Aug 16Monday, August 17, 2026 · all daysAug 18 · today »
1.Qwen 3.8 27B is excellent, but it defaults to overthinking things(simonwillison.net)
784 points by bilsbie 5 days ago | 376 comments | permalink
tl;dr: Qwen 3.8 27B is an Apache 2 licensed vision-capable model that runs impressively well from a 17GB file on consumer hardware, handling coding agents, bounding boxes, and tool calls competently. However, its default "xhigh" reasoning setting causes absurd overthinking—spending 21 minutes and 22K reasoning tokens on a simple SVG prompt—so users should dial reasoning down to low. The main drawback is speed (15-30 tokens/sec), though Multi-Token Prediction support via llama.cpp offers a ~72% boost.
HN Discussion:
  • Amazement that capable models now run on consumer/local hardware
  • Overthinking is a systemic issue from RL incentives across all current models
  • Users are building tools/forks to manually control and limit reasoning behavior
  • Current chain-of-thought reasoning approach is fundamentally a dead-end
  • Overthinking especially hurts dense models due to compounding speed penalties
2.A third world engineer responds to “RISC-V: They should have known better”(rvembedded.com)
605 points by Narishma 5 days ago | 321 comments | permalink
tl;dr: An embedded engineer in Trinidad rebuts Dmitry Grinberg's RISC-V critique, arguing that Grinberg himself derived that RV32EC is ideal for cheap microcontrollers, then complained the resulting chips exist. The author counters that RISC-V's real value isn't ISA elegance but accessibility: a single instruction set scales from 10-cent CH32V003 parts to MMU-equipped SoCs running Linux, with open toolchains and no licensing walls—unlike ARM's segmented Cortex-M/A product boundaries. For engineers outside the US/EU facing prohibitive shipping and licensing costs, RISC-V's price and openness matter more than encoding purity.
HN Discussion:
  • Author contradicts himself by citing high shipping costs while claiming 10-cent parts matter over $1 parts
  • ~The two articles talk past each other, addressing different scopes (embedded vs. general performance)
  • Shipping costs to developing nations are overstated; Asia-to-Africa/Asia routes are cheap
  • RISC-V's openness and pricing pressure on incumbents validates the accessibility argument
  • ~Both critiques are valid; RISC-V can be both an improvement and still have fixable flaws
3.Anthropic's ‘watermark’ text adulteration in Claude is a perversion of writing(daringfireball.net)
814 points by ropbear 5 days ago | 718 comments | permalink
tl;dr: Anthropic is rolling out text watermarking on all Claude models globally to comply with the EU AI Act, using a SynthID-style technique that biases token selection toward "green-listed" words detectable only with Anthropic's secret key. The author argues this deliberately degrades word choice—the essence of writing—for a scheme that harms honest users (who risk false accusations when using Claude for proofreading) while being trivially defeated by bad actors via paraphrasing tools. He contends Anthropic's claims of "imperceptible" quality impact are unfalsifiable marketing, and that applying an EU regulation worldwide reflects either technical incompetence or corporate cowardice.
HN Discussion:
  • Author misunderstands LLM sampling; randomness means there's no single 'best' token to preserve
  • Watermark detection requires sending text to providers, raising privacy and practical concerns
  • Gruber's anti-EU bias is driving a poorly-reasoned technical critique
  • ~Watermarking legitimately distorts the trained probability distribution even if no 'best' token exists
  • If word choice matters so much, users should write their own text instead of using LLMs
4.Claude: System Prompts(platform.claude.com)
743 points by tosh 5 days ago | 280 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Providing supplementary resources like git history and full prompt extractions for tracking changes
  • System prompts are too long and noisy, hurting model performance versus shorter guidance
  • Enforcing basic common sense via system prompts suggests Anthropic doesn't fully trust model intelligence
  • Heavy system prompts make models unhelpful; users prefer raw models or modular prompt choices
  • System prompts reveal interesting behavioral roadmap and safety layering strategy
5.The AI Credit Resale Economy(vectoral.com)
323 points by mlenhard 5 days ago | 128 comments | permalink
tl;dr: A gray market has emerged around reselling unused AI provider credits (Anthropic, OpenAI, etc.), with brokers offering discounts of 30-80% off list price via proxy endpoints, marketplaces like AI Credits and AICreditMart, and "bulk discount" routers like CheapCredits and Tokvana. One broker was offering $100k/day in spend, and the author estimates tens of millions of dollars in credits are circulating across marketplaces, Telegram channels, and Reddit. As providers grow cost-conscious, crackdowns on this arbitrage/abuse are likely coming.
