HN rank: 1

HN score: 871

comments: 571

original url: https://huggingface.co/Qwen/Qwen3.8-27B-FP8

评论信号：Credit where it's due. Qwen 3.8 27B is only the second local model after Gemma 4 that managed to correctly reason through one of my private benchmarks. It took 5x as many tokens to do it and 12m30s with MTP enabled, but it did do it.Gemma 4 reasoned through it more implicitly, while Qwen 3.8 reasoned more explicitly. Laguna and Muse Glimmer failed hard on it, though they're useful for other tasks.The VRAM usage seems way less efficient than Gemma 4 or Glimmer though, with 32K of context taking 2.5GB of VRAM. With those, even with MTP or a DFlash model loaded, you could still fit 256k-768k of context. With Qwen 3.8 27B I can't even fit 128k if I quantize V to Q4_0. Maybe with some trial and error I can find some settings that perform well enough with a larger context window that it's still useful for longer tasks.Lots more testing to do, though I was getting some decent results out of Muse Glimmer which was more than twice as fast and supported huge context windows, managing to solve some bugs that Gemma 4 struggled with.... reply:...

评论信号：Absolutely the best pelican I've seen from a model that runs on my laptop: https://tools.simonwillison.net/markdown-svg-renderer#url=ht...Bicycle is the right shape. Pelican beak is excellent. Nice background. Most importantly, the pelican has one leg on each side of the bicycle - that's very rare.(No chain on this bicycle though - in the reasoning trace it says "already chainstay... skip chain detail; maybe a small chainring.")I ran that on an M5 Max MacBook Pro using LM Studio and their 17GB GGUF: https://lmstudio.ai/models/qwen3.8It took 21 minutes(!) and used 22,276 reasoning tokens to produce 3,223 tokens of output.(For the "they're training on your benchmark now" crowd, all of that cheating didn't prevent it from spending 20 minutes thinking about the task first! You can see the reasoning trace in the link I shared.)For comparison, here's one I got from qwen3.8-2.4t-a95b on OpenRouter, which is pleasingly animated: https://tools.simonwillison.net/markdown-svg-renderer#url=ht... reply: For anyone who followed yesterday's Gemini 3.7 Flash pelican which rendered in Safari but not in Firefox or Chrome... https://news.ycombinator.com/item?id=49289112#49290012....

评论信号：There's a real change (compared to 3.6) in the way it writes in thinking — it drops words like "to" and "we" in "We need to", talks generally in note form, drops the/and all over the place, avoids "for"."Need be helpful concise", "Need maybe not overdo", "Need ask!" Almost caveman.I have an (unsourced, vague) suspicion that this rather unique thinking trace pattern is actually hobbling the MTP predictions, which seem to perform poorly.Other notes: it uses the trick of repeating the prompt in the thinking trace.It also worries about hidden chain of thought appearing in the final answer. It talks about "desired oververbosity 9", which is new. A bit GPT-ish.It is being extraordinarily thorough in thinking through one of my code requests, but I don't know if the net result will be any better than the 35B MoE.I asked it to ask me clarifying questions — it did, and it offered me a list of defaults I could simply agree to.I don't think it is necessarily overthinking in the looping sense, but it is in the being exhaustive sense. I need to explore how it does with a tighter reasoning budget.I am impressed but I am definitely in Camp Please-35B-A3B-When?... reply:...

HN rank: 2

HN score: 171

comments: 109

original url: https://blog.cryptographyengineering.com/2026/08/14/everything-is-about-to-go-dark/

评论信号：It's worth realizing that, before computerized central offices, telephone wiretapping required running physical wires. Back when Rudi Giuliani was prosecuting organized time, not only did physical wires have to be run, the cops were billed for them as expensive private lines. His task force was spending about a million dollars a year with New York Telephone on wiretapping. In one case, law enforcement didn't pay their bill, resulting in the person being wiretapped having the wiretap connection show up on their bill, blowing the case.That resulted in the Communications Assistance to Law Enforcement Act, which mandated that central offices offer remote wiretapping. Capacity up to 1% of lines is required.Back in the electromechanical era, the only call data that could be collected was outgoing dial pulses, using a "pen register".[1] (The one shown in Wikipedia is mine. It's a beautiful piece of antique brass telegraph technology. It records dial pulses as dashes, and has to be wound up like a clock, with a big brass key.) The Supreme Court decision allowing "pen registers" without a warrant refers to these "extremely limited" devices.... reply:...

