News for hackers. Stuff that matters.

yorwba

Reading an article takes more time than commenting, which takes more time than voting. So the typical life cycle of a HN post is that people vote on it without reading the article, comment without reading the article, vote on the comments without reading the article, and respond to comments without reading the article, until after some time people who've actually read the article show up in the comments with some clarifications, saving everyone else the bother of having to read it themselves.

If you want to escape the hype, reading posts with no upvotes and no comments on https://news.ycombinator.com/newest is an option, though somewhat hit-or-miss.

Re:Astra can read and write weights to improve Qwen on Tuesday November 03, 2026 @08:49AM (#50351946)
"action 3 left" is documented in the input, so you can find the corresponding weights in column 3 of the output weight matrix. If that's past what you expect from humans or conventional algorithms, your expectations are too low.
Re:Ask HN: Any tool to remove AI Watermark from image? on Tuesday November 03, 2026 @06:37AM (#50350726)
I'm curious how you created AI images using Claude despite Claude not supporting this: https://support.claude.com/en/articles/9002504-can-claude-pr...
It's just irrelevant for big-O complexity what happens for instances below a constant threshold, so the algorithm uses the simplest approach and moves on. If you have an exponential-time algorithm that's faster than brute force for small instances, you can use that and get a faster overall runtime at the same big-O complexity. If you can lower the switchover threshold, ditto. It's a bit weird to dismiss the subexponential algorithm just because it's leaving an obvious optimization on the table... it can only get faster from here.
Re:The SEC Should Block Anthropic's IPO to Protect the Public on Monday November 02, 2026 @06:23PM (#50345600)
Maybe the SEC should instead block funds from distorting the market by making large predictable transactions, then.
Re:Astra can read and write weights to improve Qwen on Monday November 02, 2026 @03:16PM (#50343052)
In cases where Astra succeeded, it succeeded using standard techniques also used by humans when they read and write neural network weights. It's a rare skill, mostly used for fun (e.g. https://blog.janestreet.com/can-you-reverse-engineer-our-neu... ) and basic research rather than profit, but Astra is big enough to have learned many rare skills.
At least this 1984 report on "Aspects of Soviet Secondary Education: School Performance and Teacher Accountability" https://www.wilsoncenter.org/sites/default/files/media/docum... claims that a more ambitious curriculum did not lead to more repeaters ("grade repeating has declined over the years"), but instead "most teachers have chosen the easier solution of adapting their grading practices to the results expected of them." The students who didn't get algebra were apparently filtered out later: "According to the few statistics available on entrance exam results, it seems that somewhere between 20 and 30% of all students fail tests in their chosen field of specialization."
About $22B of it, yes.
People pay money for the solar panels, though. It's not obvious to me that giving away expensive LLMs for free is meaningfully helping grow the domestic economy, as opposed to foreign inference providers who get their GPUs from Nvidia.
So far, there have not been any obvious drawbacks to open-weight models from the government perspective, so when a company wanted to release their weights, they weren't prohibited from doing so. Similarly, there have not been any obvious drawbacks to closed-weight models, so when a company wanted to keep their weights to themselves, they weren't prohibited from doing so.

So the real question you should be asking yourself is why you thought that China supported publishing open-weight models, even though they merely didn't prohibit it.

Said the Ministry of Commerce at the press conference for the draft export controls: https://www.mofcom.gov.cn/xwfb/xwfyrth/art/2026/art_d7661dbf... 对于企业利用限制类大模型对外提供API调用、APP/网页交互、SaaS应用、本地端侧轻量化部署等情形,不属于大模型技术出口管理范围。
Companies making closed-weight models typically don't export them at all, but sell access to them as a service. (Which doesn't count as an export of the model.) So it's primarily companies releasing open weights (if they pass the thresholds for the restricted category), or that host their models in datacenters outside China that will need to apply for an export license.
The proposed export controls do add some categories where export is completely prohibited. For example, sandbox detection. While the provisions specifically for AI models are generous enough, it's still possible that they trigger some other prohibition by being too capable. The statement about encouraging open source only applies to cases where export is merely restricted, not prohibited, and even then companies need to report their intention to export in advance, in principle giving the government the opportunity to deny them an export license.

So they're changing some things.