HN Discussion:
  • Providers can easily trace and crack down on relay accounts, confirming the article's warning
  • ~The article misses the huge risk/opportunity of MITM proxies harvesting training data or exfiltrating secrets
  • ~This resale abuse pattern is unremarkable and mirrors decades-old loyalty/account abuse industries
  • Using these shady resellers is too risky given security and privacy concerns, regardless of discount
  • The research is shallow and misses much larger resale ecosystems on other platforms
6.Tell HN: Cloudflare silently injects its analytics when you switch nameservers
625 points by stagas 5 days ago | 193 comments | permalink
tl;dr: After switching nameservers to Cloudflare to enable R2 bucket serving, the author discovered Cloudflare automatically injected a JavaScript analytics snippet into their otherwise JS-free static site. Disabling it required manually adding the site to the Analytics dashboard and turning the snippet off. The author argues such tracking features should be opt-in rather than opt-out.
HN Discussion:
  • Users of free MITM proxy services should expect providers to inject content eventually
  • The injection only happens because the author enabled Cloudflare proxying, not just DNS
  • Cloudflare defends the default-on behavior as beneficial RUM data that's easy to disable
  • Users can mitigate script injection using CSP headers as a technical workaround
  • Shock and outrage that Cloudflare injects code, possibly raising legal/ethical concerns
7.Stripe will reportedly acquire OpenRouter for $7B+(techcrunch.com)
459 points by zacharyozer 5 days ago | 289 comments | permalink
tl;dr: Stripe has reportedly agreed to acquire OpenRouter for over $7 billion, according to Bloomberg. OpenRouter, which raised a $113M Series B in May at a $1.3B valuation, provides a unified API for accessing 400+ AI models, letting customers switch between them based on task and budget. The startup, often called "Stripe for AI," claims 8 million users worldwide.
HN Discussion:
  • Stripe's API expertise makes it a natural fit to abstract LLM routing like it did payments
  • The acquisition is primarily a defensive move to capture AI payment volume Stripe was losing
  • The $7B valuation seems disproportionately high for a thin API middleman
  • Concerns about privacy and customer outcomes post-acquisition, prompting search for alternatives
  • Stripe can leverage OpenRouter's data for routing optimization, fraud detection, and cross-business synergies
8.Firefox for iOS now has a native adblocker(support.mozilla.org)
701 points by pentagrama 5 days ago | 273 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ~Firefox iOS still lacks key privacy features like auto-clearing cookies that competitors have
  • uBlock Origin Lite for Safari is a superior alternative adblocker on iOS
  • ~This feature is redundant since Firefox Focus already offered system-wide ad blocking
  • Apple's WebView restrictions prevent alternative browsers from properly supporting extensions
  • Ad blocking should be default behavior in all modern browsers
9.Nvidia dramatically reduces amount of OpenAI infra financing it may guarantee(reuters.com)
245 points by root-parent 5 days ago | 151 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Headline is poorly worded and mangled
  • Deal was never actually signed, adding context about scale of infrastructure
  • Nvidia is becoming a risky financing entity beyond its chip business
  • ~The scaled-back backstop may still be a profitable arrangement for Nvidia
  • This reflects broader circular AI financing problems and bubble dynamics
10.The federal keyword lists that canceled billions in research funding(highereddive.com)
278 points by walrus01 5 days ago | 179 comments | permalink
tl;dr: Court documents from a University of California lawsuit reveal that federal agencies including NIH, NSF, DOD, and NEH used keyword searches—rather than individual review—to terminate research grants deemed misaligned with Trump administration priorities. NSF's list, drawn from a 2024 Ted Cruz report, flagged terms like "minority," "prejudice," and "historically," while DOD targeted climate-related terms and NIH cut grants referencing "structural racism" or containing diversity supplements. UC researchers argue the terminations violated their First Amendment rights and have won two preliminary injunctions blocking the cancellations.