评论信号：> In the real world, it does feel likely that we’re going to hit some sort of a ceiling on the number of useful bugs, and probably we’ll hit it soon.This doesn't resonate with me. I see companies adding more sloppily written features with AI. I see more bugs in the software I use, not less. While it's plausible that software is getting both buggier and more secure, I suspect those two move in the same direction not opposite.My guess is that we're getting better at finding existing security issues with AI (and thus fixing those issues), but simultaneously adding more insecure surface areas at a faster rate. reply: One way to resolve the tension here is to note that CNE and lawful-intercept access to phones depends generally on platform vulnerabilities, not application code vulnerabilities....

评论信号：I've always loved the ridiculousness of the "going dark" label when law enforcement can't access encrypted chats or a back door isn't built into a piece of software. When there are security cameras on the vast majority of houses, stop lights and in people's hands, and when so much meta data about people's associations are shared from Google, Facebook, any other social platform, how in the world can they say they are "going dark". How did they ever solve crimes before these things?

HN rank: 3

HN score: 100

comments: 51

original url: https://dmitry.gr/?r=06.%20Thoughts&proj=12.%20RV

评论信号：RISC-V is... fine. It satisfies my two requirements for an ISA as a hobby CPU designer, which are:1. Supported in mainline LLVM and GCC.2. I can implement it without lawyers sending me a love letter.Everything else, I can fix in post. There are enough good ideas spread across the extensions that I can assemble a reasonably put-together, curated embedded ISA with competitive performance and code density that admits a simple implementation.I think Dmitry's points are largely on-target, though I have filed my usual statutory complaint that every rant that includes a bitfield diagram for the RISC-V J format should accompany it with a similar diagram for the Arm T32 BL encoding. reply: > RISC-V is... fineExactly.> It satisfies my two requirements for an ISA as a hobby CPU designer...You probably have some unstated requirements as well, such as available toolchains and "vetted well enough to actually be able to run code."Risc-V now occupies the Schelling point for people who, for whatever reason (rent-seeking and security top the list) want to leave the x86 and Arm ecosystems.

评论信号：the significance and allure of risc-v, the reason china is investing heavily in it right now, has little to do with the technical details of how it works under the hood, it's the fact that it is an open standard not encumbered by intellectual property law. even if it isn't technically the best general-purpose processor architecture, it sets an important precedent by proving that it is possible to develop an open public architecture that the world can use to build computing devices without being extorted by a multinational corporation charging licensing fees or a geopolitical superpower enacting tariffs and sanctions. reply: > it's the fact that it is an open standard not encumbered by intellectual property law.There are actually many of those. But Risc-V has become, through effective marketing, the Schelling point for anybody who wants to avoid the x86 and Arm ecosystems, both for the rent-seeking behaviors you mention, and also, in some instances, for security reasons.And, as others have mentioned, the ISA doesn't really matter....

评论信号：It's basically MIPS all over againThe conclusion is honest, and you can of course brute force any ISA into any role. I used to loathe x86 for that reason, but now that I'm older I respect the game. reply: I think MIPS is a great example, and even there I don't think there's the bizarre bifurcation of ISA options RISC-V brings to the table.As a fellow olderster, I can't help but think that after almost 50 years of "ISA X is sooooo much better than x86 it's obvious ISA X is the future and x86 will be dead Real Soon Now (for whatever todays version of x86 is)" I can only shake my head ruefully and say "ping me when that happens".Controversial Take (that history proves isn't): Software matters; ISAs don't.

