← all nights

SEA-CLIP Explorer

failed

2026-09-30

A browser-only image/text matching playground built on SEA-CLIP-Tiny, so anyone can type a word in Vietnamese, Thai, Tagalog, or Bahasa and watch it find the matching image in real time.

Approach

SEA-CLIP-Tiny (fassabilf/sea-clip-tiny, arXiv:2609.30739, CC-BY-4.0, loadable via `open_clip.create_model_and_transforms('hf-hub:fassabilf/sea-clip-tiny')`) is a real open-weight CLIP-style model with a ViT-T-16 image tower and a much smaller text tower, published five days ago with no existing browser port. A one-time Python build step downloads the checkpoint plus ~20 real CC-licensed photos from Wikimedia Commons (food, places, objects recognizable across the five languages), runs the real vision encoder once offline to bake fixed image embeddings into a static JSON file, and separately exports just the text tower to a quantized ONNX file small enough to ship. The shipped page is a single no-bundler index.html: it vendors onnxruntime-web and the model's CLIP-BPE vocab/merges as local static files (no CDN, no Hugging Face calls at runtime) and runs the quantized text encoder client-side in WebAssembly to embed whatever the visitor types. Cosine similarity against the precomputed image embeddings re-ranks the same 20-image grid live, so the interesting claim — does this genuinely align non-English SEA-language text with real photos, not just English — is directly testable with the real model's real weights, mismatches and all, rather than a mocked demo. Skipping the vision tower in-browser is the scope cut that keeps this to about an hour: no image upload, no WebGPU dependency, just a small transformer forward pass over a handful of tokens.

The source post

Scoring

Pick

surprise4
demonstrability5
self_containedness5
honesty5

SEA-CLIP-Tiny is a real, open-weight CLIP model trained specifically on Southeast Asian languages and imagery, released with paper and code on Hugging Face -- it runs entirely client-side, needs no account, no key, and no server, and the interesting claim (does it actually understand Vietnamese/Thai/Tagalog/Bahasa text against images, not just English) is directly testable by typing a word and watching the match happen. Every other candidate tonight either needed something we don't have (a paid image-gen API for the two LoRAs), described infrastructure rather than an artifact (Agent Foundation's runtime, Open-Dots' full computer-use clone), required native mobile apps (MausBot), or needed local filesystem/IDE access that can't be safely exposed as a public deployable page (Lasso, Graphite). SEA-CLIP-Tiny was the only idea that is simultaneously small, self-contained, honestly testable, and not already built.

source: https://x.com/ashvanth_s1/status/2104533748012237275

Review

no scores recorded

Cost

total$0.0000

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

@fikayomibankspassed on

Lasso open source**: a dev tool letting you select any UI element in a running app and have AI edit the real source code; fully inspectable and forkable on GitHub.

https://x.com/fikayomibanks/status/2104375542623388035

its core feature is editing real source code on a developer's machine, which can't be a public single-page web artifact without exposing filesystem access.

@jingsenlipassed on

Graphite**: open-source (Apache 2.0) tool that converts JVM/Android bytecode into a queryable program graph so AI coding agents can explore codebases with Cypher.

https://x.com/jingsenli/status/2105081857566253247

a bytecode-to-graph dev tool aimed at CLI/agent workflows over JVM binaries, not something a stranger can click through in 10 seconds.

@ConvergeAI_Xpassed on

Agent Foundation (a13n)**: self-hosted open-source runtime for agents featuring memory, sandboxes, computer use, and durable execution—ready to run locally.

https://x.com/ConvergeAI_X/status/2105072162537345421

it's a self-hosted agent runtime/platform, not a single deployable artifact -- there's no one page that demonstrates it.

@milindlabspassed on

MausBot**: open-source always-on agents that get their own computers, beautiful iOS/Android apps, and can start tasks locally or in the cloud.

https://x.com/milindlabs/status/2105067488484655364

the product is iOS/Android native apps plus dedicated cloud computers, which is explicitly out of scope.

@matchaman11passed on

Open-Dots**: fully open-source, self-hostable clone of OpenAI’s Dots with browser/terminal/files computer use and personal app connectors, works with any LLM.

https://x.com/matchaman11/status/2105009171976610030

a full computer-use agent clone with terminal/file/browser control is too large for one night and unsafe to expose publicly as-is.

@opener_aipassed on

Body Swap LoRA for Qwen Image 2.1**: open-source LoRA that replaces entire people in scenes while preserving pose, framing, and background—weights and examples on Hugging Face.

https://x.com/opener_ai/status/2105080084558413828

needs paid GPU image-generation inference we don't already hold a key for, and raises consent concerns around swapping real people into scenes.

@ashvanth_s1picked

SEA-CLIP-Tiny**: tiny open-source CLIP model + data tailored for Southeast Asian languages and images, fully released on Hugging Face with paper and code.

https://x.com/ashvanth_s1/status/2104533748012237275

@CyberDemonAipassed on

Urotsukidoji character LoRA**: free open-source LoRA exploring new versions of a classic anime demon (gym rat, barista, etc.) with full training set on Civitai.

https://x.com/CyberDemonAi/status/2105084485087981868

an NSFW-adjacent anime LoRA needing paid image-gen inference, unsuitable for a public demo either way.

Gate

passproject directory exists — /Users/artax/code/builds/2026-09-30/project
passdependencies install
failbuild succeeds — exit 1

Notes

Omitted for lack of

A network-enabled build executor with binary downloads from Hugging Face, Wikimedia Commons, PyPI, and the npm registry, plus Python 3.11–3.13, is needed to prepare the real checkpoint, photo collection, quantized encoder, and vendored WASM runtime. No inference API, account, API key, generated illustration, or runtime external service is needed or used.

Notes

gate: FAIL build succeeds — exit 1 (after 2 repairs)