HN Discussion:
  • Absurdity of the banned words shows the administration is anti-science and regressive
  • Researchers are self-censoring proposals to avoid keyword triggers, illustrating the article's harm
  • The cuts are hypocritical since deficit spending is still rising, undermining any fiscal justification
  • Free speech advocates are hypocritically silent about this government censorship
  • This resembles historical authoritarian ideological purges like Lysenkoism or cultural revolutions
11.Research papers using "kidney disappointment" instead of "kidney failure"(scholar.google.com)
429 points by Alifatisk 5 days ago | 152 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Shares similar amusing examples of AI or paraphrasing tool mistranslations in papers
  • Attributes tortured phrases to paraphrasing tools used to evade plagiarism detection
  • ~Suggests translation issues from non-native speakers rather than AI as the cause
  • Questions the AI/translation explanations given the phrase predates modern LLMs
  • Argues AI contributions should be cited like third parties to address plagiarism
12.The weekend is 100 years old(theguardian.com)
217 points by lentil_soup 5 days ago | 154 comments | permalink
tl;dr: The two-day weekend turns 100 this year, having been popularized by Henry Ford in 1926 (and Boots in the UK in 1933) as a productivity and consumer-capitalism play — while a Soviet experiment with staggered days off from 1929-1940 failed because shared time off is what makes leisure meaningful. Today the weekend is fraying: remote "time owners" have quietly turned Friday into a half-day ("Skiveday Friday"), while shift workers remain clock-bound, and smartphones let work bleed into Saturday and Sunday. The four-day week is gaining traction, with 56 of 61 UK pilot companies sticking with it.
HN Discussion:
  • Weekend conventions vary globally and have shifted recently in Gulf states
  • Questioning how to escape the industrial clock's rule over daily life
  • The week and other time units are recent social constructs, not natural
  • Excitement about the four-day workweek and shared time off mattering
  • Awe at how radically industrial-era life differs from pre-industrial life
13.Qwen 3.8 27B(huggingface.co)
1423 points by erdaltoprak 7 days ago | 790 comments | permalink
tl;dr: Qwen has released Qwen3.8-27B, a 27B-parameter dense vision-language model with a hybrid Gated DeltaNet/Gated Attention architecture, native 262K context (extensible to 1M via YaRN), and FP8 quantization. It features toggleable thinking mode with tunable reasoning_effort, preserved thinking across turns, and native image/video understanding. Benchmarks claim it outperforms prior Qwen models and competes with Opus 4.6 Max on coding (SWE-bench Pro, Terminal Bench) and agentic tasks, while trailing on some reasoning benchmarks like HLE.
HN Discussion:
  • ~Model successfully passes personal benchmarks but has efficiency tradeoffs like high VRAM and token usage
  • Impressed with quality of output for a locally-runnable model, praising specific creative generation results
  • Model overthinks and second-guesses, making it less practical than competing efficient models like Gemma
  • Benchmarks approaching frontier closed models suggests local models will soon match top-tier capability
  • Small dense open-weight models like this benefit the public most by being widely accessible
14.Asus Bike Booster(asus.com)
622 points by wiradikusuma 10 days ago | 447 comments | permalink
tl;dr: ASUS's Oxiis E250G1 is a friction-drive motor that clamps onto a conventional bike's seat post to convert it into an e-bike, delivering 250W rated (500W peak) power with a 158Wh removable battery good for ~50km in eco mode. It fits most bike types (16-29" wheels, tires up to 60mm, seat posts 25.4-34.9mm), weighs 3.7kg, charges in 2 hours via 100W USB-C PD, and offers three assist modes controllable via app.
HN Discussion:
  • Friction drives are inefficient, wear tires, and perform poorly in bad conditions
  • Product targets almost no one since bikes without fenders won't want this attached
  • This is not a novel idea; similar products existed decades ago (Sinclair, VéloSoleX)
  • Smartphone app control and limited waterproofing are serious usability drawbacks
  • USB-C PD charging is a genuinely great feature that other ebikes should adopt
15.Asynchronous I/O in DuckDB: Work, Thread, Work(duckdb.org)
280 points by pdet 11 days ago | 31 comments | permalink
tl;dr: DuckDB v2.0 (fall 2026) introduces asynchronous I/O for Parquet and CSV reads, using a separate ASYNC thread pool and a read-ahead queue with memory governance to keep remote fetches in flight while worker threads decode. Benchmarks on TPC-H Q6 against S3 show ~3x speedup for Parquet (up to 3.7x when tuned to saturate 25 Gbit/s), ~20x for CSV, and ~1.5x for cold local reads. Concurrent query workloads see CPU utilization jump from ~6 to ~48 cores on a 64-vCPU machine, with JSON and DuckDB-native format support planned next.
