scosman

19 ms·

331.
by scosman 1y ago
Lighthouse also warns against generic link text, like “Learn More”. It’s one of the few lighthouse warnings I just ignore.
332.
by scosman 1y ago
Ah, great background. One more tidbit I just learned: there was a Netscape/Sun deal around the name, so the registered trademark has some legal history. It's not that Sun (and then Oracle) just claimed rights to something Netscape
333.
by scosman 1y ago
Java is a Sun product, but Java has nothing to do with Javascript except a confusing name overlap. Javascript was written at Netscape.
334.
Many small evals beat one big eval, for AI Products(getkiln.ai)
2 points by scosman 1y ago|0 comments
335.
by scosman 1y ago
Any good YouTube options (including self host)? I’ve tried a few and they always seem to be down more than up.
336.
by scosman 1y ago
Works on MacOS as well. Not just iOS.
337.
by scosman 1y ago
oh happy day
338.
by scosman 1y ago
I’d like to use it anonymously too but it doesn’t feel reasonable to ask for access to a $80k piece of hardware anonymously. I needed to give ID to rent a movie at blockbuster.
339.
by scosman 1y ago
Slower tokens, maintenance time (installing new models/updating tools), lower SWE bench scores. For me it’s not worth it - optimize for speed. If I worked on a codebase I couldn’t send to the hosted tools I’d go this route in a heartbe
340.
by scosman 1y ago
Lot's of tools for each of those separately (RAG and fine-tuning). We're working on combining them but it's not ready yet. You don't need a big GPU cluster. Fine-tuning is quite accessible via both APIs and local tools.
341.
by scosman 1y ago
Context window + prompt caching if you can't use RAG. Can add a lot to long context models, and their needle in haystack metrics keep getting better. Why can't you use RAG?
342.
by scosman 1y ago
Yup! I'll have to write some of these up. I can probably do open datasets and evals too. If you have use cases you'd like to see let me know! Some quick examples (task specific performance): - fine-tuning improved performance of L
343.
by scosman 1y ago
Got it. Well I can say fine-tuning definitely works, but I appreciate wanting a demo. We'll work on something compelling. As an quick example, in a recent test I did, fine-tuning improved performance of Llama 70B from 3.62/5 to (w
344.
by scosman 1y ago
We don't sell fine-tuning tools - we're an open tool for finding the best way of running your AI workload. We support evaluating/comparing a variety of methods: prompting, prompt generators (few shot, repairs), various models
345.
When Fine-Tuning Makes Sense: A Developer's Guide(getkiln.ai)
157 points by scosman 1y ago|62 comments
346.
by scosman 1y ago
I had this happen on a consumer startup with referrals. Every month like clockwork one person would fake referrals to get a free moth, which involved jumping through non trivial hoops (re-installing all, creating content in the fake account
347.
by scosman 1y ago
0.2s on ty compared to 4.7s on pyright. Not even close.
348.
by scosman 1y ago
Found 275 diagnostics 0.56s user 0.14s system 302% cpu 0.231 total Damn that's fast. And zero issues in this project with pyright so curious that these are...
349.
by scosman 1y ago
What’s the thinking of spreadsheet first? Just making it super accessible for people who already have data? I’m building a UI for fine tuning (and evals, and synthetic data gen) - https://github.com/Kiln-AI/Kiln - and
350.
by scosman 1y ago
Anyone manage to sign up for the waitlist? I just get a redirect loop back to the login when requesting access.
351.
by scosman 1y ago
It's just damn cool and worth the $2. Plus more reliable, no syncing, no web-connection. New content is just new content. Not how you'd build it at scale, but that's kinda the point.
352.
by scosman 1y ago
This is brilliant. Great design and great work. I love that the cartridges actually hold the content and aren't just NFC tags. For lazy parents like myself: check out a Yoto player. It's web-connected, has limited internal storage
353.
by scosman 1y ago
But read up on scaling limits before going down this path. Last time I checked it capped out at thousands of schemas, but no where near millions.
354.
by scosman 1y ago
What’s the disk behind this strategy? Lots of sqllite DBs is easy until you have lots of servers, no?
355.
by scosman 1y ago
This is just beautiful. A small, functional, electric truck. Not a luxury SUV with a tiny truck bed for cowboy cosplayers, or a cyberpunk glue heap. I hope they sell millions.
356.
by scosman 2y ago
You can use both properly and together. The service account should only be accessed on the service. If using Auth+Server, you can check the verified user identity via Auth JWTs (or something, see the docs). Yeah, don't use the server c
357.
by scosman 2y ago
It’s normal Postgres. There’s no need to handle everything on the front end. The tutorials nudge you to learn RLS and use their SDKs for the client, but you can write perfectly normal server side code as well.
358.
by scosman 2y ago
> There is no evidence in the fork that he is the copyright holder of the original code and it looks like Microsoft is. Only because they removed the license and copyright. If they were willing to do that in 1 file, they are willing to d
359.
by scosman 2y ago
If they forked from before the author had a license, it’s worse. MS had no right to use it. I’ve contributed to plenty of project that don’t have the per-file copyrights. It’s a choice not a mistake.
360.
by scosman 2y ago
Actual Deepseek. 500gb of memory and a threadripper works. Not a standard PC spec, but a common ish home brew setup for single user Deepseek.

More ›