HN rank: 1

HN score: 108

comments: 38

original url: https://tencent-hunyuan.github.io/Hunyuan3D-WorldClaw/

评论信号：In case it's not obvious, this isn't a model, this is python scripts that call out to models (code not available).It's mostly "did you know you can attach an LLM to a PCG system?", but there is one idea here you don't see much of: an image model performs the composition (which image models are really good at), and then you extract the objects into 3d via things like SAM3D before placing them in the world, which is pretty interesting.The rest is standard stuff you'll find in your favorite PCG system/game engine. Still, think most people don't realize how good LLMs are at 3D these days, especially Fable 5 (which this project predates). reply: That’s exactly how I’ve used tools like Tripo and Meshy.Use any image generation tool (or any image) and have a multimodal LLM like Gemini extract individual parts of the image “isolate with a transparent background” - then you can use those images for image-to-3d in those tools.Don’t forget to use low or “smart” poly features otherwise you get too many vertices to the point you can’t performantly raycast etc.But yeah - it’s there.

评论信号：It looks impressive. However open worlds are at their best with hand placed details and environmental storytelling. Look at Skyrim/Cyberpunk vs. Starfield where most of the world is proc gen’d.I get that you run this, and then edit it. But the generated villages just aren’t interesting in my opinion. It’s probably great for tencent’s market where you’re mass producing gacha style games, but I don’t think worlds made in this fashion will scratch the open world itch like the best open world games out there. reply: You know what's also good at storytelling? AI. Or else we wouldn't have writing awards being revoked when it comes out it was written by AI.I wouldn't say we don't need artists or writers, but I think most people will be surprised how quickly fully AI driven pipelines will advance here.

评论信号：In both the Autumn and Winter examples in the hero images, the algo seemed to have placed buildings on the water in the foregroundIn the summer example in the hero images, the building placement + small pockets of water on the left looks odd and low attention to detail. A similar poor quality result as if an uncaring human used a scatter brush...Curious if the examples are cherry-picked and by how much, or if this is one-shoted reply: I am sure they are cherry picked. No one would spend time trying to publish results like this without picking the best results they possibly could to showcase.

HN rank: 2

HN score: 87

comments: 49

original url: https://www.cftc.gov/PressRoom/PressReleases/9281-26

评论信号：EDIT: see DannyBee's comment below ( https://news.ycombinator.com/item?id=49266746 ). It does look like the CFTC has extrapolated the "nationwide" out of either some novel interpretation of the filing or just entirely fabricated it as their justification for their action.I've got no love for Kalshi, but "orders Kalshi to continue operate in New York" doesn't seem to be present anywhere in the actual release.The article presents the sequence of events as:1. The State of NY files a lawsuit against Kalshi under the theory that it can be regulated by state gambling laws.2. The State of NY files for a temporary restraining order requiring Kalshi to halt trading nationally, not just in NY.3. Kalshi reaches out to the CFTC to claim that NY doesn't have the authority to regulate interstate commerce.4. The CFTC agrees and uses their authority to override the TRO.That seems pretty aligned with how interstate commerce is regulated and managed in the US. reply: It seems that New York is asking the court for a temporary restraining order that would prohibit Kalshi from offering all event contracts nationwide....

评论信号：Am I reading this right? The state of New York wants Kalshi to stop offering prediction gambling, excuse me, "event contracts" in New York, and the federal Commodity Futures Trading Commission just ordered Kalshi to keep operating in New York regardless? reply: Here is a link to a post about the State of New York's case: https://www.shb.com/intelligence/newsletters/securities-liti...I am not sure about how the state regulation of betting will turn out (though I would have guessed that it is indeed pre-empted), but the nationwide injunction seems shaky given Trump v Casa: https://www.supremecourt.gov/opinions/24pdf/24a884_8n59.pdf

评论信号：This seems correct, even though I think it's unfortunate that these prediction markets are ruining so many people's lives.> In the lawsuit, filed on July 31, New York seeks a temporary restraining order prohibiting KalshiEX, LLC from offering all event contracts nationwide and more than $36 billion in damages.This seems naturally the territory of the CFTC. They have exclusive right to regulate futures and derivatives contracts, which Congress handed them. Also, it seems straightforwardly anti-commerce-clause to allow NY to prohibit Kalshi from offering these contracts nationwide.

HN rank: 3

HN score: 242

comments: 103

original url: https://ngrok.com/blog/compression-is-prediction

评论信号：This is the thesis behind the "Information Theory, Inference, and Learning Algorithms" course that was taught at Cambridge University.> Why unify information theory and machine learning? Because they are two sides of the same coin. In the 1960s, a single field, cybernetics, was populated by information theorists, computer scientists, and neuroscientists, all studying common problems. Information theory and machine learning still belong together. Brains are the ultimate compression and communication systems. And the state-of-the-art algorithms for both data compression and error-correcting codes use the same tools as machine learning.Book (creative commons): https://www.inference.org.uk/mackay/itila/book.htmlLectures: https://m.youtube.com/playlist?list=PLruBu5BI5n4aFpG32iMbdWo... reply: I had a long ranting comment I deleted....

