Clint Herron @HanClinto

Artificially Intelligent, Naturally Curious hanclinto.com Ohio Joined March 2009

Tweets

331
Followers

100
Following

229
Likes

5K

Clint Herron @HanClinto

a week ago

@washburnello @Pixel_Dailies Excellent choice! :)

0 0 1 7 0

View Details

Clint Herron @HanClinto

2 weeks ago

I got a kick out of this short story last night -- reposted it on my blog: blog.hanclin.to/posts/gh-33/

0 0 2 8 0

View Details

@washburnello @Pixel_Dailies "Rays" has surprisingly good depth given the resolution. I like the lack of an outline on folding chair -- it really works there. "Loaf" and "Rabbit" are great. Some really good ones in here

0 0 1 11 0

View Details

Clint Herron @HanClinto

2 weeks ago

@IsaacKing314 "Spikey" is a very good word for it. I've been trying to get rid of some of the spikiness by trying to get away from prompting everything and moving towards more... heuristic / repeatable evaluations? Not sure that really applies to your example. More scripts, less prompting.

0 0 1 14 0

View Details

Clint Herron @HanClinto

2 weeks ago

@washburnello @Pixel_Dailies "Hill" -- lol

1 0 1 9 0

View Details

Clint Herron @HanClinto

2 weeks ago

@washburnello @AlloyAmi008 Nice, thank you for the recommendation

0 0 1 19 0

View Details

clem 🤗 @ClementDelangue

3 weeks ago

llama.cpp with MTP support makes local models fast enough to use as daily drivers 🚀 Qwen3.6-27B dense generation below on A10G: From 25 tok/st to 45 tok/s (+78%)!

83 117 1K 138K 877

View Details

Georgi Gerganov @ggerganov

3 weeks ago

Highlighting the new WebGPU backend in llama.cpp/ggml The work to bring full-fledged WebGPU support in llama.cpp started about an year and a half ago. It has been lead by @reeselevine and team at USCS. For more information, checkout the interactive blog and paper in the quoted post. Here are 2 excerpts from the paper, summarizing the implemented software architecture.

Reese Levine @reeselevine

4 weeks ago

WebGPU support in llama.cpp is here! Check out our blog post introducing it: reeselevine.github.io/llamas-on-the-… Run local models in your browser, with GPU acceleration. No data leaves your computer! Thanks to everyone who's made this possible, especially @ggerganov

5 27 126 737K 61

13 109 618 91K 317

View Details

Victor M @victormustar

4 weeks ago

llama.cpp with MTP support makes local models fast enough to use as daily drivers 🚀 Qwen3.6-27B dense generation (on A10G): From 25 tok/s → 45 tok/s (+78%). Two flags on llama-server: --spec-type draft-mtp --spec-draft-n-max 2

Georgi Gerganov @ggerganov

4 weeks ago

llama.cpp adds MTP for the Qwen3.6 family This is a significant milestone for the local AI ecosystem. The performance jump with these changes is massive and elevates local inference on commodity hardware further. Special thanks to Aman Gupta for leading this development!

48 180 1K 273K 533

41 124 1K 171K 1K

View Details

Clint Herron @HanClinto

2 months ago

CollectorVision is an open-source, open-weight, state-of-the-art image recognition library for collectible trading cards (Magic, Pokemon, Yugioh, etc). It is not a standalone product -- it's meant to be integrated into other sites and tools that would benefit from having fast, accurate, fault-tolerant card recognition. This demo is 100% local -- images never leave the client. The only thing the network is used for is to look up card names and prices from a third party site. Try it out and let me know what you think! #mtg #opensource

1 0 2 32 0

View Details

Clint Herron @HanClinto

2 months ago

@esrtweet @TONYxTWO Seriously. For a second I missed the scaling on the graph and thought it was saying 120 firearms per person, and I got really self-conscious that I was a bit behind the average. Then I saw the scaling, and I was disappointed in my neighbors.

0 0 1 17 0

View Details

Clint Herron @HanClinto

2 months ago

@W1stCallieFan @Ava_AM12ff absolutely. Some of my favorite friends are people who disagree with me. One needs that. Proverbs 27:17 -- iron can't really sharpen iron if it's not held at an angle to it. We need disagreement to refine us, and ideally we are both better for it.

0 0 0 11 0

View Details

Clint Herron @HanClinto

2 months ago

Two mathematicians are talking. One says a theorem is trivial. After two hours of explanation, the other agrees that it is indeed trivial.

0 0 2 15 0

View Details

Clint Herron @HanClinto

2 months ago

@TheCinesthetic Rescuers Down Under

1 1 6 544 0

View Details

Clint Herron @HanClinto

3 months ago

0 0 1 14 0

View Details

Clint Herron @HanClinto

3 months ago

@RealAstropulse Community sharing sounds amazing. Can we use the styles to generate a similarity embedding, so that we can search public styles by "show similar"? This could provide really fun exploration.

1 0 1 31 0

View Details

Clint Herron @HanClinto

3 months ago

This is the sort of thing that I find so inspiring and reminds me we can live in the "flying cars" future I always dreamed about as a kid. This is absolutely the sort of thing that makes me thrilled about working on building the future of AI, and is the sort of thing I want to enable more of.

Paul S. Conyngham @paul_conyngham

3 months ago

x.com/i/article/2036…

245 931 5K 3.4M 5K

0 0 0 16 0

View Details

Clint Herron @HanClinto

3 months ago

This new algorithm felt cute. Might delete later. Playing around with neural networks and some new scanning algorithms. Uses a couple of custom-trained neural networks -- one to predict card corner points (the "detector"), and another to hash / vectorize the image (the "identifier") and do nearest-neighbor search to find the closest match in our image database. It does not need to be re-trained every time new cards are released. Both models are sized to be pretty small and lightweight -- should be able to run locally in fairly limited resources (in a browser, or on a phone, or in a Raspberry Pi). This current demo is just slapped together, and streams images up to a local web server running on my home network. Not optimized at all -- the FPS I'm sure could be much faster, but I set the framerate to 3 fps out of convenience and didn't think to bump it up before recording. This is all still extremely preliminary and not super-tuned yet. The corner detector only has 15 epochs into it, and I think it can get a lot better. The identification network is looking good, and still has a lot of headroom to improve, but I stopped training because I didn't want to overfit too much on Magic, and wanted to make sure I got some other CCGs into the mix before training it too much more. This is not tied to any of my previous work, and could be licensed for use in your product. Interested? Send me a message!