@jeresig

Creator of @jquery, Chief Software Architect at @khanacademy, Creator of https://xitter.cf/t.co/UcOO6CdK7c

 Hudson Valley, NY
johnresig.com
Joined February 2007
  • Tweets10.4K
  • Following2.5K
  • Followers233K
502Media
I love how good LLMs are at training other models. I used sol & astra to do annotations and classification on 10s of thousands of prints, then used opus-5.5 to manage the actual training, and now I have a fine-tuned detector and a classifier I can use to help categorize prints!
A view showing a Japanese print of kabuki actors. The print itself is annotated with boxes around faces, title, signatures, and the print itself. Next to it are classifications confirming that this is, in fact, a print, we're seeing the front of it, it's one sheet, it's unmounted, vertical, rectangular, very likely to be a Japanese print, and depicts kabuki actors show are playing famous warriors from history.

ALT A view showing a Japanese print of kabuki actors. The print itself is annotated with boxes around faces, title, signatures, and the print itself. Next to it are classifications confirming that this is, in fact, a print, we're seeing the front of it, it's one sheet, it's unmounted, vertical, rectangular, very likely to be a Japanese print, and depicts kabuki actors show are playing famous warriors from history.

15
3,823
There's this one print series by Hiroshige & Kuniyoshi that has this huge title cartouche that causes my image similarity models to match every other print in the series. 15% of my DB is matches from this one series! I have to hardcode an ignore for this specific cartouche!
A grid showing prints by Kuniyoshi and Hiroshige that have a large black and white title cartouche in the top right, all with the same text. The prints themselves tend to have landscapes and pictures of beautiful women.

ALT A grid showing prints by Kuniyoshi and Hiroshige that have a large black and white title cartouche in the top right, all with the same text. The prints themselves tend to have landscapes and pictures of beautiful women.

2
1
1
8
4,499
Was really hard to narrow this down. Many of these games led me to playing other similar games, which is why they're so important to me. Nowadays I'm mostly playing roguelikes, narrative games, and CRPGs. Not listed: Tetris-likes, Minecraft, and Platformers!
Sim City 200, Civilization II, Age of Empires, Unreal Tournament, Animal Crossing, Hearthstone, Slay the Spire, Disco Elysium, Baldur's Gate III

ALT Sim City 200, Civilization II, Age of Empires, Unreal Tournament, Animal Crossing, Hearthstone, Slay the Spire, Disco Elysium, Baldur's Gate III

4
30
6,060
If you were confused by this announcement like I was (I already use Claude Code cloud a lot!) this is a credit to pay for some token use in the cloud. I spent 30mins trying to find the new pricing (that didn't exist). Cloud VMs won't cost anything extra (as has been the case).
ClaudeDevs
@ClaudeDevsSep 23
Cloud sessions are officially available and out of research preview! They let you keep Claude Code working, even when your laptop is closed. Existing subscribers get a one-time credit to try them: $100 on Pro, $250 on Max.
5
1
1
14
8,097
gpt-6-luna (flex, low thinking) is phenomenal for text extraction. Half the cost of the next cheapest and just as accurate as gpt-5.6-luna. I switched from gemini-2.5-flash-lite when I started getting refusals for text extraction (?). Much better quality with luna in general.
Scatter chart titled "Comparing text extraction with gpt-6 and 5.6 luna." It plots the average of accuracy and completeness (0–10), graded by Claude Opus 5 on 47 Japanese print listings, against cost per 1,000 listings on a log scale. gpt-6-luna scores 8.1 at $0.20 with the original prompt and 8.3 at $0.24 with a tuned prompt, joined by a dashed line. gpt-5.6-luna scores 8.5 at $0.51 originally and 8.4 at $0.57 tuned. Both Luna models ran on OpenAI's flex tier. Other models: gemini-3.5-flash-lite 8.1 at $1.39, gemini-2.5-flash-lite 7.7 at $0.34, glm-5.3-flash 5.8 at $0.72, mimo-v2.6-flash 5.7 at $1.53, deepseek-v4.1-flash 5.5 at $2.79, and minimax-m3 4.5 at $0.73. The tuned gpt-6-luna nearly matches gpt-5.6-luna at under half the cost.