评论信号：Grant Sanderson has an excellent video on the same topic [0]. It's part of a series that is ongoing.[0] Compression is Intelligence Part 1 - https://youtu.be/l6DKRf-fAAM?si=yyLWq8x4sSRkWd98 reply: I wonder if the author of the article knew about the series, or do they both just independently came across this topic to talk about it.

评论信号：Nope; there is a bit more nuance and the distinction is important.Compression is functionally equivalent to prediction when the data distribution is exactly representative of all future problems. The story changes drastically if you want generalization -- because the test distribution could be arbitrarily different, even if it had the same support! Eg: you observe a rare edge case in your training data and (lossy) compression could simply ignore it. But if you wanted generalization in that particular part of the space -- either because an adversary was testing you, or for design freedom where you choose to build in that specific corner -- then you don't just want data compression, but good prediction performance on a test distribution which peaks in that corner.Assuming that the training data distribution is exactly the distribution you will ever care for is implicitly doing a lot of the heavy lifting in the claim that compression = prediction, and I'm peeved at how much this statement is unthinkingly repeated like a manifesto.There is nothing natural about the training data distribution, especially if the data generation process is exploratory while the downstream usage will be ex...

HN rank: 4

HN score: 176

comments: 90

original url: https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/

评论信号：Coincidentally I've been playing with small (~30B) self-hostable models for coding tasks today -- specifically plugging them into Cloudflare OS (which I work on) and asking each to build a collaborative whiteboard.I'm finding that the Mixture-of-Experts (MoE) models (Qwen 3.6-35B, and Nemotron 3.5 Lightning) are, well, terrible at this. They just couldn't get the job done at all. Went way off the rails. They are really fast though!Whereas ~30B dense models (not MoE) are pretty decent. I tried Muse Glimmer, Gemma 4-31B, Qwen 3.6-27B, and Laguna XS[0]. They were all able to build a working collaborative whiteboard app, without any guidance (other than feeding back error logs to the model). I also asked each to then draw a monkey by calling the API of the whiteboard it has just built. Laguna drew random scribbles but the rest all managed to produce something monkey-like.(Frontier models in comparison will write the app in one shot with no errors at all.)Note that both Qwen 3.6 and Gemma 4 each have both MoE and dense variants. I find this very confusing, because e.g. ollama's model index typically only distinguishes variants by their size, but MoE vs.... reply:...

评论信号：One major consequence of the ramapocalypse, I think, is an even higher focus on small efficient models. I personally believe that the multi-trillion parameter models are fundamentally missing things and the push to smaller, more efficient will drive evolutionary structural changes that will lead to future gains reply: I'd assume the closed weight models are all working on shrinking their parameter counts anyways. They too benefit from smaller models. It'd be foolish for these SOTA labs to not be working at reducing parameter counts.

评论信号：> NeMo Switchyard, an open source library for smart routing> When deployed, NeMo Switchyard can intelligently direct each request to the most capable and suitable model for the jobHow do routers like this handle prompt caching when you send the second request?Sticky models per session? but then the second message of that session won't be sent to a suitable model, and will only be sent to the same model as previous one. reply: I've seen ones that are configurable to pick a trade off point between lower cost (cache stickiness) and routing performance (best model for that turn).But yeah I'm skeptical all this overhead is worth it.

HN rank: 5

HN score: 283

comments: 128

original url: https://www.modular.com/blog/modular-26-5-mojo-1-0-is-here

评论信号：I feel like this language would really benefit from some sort of 1-pager overview.I just spent a fair bit of time on the official site, and I still don't think I have a very good grasp of what problem this language aims to solve, or why I would select it over other similar languages reply: Having written a lot of Mojo over the last two year, just for fun, it's a really cool language. Ownership model adjacent to Rust, comptime in the realm of Zig, rich type system, first class SIMD support, etc. Performance wise it's the first language in long time that isn't just an LLVM wrapper. LLVM is still involved, but they are using it differently than say, Rust or Zig.Very excited for Mojo once it's open sourced later this year.

评论信号：Don't see the value of using a language with a closed source compiler...Much better options out there. Python already has libraries like Pydantic that offload performance to functions written in Rust under the hood. reply: The point is that Mojo compiles to MLIR for programming on heterogenous compute and GPU so something like Pydantic isn't really comparable here.The alternatives would be stuff like CUDA

评论信号：AI generated first image does not give me much confidence.Lastest OpenCV 5 release notes also had a lot of LLMisms.I guess that's the new normal. Still, I am very hopeful for Mojo. reply: I know right? Would be better to have a banner with the language logo than some totally out of place aislop image, almost makes you think this is a random person's blog and not the official Mojo site