HN rank: 1

HN score: 35

comments: 14

original url: https://jross.me/open-source-as-we-know-it-is-dead/

评论信号：I could not imagine the pain being being the head of a Libre/Open project nowadays.It used to suck when there was almost nobody contributing because it felt isolated. Worse was if your project did become popular, you would be managing loads of code AND spending loads of time managing the community than the code.Nowadays, you probably don't have the community to push back and converse with, just a flood of code with few people you can't actually question it with. So back to isolation but with a million lines of code.It feels like you would have to have a project with an direct artificial limit to increase the quality of submissions. Like, all submissions must be able to run on a 486 25Mhz and 4MB of RAM, this would keep the code small and manageable. Don't do that, but you get the idea. reply: > Like, all submissions must be able to run on a 486 25Mhz and 4MB of RAM, this would keep the code small and manageable. Don't do that, but you get the idea.Why not? It works for Dusk OS!

评论信号：I think the answer is straightforward. You complain that a substantial PR is no longer evidence of good faith. What is evidence? Participation in the community.I think mailing lists (or an equivalent) are coming back. In a big way. reply: Goodhart's Law. If community participation becomes a hard requirement we'll see AI agents swamp them too.

评论信号：There is also the fact that Maintainers simply don't need contributions any more. This I've noticed also happens at work. Everyone is maintaining their own mini stacks of things. Contribution is up cooperation is down. Why go talk to a colleague when you can talk to Claude? Change is everywhere.

HN rank: 2

HN score: 21

comments: 15

original url: https://www.vivienhenz.com/common-lisp

评论信号：> In Common Lisp your program won’t crash, it’ll stop and open a debugger with the whole stack and all the variables. You can just point your LLM at the debugger, and it’ll make its fix and resume the program.> To my knowledge Common Lisp is the only mainstream language that does all of this.Sounds like someone who has never used C# and Visual Studio? Even JavaScript is capable of doing this, honestly JavaScript might be the one language with the richest developer tooling of all time (possibly?), sad to say because there's nicer to work with languages out there. reply: IF you're running a debug build. I guess Common Lisp can do this in prod?

评论信号：related: https://opusmodus.com/forums/blogs/entry/4-connecting-an-ai-...

评论信号：cant you do most of this with smalltalk too, or erlang?I dont really understand what macros get you when llms exist since a llm doesnt really need to create dsls to get work done reply: Smalltalk, in image based systems, absolutely. Erlang not as much, it's more of a "crash the actor and try again" than "rewind, pause, and handle errors in-system"

HN rank: 3

HN score: 369

comments: 110

original url: https://reflection.ai/blog/introducing-beam

评论信号：Always glad to see more open-weight models, but this caption on the 2nd demo image had me do a double-take: "Land or Water Generalization Experiment: We recreated the viral X puzzle by asking Beam to create a fixed 180×90 grid for longitudes -179° to 179° and latitudes -89° to 89°, with 16,200 points. This puzzle is a few days old, so could not appear in the training data, thus testing the model’s generalization. Beam gets 95.5% coverage right, putting us between Opus 5 (92.5%) and Fable 5 (97.8%), which shows how well it generalizes to novel new tasks."Oof, no, this "puzzle is a few days old" is incorrect even if it's a social media trend just recently. Asking a model to generate a world map in this way is at least from August 2025 as it appeared on LessWrong at that time: https://www.lesswrong.com/posts/xwdRzJxyqFqgXTWbH/how-does-a... reply: Yeah I remember when the original post about this came out. Def not recent. Though I think their point survives in that they didn't exactly RL on this.

评论信号：> Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads.> Beam’s capabilities come from major investments in both pretraining and reinforcement learning (RL). We pretrained the model on 23.8 trillion diverse, curated, high-quality tokens from the web and proprietary licensed datasets, matching or outperforming available similar-sized open base models. In parallel, we developed the algorithms, training environments, and infrastructure needed to sustain high-compute RL at exceptional scale. Our high-compute RL run generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks of training.Early access, no weights no tech details, just a sign up here for info reply: I'm all for more open models, but talk is cheap and this is a rather pointless announcement without anything backing it up. Publish your weights and HF repo or shut up IMO.