ALT Scatter chart titled "Comparing text extraction with gpt-6 and 5.6 luna." It plots the average of accuracy and completeness (0–10), graded by Claude Opus 5 on 47 Japanese print listings, against cost per 1,000 listings on a log scale. gpt-6-luna scores 8.1 at $0.20 with the original prompt and 8.3 at $0.24 with a tuned prompt, joined by a dashed line. gpt-5.6-luna scores 8.5 at $0.51 originally and 8.4 at $0.57 tuned. Both Luna models ran on OpenAI's flex tier. Other models: gemini-3.5-flash-lite 8.1 at $1.39, gemini-2.5-flash-lite 7.7 at $0.34, glm-5.3-flash 5.8 at $0.72, mimo-v2.6-flash 5.7 at $1.53, deepseek-v4.1-flash 5.5 at $2.79, and minimax-m3 4.5 at $0.73. The tuned gpt-6-luna nearly matches gpt-5.6-luna at under half the cost.

6
5
79
9,025
Switched our internal @KhanAcademyEng PR review bot over from Opus 5 to Opus 5.5: Cost decreased by 50% Tool calls decreased by 34% Execution time decreased by 62% No decrease in quality.
10
6
1
177
12,596
I just had to explore using Jev with my existing Japanese print metadata extraction pipeline (where I use gpt-5.6 luna). Turns out that in some cases I could replace luna completely and save a bunch of money - in others I could augment what I had for higher quality!
Table titled "Luna vs. Jev: Japanese Print Metadata Extraction" comparing GPT-5.6 Luna with Jev on three decisions. Which artist is this? Luna: 6 wrong, $0.29 per 1k, 1.1 s. Jev: 2 wrong, $0.04 per 1k, 0.28 s. Is this translation actually Japanese? Luna: 0 of 60 Chinese or untranslated answers caught. Jev: 60 of 60 caught. Medium and format of the object: Luna 8 wrong of 124. Jev 0 wrong, abstains on 14.

ALT Table titled "Luna vs. Jev: Japanese Print Metadata Extraction" comparing GPT-5.6 Luna with Jev on three decisions. Which artist is this? Luna: 6 wrong, $0.29 per 1k, 1.1 s. Jev: 2 wrong, $0.04 per 1k, 0.28 s. Is this translation actually Japanese? Luna: 0 of 60 Chinese or untranslated answers caught. Jev: 60 of 60 caught. Medium and format of the object: Luna 8 wrong of 124. Jev 0 wrong, abstains on 14.

6
1
1
40
7,843
I love giving Fable problems that I 100% would've never done otherwise, such as collecting Japanese prints out of static image auction catalogs written entirely in Japanese. Uses OpenCV to do segmenting, GPT-5.6 Luna to do text extraction + translation.
An image from an auction catalog showing prints with captions written in Japanese. Then the same image but segmented by opencv into distinct groupings. The a final image showing all the images extracted in aggregate in a single database, in English.

ALT An image from an auction catalog showing prints with captions written in Japanese. Then the same image but segmented by opencv into distinct groupings. The a final image showing all the images extracted in aggregate in a single database, in English.

3
33
7,219
A little experimental game I'm making: carving Japanese-style woodblock prints! Wanted something chill, atmospheric, but most importantly: accurate to the actual experience of carving woodblocks! Still messing around with it but I've been having fun building it!
3
1
36
7,857
Incredible - thank you for the reminder, @HenriHelvetica! 20 years!!
Henri Helvetica 🇭🇹 👨🏾‍🚀 🚀
@HenriHelveticaAug 26
🎂 Many of those responsible are quite possibly no longer on this platform, but 20yrs ago TODAY the mythical @jquery launched v1.0. As usual, sharing a thread I posted after a short email exchange I had w/ @jeresig a few yrs back, ICYMI. xitter.cf/HenriHelvetica/status/…
19
51
2
488
38,438
Ripped Google Analytics out of ukiyo-e.org this past week and replaced it with a better, more detailed, home-grown setup that doesn't use any cookies! I didn't realize how much traffic I'm sending to museums - about 3,700 clicks per day!
A dashboard showing the traffic to ukiyo-e.org over the past week. 1,840,471 image impressions, 91,418 image views, 24,984 print visits, and 1,071 site visits.

