scosman

20 ms·

361.
by scosman 2y ago
Some people run Deepseek on CPU. 37B active params - it isn't fast but it's passible.
362.
by scosman 2y ago
They have usage data and I'm sure they did the math. A few users will max this, but most won't touch it or only do a few requests a month. If it bumps their conversion rate it should be worth it. I hope!
363.
by scosman 2y ago
Okay - this is amazing. I’m a happy paying Kagi user. Getting to apply almost all of my payment towards for cross-platform tokens as a bonus - wow. This will replace a chatGPT and Anthropic sub for most everyday users. Their assistant is be
364.
by scosman 2y ago
> These models are our best yet thanks to distillation from Llama 4 Behemoth, a 288 billion active parameter model with 16 experts that is our most powerful yet and among the world’s smartest LLMs. Llama 4 Behemoth outperforms GPT-4.5, C
365.
by scosman 2y ago
Oh for perf reasons you’ll want it all in vram or unified memory. This isn’t a great local model for 99% of people. I’m more interested in playing around with quality given the fairly unique “breadth” play. And servers running this should b
366.
by scosman 2y ago
At 109b params you’ll need a ton of memory. We’ll have to wait for evals of the quants to know how much.
367.
by scosman 2y ago
128 exports at 17B active parameters. This is going to be fun to play with!
368.
by scosman 2y ago
Yes. But is that the line? Crawling makes sense (automated traffic) but this isn’t automated, it’s user initiated. Search indexing makes sense (this isn’t that). Training makes sense (this isn’t that). It should have a honest user agent so
369.
by scosman 2y ago
Do you think that should be an option? I totally get opting out of crawlers, search or training but this is different. But should the author be able to opt out of a tool used for manually initiated queries? I can’t say “don’t use grep” on m
370.
by scosman 2y ago
It’s one of my favourite MCP use cases. I have cloned projects and used the file browser MCP for this, but this looks great. It allows you to ask questions about how an entire system works. For example the other day “this GitHub action requ
371.
by scosman 2y ago
How does it compare to Apple Mail? That’s my reference local email client.
372.
Show HN: Kiln – GitHub+Jira+Figma for AI product development
2 points by scosman 2y ago|0 comments
373.
Using Computer Fans for Home Air Movement(scosman.net)
4 points by scosman 2y ago|0 comments
374.
by scosman 2y ago
Agreed if you are building a typical service. The abstractions will slow you down and don’t and anything. The use case where they are helpful is “bring your own keys” apps. I maintain https://github.com/kiln-ai/kiln wh
375.
by scosman 2y ago
I tried this years ago. Some advice: target your marketing and product at “beautiful+easy portfolio” to start. See if you can make something beautiful people want to share, and when other people see one they want to create their own. You ca
376.
A prompt engineer's guide to fine-tuning(getkiln.ai)
5 points by scosman 2y ago|0 comments
377.
by scosman 2y ago
If you expand into tools for developing AI models, check out Kiln: https://github.com/Kiln-AI/Kiln It includes synthetic data generation, fine-tuning and evals to help build your own models.
378.
by scosman 2y ago
Did Karelia Software pay the writer of `find` to use the concept?
379.
by scosman 2y ago
I've never understood this Apple criticism (scherlocking). Someone built a search for your files, so it's not right for Apple to build a pretty key feature into the OS? There's a lot of fair criticisms of Apple, but they don&
380.
by scosman 2y ago
To answer my own question: "To use YouTube.js in the browser, you must proxy requests through your own server"
381.
by scosman 2y ago
Potentially dumb question: if YouTube.js works in browser - can/has someone made a YouTube player that’s just a static page? Is there a need for a backend?
382.
by scosman 2y ago
What do folks consider helpful in .curorrules? So far I've found: - Specify the target language/library versions, so it doesn't use features that aren't backwards compatible - Tell it important dependencies and versions
383.
by scosman 2y ago
+1. But cursor now support sonnet 3.7 thinking, so maybe the team adopted that so thinking is separate from response?
384.
by scosman 2y ago
Do that. Also: set up a widget on your home screen that shows a bunch of random old photos. You won’t “find them”, you want them to find you. There will be misses but the hits will be worth it. Apple’s is decent (disclaimer: I made it). I l
385.
by scosman 2y ago
> Because our phones are pulling from a huge, contextless pool There is so much context in your camera roll. So so much. Not all of it is understood yet, but it’s finally getting good. > The physical archives we amass in drawers or ol
386.
by scosman 2y ago
Exactly. Benchmarks != evals. Task specific evals are how you determine if an AI system works. Expanding your evals as you test, find new issues, and change techniques is the only way to ensure each change doesn’t regress a bunch of your pa
387.
Show HN: Build your own AI evals in a minutes – benchmark to human prefs(docs.getkiln.ai)
1 points by scosman 2y ago|0 comments
388.
Ask HN: What does "Used by engineers at [Company]" for OSS mean? Stars/emails?
2 points by scosman 2y ago|0 comments
389.
by scosman 2y ago
Can someone explain how it works under the hood. A Rust server for python obviously needs to still run the python interpreter. Is it running many instances to get around the GIL, and doing the request management / reverse proxy /
390.
by scosman 2y ago
If we are lucky, in 30 years we will have reinvented the stock exchanges of the 1950s!

More ›