HN rank: 4

HN score: 22

comments: 9

original url: https://www.economist.com/by-invitation/2026/08/13/the-case-for-overhauling-american-science

评论信号：This is the full proposal: https://www.whitehouse.gov/wp-content/uploads/2026/07/Scienc...It's pretty thoughtful about diagnosing the problems of the current system, but I don't know about the solutions. Disbursing money to researchers directly (or via industry) seems captive to the same incentives as disbursing money via universities.

评论信号：Achieve link (https://archive.is/20260813145232/https://www.economist.com/...)Maybe I read it wrong but it seems the case being made is to overhaul the current NSF...with a different NSF? Like I'm failing to see what specifically the current government is doing to hamper AI progress that there should be a "focus on harnessing AI and out-competing China", and what harnessing AI actually means. reply: It means funneling public funds to AI and AI-adjacent companies that grease the wheels enough with the administration. It's like everything else: the administration decides it can control funds however it wants, regardless of what was legislated, and leverages those funds for quid pro quo transactions.Say what you will about how to improve academics and scientific research, but the current administration's modus operandi is not the way.

评论信号：It's a surprisingly short article https://archive.ph/DhZkR

HN rank: 5

HN score: 767

comments: 701

original url: https://mun-logadan.github.io/why-does-opus-5-feel-worse/

评论信号：The single biggest annoyance with Opus 5 is that it writes too elliptically.Sentences that orbit a point, then jump to it like it's a revealed insight.Unnecessarily abstract phraseology. Constantly using inanimate nouns as the subjects in sentences in order to unlock variety in verb choice, especially when it helps construct a sentence where the real action can 'land' like a surprise at the end.It is definitely more capable, and yes, I've found it can make unwarranted decisions, but actually I've found Fable worse for that, particularly if it's off in a subagent somewhere out of sight.And comments are out of control. I have a subsystem in my hobby app that I wrote over a couple of weekends with Opus + Fable. After ~30 or so commits it apparently started instructing subagents to copy the "existing verbose comment style of the codebase" - a verbose style it initiated. A review of the code showed it was approaching 3:1 comments to code ratio. I spent a day's worth of tokens (5x) rephrasing and eliminating comments. reply: Everything that claude writes fits into the same aesthetic structure. The aesthetic is that of an expert slowly revealing an insight to the user....

评论信号：I’ve been doing some heavy work on a personal project lately. I burned through the limits on Claude, the plus a few hundred dollars in credits, and ultimately decided to move to an OpenAI account just so I can keep going.I was surprised to find that OpenAI Sol is much much nicer to work with than Opus 5 or Fable at the moment. Especially on Opus 5, the way it communicates is just exhausting. It keeps “being honest” and “confessing” mistakes and just generally talking a lot. I felt like I had to really dig to see what it’s doing.The project involves OCR, and despite repeated instructions not to, both Claude models keep spinning out a bunch of agents to re-invent the OCR setup, and they inevitably seem to invent a primitive serial version that takes 20x the time, or longer, to complete, and then running it against thousands of docs. Basically I have to watch it like a hawk or it just spins out on red-teaming tasks that take hours and hours.I don’t know what its system prompt is, but Sol/Codex is just so much nicer to talk to. It only asks exactly what’s needed, it tells me only what I need to know, and it is just generally workmanlike.... reply:...

评论信号：I'm with the author and others in this comment thread, speculating that effectively the balance has tipped to where humans are no longer the target audience of post training - other agents are. Whether it's through the reasoning / CoT, or whether it's in handing off to subagents etc, the focus has moved to agents communicating in "agent-speak" to themselves or other agents. And human niceties are just kind of, noise in the way of getting work done.This will probably bring us to a cross roads where the folks that want to remain in oversight and control of the AI work will bifurcate from those that want to skate straight to the future where nobody looks at anything and outcomes are evaluated purely empirically. reply: Every round of models (plus all the secret tweaks) require new strategies to stay afloat as a human. My new tactic for Fable and Opus is to give them a line limit, both during planning and code creation. It os amazing how well that works for keeping them on task and avoiding premature optimization, pointless tests or any of those "robustness" ideas that are not planned or asked for.