ALT A dashboard showing the traffic to ukiyo-e.org over the past week. 1,840,471 image impressions, 91,418 image views, 24,984 print visits, and 1,071 site visits.

8
2
46
8,867
"Impressions" is when someone sees an image inside a search result -> 5% conversion to a viewing a print image on my site -> 27% conversion clicking onto the final museum site. Averaging about 1,500 manual searches and about 200 search-by-image per day. Neat!
1
1
2,012
I've finally found a Wordle-like game that I'm good at! You have to associate historical artifacts with their origin and time of creation. Very fun! Anthropeum.com · Jul 31 2026 🟦🟩🟩🟨🟩🟩🟩🟦🟦🟦 82,910 · top 2% of players today!

Anthropeum

A daily game that challenges our understanding of human cultures. Ten objects. 5,000 years of human history. Guess where and when each artifact was made. See how you rank.

anthropeum.com
2
2
16
7,721
The new Github Stacked PRs preview is incredible. Landing 5 stacked PRs directly to a merge queue all at once! A+++! This removes so much friction (and the gh cli tools + agent skill help a ton) github.github.com/gh-stack/
A display of 5 pull requests queued to merge into a merge queue all at once.

ALT A display of 5 pull requests queued to merge into a merge queue all at once.

4
15
116
31,072
I have a physical object that I can take photos of from all angles and I want to turn it into a relatively low-poly 3d model for a game (mapping the photos onto the surface). I feel like there's gotta be some simple non-AI tool to handle this?
10
16
9,515
The Washington Monument on the Potomac River, Kawase Hasui, 1935 - depicting the cherry trees that were a gift from the Japanese government and planted in 1912.
A Japanese woodblock print showing the Washington monument with the tidal basin below with a row of bright pink cherry trees in-between.

ALT A Japanese woodblock print showing the Washington monument with the tidal basin below with a row of bright pink cherry trees in-between.

12
34
4,963
My pie-in-the-sky requests for @CloudflareDev to overhaul corporate web dev: * Fully hosted private Git + CI + "Pull Requests" * Playwright end-to-end test running and result collection * Storybook hosting + shapshot comparisons * Private NPM package registry hosting
6
2
1
52
6,663
I feel like too many Github competitors try to focus on the public-facing part of Github. There are too many security problems there. So much can be gained by keeping things private and protected, especially with agents running in there.
1
7
2,232
This is a super-smart move by Cloudflare - really excited to see these teams work together even more!
Cloudflare
@CloudflareJun 4
VoidZero, the team behind Vite, Vitest, Rolldown, Oxc, and Vite+, is joining Cloudflare. Vite stays open source, vendor-agnostic, and built for everyone. cfl.re/3Q1XYSX

VoidZero is joining Cloudflare

VoidZero, the team behind Vite, Vitest, Rolldown, Oxc, and Vite+, is joining Cloudflare. Vite stays open source, vendor-agnostic, and built for everyone.

blog.cloudflare.com
1
2
33
6,503
deepsec is very cool! I ran it against our main backend repo at @khanacademy and it found some really interesting edge cases. Spun up an agent team in Claude and addressed all the high-impact issues. This likely would've been a multi-week, or month, effort otherwise!
Vercel Developers
@vercel_devMay 4
Introducing deepsec, an open source coding security harness. • CLI-first • Sandbox-based scaling • Pluggable coding agents • Designed for large-scale repos • Use AI Gateway or your own subscription After months of successful internal use, we put it to the test on some of the largest open source codebases. vercel.com/blog/introducing-…

Introducing deepsec: The security harness for finding vulnerabilities in your codebase

Today we're open sourcing deepsec, an AI security harness that runs on your infrastructure, with your keys, against your code.

vercel.com
3
10
1
153
59,931
Load more