Their method uses two models trained separately on images and captions and turns images into captions without ever having seen a captured image. So of course it's not very good at the task. Almost all the results in the paper are bad in an absolute sense. But they're surprisingly good considering the enormous handicap of not having any paired training data. So this might be not completely useless in settings where paired data is actually hard to come by.
Re:Domain hacks with unusual Unicode characters (2018) on Sunday November 01, 2026 @02:22PM (#50328772)
The URL from the top comment has been mangled by HN's "no fancy text" policy. You'll have to use the URL given in the post itself.
If you insist on the identity transformation, you'll fail even the basic "are these two different works or the same?" test that decided the court case you cited, so that's not going to work.
It comes with a reference implementation. I didn't see any obvious malware in there, but maybe the weights exploit a buffer overflow or something.
There is a ≈7 second gap between steps, suggesting the run had already been going for almost 5 months at that point...
Again, the court didn't rule on AI authorship, so your interpretation "1) AI chatbots are the originators of copyright" is wrong. It only ruled that the AI output was not the plaintiffs' work, therefore they couldn't claim that their work was being distributed, although they could've claimed (but didn't) that a work infringing on their copyright over their work was being distributed.

You can get what you call the "billionaire's treatment" by hiring a lawyer who understands the distinction between those two situations, who can accordingly pick the right law for you to sue under.

I'm confused by the framing that copyright protecting creative expression and not hard work is something that would come as a shock to media bosses. If you're in the business of buying up creative stories and paying for the hard work necessary to sell them to a large audience, then being able to automate the expensive part is good news. Imagine film credits listing just the director, screen writer and some assistants, with everyone whose job was never primarily about making creative decisions being eliminated.
The court in that case didn't rule on AI authorship at all.

"The coders describe an AI tool that doesn’t look up and reproduce stored work, U.S. Circuit Judge Eric Miller determined, but rather creates new work. That new work, the Donald Trump appointee said, may or may not infringe plaintiffs’ copyrights in their code, but it cannot reasonably be described as a copy of that code from which the copyright management information has been removed or altered."

... then all elements of the prompt that come from your human creativity and that survive into the final result remain under your copyright. Which in most cases should be almost as good as owning the full copyright, although it might influence fair use defenses if even a full copy infringes on only a minor portion of the work.
Were you under the impression that research costs can only be incurred in the US? Dizal had to spend money to develop the drug too. They recoup it by selling for the highest price people are willing to pay. From the article: "AZ paid Dizal $600 million upfront, and will owe royalty payments to Dizal based on global sales of the drug through AstraZeneca’s sales channels outside China." With those $600 million, they can now do even more drug research, which they will again sell at the highest price people are willing to pay and so on.
Re:Toxic experts (2017) on Thursday October 29, 2026 @01:36PM (#50292937)
The comments are preserved here: https://web.archive.org/web/20171108220351/https://nedbatche... It's a more subtle nitpick about how amortized runtime analysis is done.
They probably prefer being used by the deep-pocketed organizations that have access to Gemini 4 Carbon even if it means losing the entire rest of the market.
https://blog.google/innovation-and-ai/models-and-research/ge...

"We're taking a deliberate approach to rollout and are actively engaged in the U.S. government’s voluntary process for pre-release model testing. Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program. We’re continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via our product surfaces as soon as possible."

Re:Scaling Hierarchical Networks on Thursday October 29, 2026 @05:57AM (#50286641)
The whole point of this architecture is that it can also make use of multi-char patterns, but unlike BPE, they aren't hard-coded. And you can still use speculative decoding. Whether it works better or worse is something that you have to actually measure.
It's hardly surprising that ≈0.01% of the population of Barcelona are ready to do property damage. The only surprising thing is that they managed to all be at the same place at the same time. Identifying Tiktok as one link in their organizing process doesn't mean Tiktok is to blame, but it suggests that one way to prevent future occurrences could be e.g. to have a dedicated police unit monitoring social media content so they're not caught unawares.
Image-generation platforms can be subpoenaed for identifiers associated with the account that produced those images.
Re:Artificial Analysis Intelligence Index v5 on Wednesday October 28, 2026 @05:52AM (#50271145)
Their problem is really that individual benchmarks are not that informative. Bad enough models all perform at chance level, good enough models all saturate the benchmark. And the number of tasks is often low enough that differences of several percentage points might just be noise. Artificial Analysis' approach of aggregating scores in a weighted sum is largely just choosing which noise to amplify more.

I prefer Epoch's https://epoch.ai/eci methodology as they explicitly model benchmark saturation and also provide error bars for their estimates.

Older »