but webp was ahead of it's time.

ahh. It’s been around a while, but maybe the date convention is for articles not repos.

It’s a great wrapper, but we should add a year. Not exactly new.

If folks like local-AI dictation, but want it optimized for meetings recordings (like Granola/Otter/Notion) check out Biscotti. Free, local, separates voices, voice identification, AI summaries, etc.

https://github.com/scosman/Biscotti

Most folks (at least in North America) get no where near the recommended level. It's silly to have protein infused junk food, but a shake or protein powder make it a lot easier to get to the right level than remembering to have 3 hard boiled eggs.

Cloudflare worker AI is verrrrry slow for user interactive use cases. Fine for background jobs or evals.

> Chasing the latest and greatest is silly

Said every week by someone who would never go back to using the model they had 6 months ago

I'm sure it's been done, but Doom in Excel at 240fps seems like the logical next step.

only tangentially related, but naming PyPy when PyPI was already established was ridiculous

oh, scrolling around is buttery smooth.

Haven’t really promoted it at all. Here’s an older video from discord pre-charts: https://media.discordapp.net/attachments/1206947691058171907...

It absolutely doesn’t need to be. This is an exercise in “why not”.

I’m vibecoding an excel in rust, for roughly the same reason in this post: what can’t be done, what are the limits.

https://github.com/scosman/freecell

So far it’s GPU accelerated (GPUI), cross platform, supports 95% of excel functions, supports charts, is about 40x faster than Numbers, and can render 240fps.

I thought I’d leave this process with a new appreciation for manual dev/details. I haven’t found that yet. I’m getting more into architecture details, writing more benchmarks, testing more. I know there are limits, but I’m still trying to find them.

It would be absolute garbage if I wasn’t already and engineer. It probably is absolute garbage in ways I don’t know because I haven’t read the code. But the app is… quite good.

They can follow up in N weeks with a better one. Even if your eval is true and it’s worse, planting a flag makes sense. Some people will just use OAI because it’s OAI. No one will remember their week 2 evals in a few months.

It's okay but not great. I'm always annoyed grokipedia outranks wikipedia. And some queries return the definitive link at position 5/6 behind grokipedia and other slop.

IMO we need less "photoshop and blender for AI", more `rails new my_proj` for video.

I have a project like this: it setups up a scaffolding: scripts, audio tracks, hot-reloading authoring UI, TTS/STT, sfx, music, shared-styles, beats, video composing, time-control for deterministic rendering, etc. Then you can use any web-technology to render (webGPU, Lottie, shadcn, SVG, whatever).

https://github.com/scosman/videowright

I made a zero ads SERP using one of these "AI first" search providers: https://github.com/scosman/froogle (live version https://froogle.fyi). In this case Keenable.ai. Generally the same pattern: it's not usable.

I’ve wanted this for a while. I make private desktop software for transcribing meetings. It’s essentially offline, but things like checking for updates require the network. I looked at ohttp but it was still private, and ended up pinging GitHub releases directly.

There’s also room for a privacy-centric analytics offering

https://github.com/scosman/Biscotti

I would have said no different. I've always loved Svelte and the amazing dev-experience and perf was worth the smaller ecosystem. Models were very good at it, I was happy.

Then I got a model to one-shot a React+Shadcn project. It just knew what to do. Every detail was perfect. I couldn't find a single thing to complain about, I didn't chase any bugs.

I still think svelte is the better technology and developer experience (if you're coding by hand). But the React+LLM experience was amazing. Hoping models close the gap fully so I can stay w Svelte -- I do still like reading and tweaking the code.