评论信号：I thought it would be interesting to look at some key figures vs another contemporary model in the same weight class (DeepSeek V4.1 Flash) DS V4.1F Beam LM total params 552B 501B LM active params (prefill) 8B 23B LM active params (decode) 16B 23B N-gram/PLE params 196B 0 Pretrain tokens 45T 28T Disk KV bytes/token (FP4) 890 No information Vision Yes (pretrain) No Weights available Yes (launch day) "This month" Weights licence MIT Apache 2.0

At first blush the benchmarks are impressive, but to paraphrase Linus: "Talk is cheap, show me the weights." :-) reply: yeah agreed, let's normalize "show me the weights" in this era

HN rank: 4

HN score: 142

comments: 47

original url: https://flattensf.com/

评论信号：Checkout https://bikehopper.org for solid elevation, bike infrastructure and transit-aware routing around the Bay Area. Some friends and I built this about 5 years ago and continue to maintain it. We use 1m DTM elevation data for SF and 50m for rural areas. DTM is a must for elevation based routing in SF as there are so many sizable buildings, large trees the other models fail hard. reply: Very Cool. Thanks for making it!I've checked out a few bike mapping apps recently and one feature I found absolutely invaluable was grade indications like http://hillmapper.com (which unfortunately seems unmaintained.)Google Maps elevation diagrams comparing the various selected routes are nice but they do not do a good job of highlighting extreme grades. I feel like you actually want to graph the change in elevation rather than the elevation itself.Using a route I cycled the other day (near Mission Bernal Safeway to near Diamond Heights Safeway). I think the routing may overweight bike lanes too much....

评论信号：It would be neat if it there were an option which added distance and elevation gain, but minimized the grade.For example, a route from SOMA to Nob Hill can be made flatter by approaching from either the east or west. Technically the "flattest" as measured by elevation gain between the two is straight up Taylor from Market, but it's much nicer to bike to the top of Nob Hill by going out of the way (e.g. Polk → California, Embarcadero → Broadway, etc.) reply: I suspect that this would result in absurdly long routes. Imagine trying to go up a mountain while minimizing grade... you'd end up zig-zagging back and forth endlessly.

评论信号：It's not accurate. I mapped it from my house on Cabrillo St in the outer Richmond to 4th avenue and it says I need to climb 25th avenue then walk Geary instead of correctly telling me to use 23rd avenue that is completely flat. reply: I think this comes down to the DEM they are using. I think it would have come out better if they had used USGS 3DEP DTM at 25cm.Edited: which apparently is hosted for free on AWS S3.

HN rank: 5

HN score: 135

comments: 30

original url: https://qlabs.sh/research/dust

评论信号：Every few years, a derivative-free neural network optimization algorithm gets some hype. I'd bet my life savings that none of them ever make an impact.Derivative-free optimization can be useful for genuinely discontinuous objectives [1], but common neural network objectives are smooth and/or Lipschitz.The gradient is useful. Instead of trying random directions and hoping that one of them is an improvement, it tells you where to go. The more parameters you have, the more useful it becomes.A strict complexity gap between gradient-based and derivative-free Lipschitz convex optimization has been suspected for decades and recently proved (using AI, [2]). Neural net optimization is nonconvex, but not radically different.IMO, a more promising direction is gradient-based optimizers specialized to the neural network structure, like Muon [3].[1] https://arxiv.org/abs/2202.00817[2] https://arxiv.org/abs/2607.13335[3] https://jeremybernste.in/writing/deriving-muon reply: Random selection seems to favor post traininghttps://arxiv.org/pdf/2603.12228

评论信号：> There are many interesting open questions. The first is whether, and how, Dust can find better directions than backprop’s first-order gradientBoth algorithms are bound by the same Pareto frontier based on the Empirical Risk Minimisation Principle, so they’re already on the same trajectory. Interestingly backprop is limited by conditioning of the Hessian matrix in order to converge (differentiate correctly). So removing this limitation is actually a great step. I’m excited to see a comeback of evolutionary methods because they’re much more general, albeit costly and naive. We’re now very close to what can be described best as brute forcing the Pareto frontier out of our datasets. Not sure that’s what we want but I have no better ideas either. reply: Pareto frontier the Pareto frontier and then with a little Pareto frontier you might get to the Pareto frontier

评论信号：It's computationally expensive and infeasible but what's the upside?Genuine question due to unfamiliarity with the subject.