HN Discussion:
  • User confirms significant speedup with real-world benchmark on their own setup
  • ~Questions technical design choices like thread oversubscription and benchmark memory configuration
  • Praises DuckDB's trajectory toward becoming the fastest analytical query engine
  • Asks technical clarification questions about implementation details like CSV quoted newlines
  • Expresses enthusiasm that this long-awaited feature has finally arrived
16.Models Are Getting Dumber on Purpose(w4g1.dev)
329 points by hruvhwe 5 days ago | 187 comments | permalink
tl;dr: Newer models are shrinking active parameters while boosting reasoning benchmarks, but they're trading factual recall for procedural skill—Qwen3.5's small models hallucinate 80%+ of the time on knowledge questions, and even Gemini 2.5 Pro only hits 53% on SimpleQA. The bet is that facts (which rot and consume ~2 bits/parameter) belong in external retrieval systems, while compact reasoning procedures stay in weights. If it pans out, frontier-quality reasoning could run on a single 24GB consumer GPU, with hallucinations becoming traceable data bugs rather than unfixable weight errors.
HN Discussion:
  • Envisions modular, pluggable domain-specific knowledge modules composed with reasoning cores
  • Article is AI-generated and relies on outdated benchmarks/facts, undermining its credibility
  • Reasoning and factual knowledge cannot be cleanly separated since reasoning depends on grounded facts
  • Hallucination cannot be fully solved even with external retrieval, as models still fill gaps imaginatively
  • The vision is speculative sci-fi dreaming disconnected from practical constraints
17.Cultivating a state of mind where new ideas are born (2023)(henrikkarlsson.xyz)
273 points by felixbraun 6 days ago | 64 comments | permalink
tl;dr: Creative breakthroughs require a "solitary state of mind" where fragile early-stage ideas can survive without being killed by social pressure or peer judgment—a key reason Y Combinator avoids coworking spaces and why Grothendieck credits three years of isolated work for his mathematical originality. Sustaining this state is difficult and requires deliberate techniques: strict rituals (Bergman's daily workbook), working faster than self-censorship (Knausgaard), subverting one's own excitement (Tallest Man on Earth), or framing work as religious service. The ability to linger in confusion and surface original questions matters more than raw talent at answering them.
HN Discussion:
  • Personal experience confirms ideas are fragile and external validation nearly killed a promising project
  • Academic and collaborative environments can also produce great work, contradicting the solitude thesis
  • Ideas aren't fragile; creators just need conviction, and isolation isn't a prerequisite for creativity
  • ~Solitude must be balanced with collaboration with the right team or people
  • Solitude is really a state of mind of being unbothered by others' opinions, echoing the article
18.GLM-5.3: Frontier coding with emergent cyber capabilities(z.ai)
1164 points by pella 8 days ago | 578 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Hands-on user confirms strong capabilities including offensive security tasks
  • ~Skepticism about cyber capabilities disclosure and comparison to competitors' efforts
  • Impressed by honest, non-hyped tone and research-oriented writing style
  • Praise for Chinese labs closing the gap with closed frontier models at lower parameter counts
  • Argues closed labs should release cyber models to defenders to level the playing field
19.At-home test for infected ticks could improve Lyme Disease diagnosis(smithsonianmag.com)
296 points by gmays 6 days ago | 125 comments | permalink
tl;dr: LymeAlert, launching in August for around $50, is an at-home test kit that detects Borrelia burgdorferi in ticks removed from people or pets, delivering results in 15 minutes via a pregnancy-test-style strip and companion app. Developed by an MIT Sloan team led by pediatric PA Erin Dawicki, it aims to close the gap left by mail-in tests that miss the CDC's 72-hour prophylactic antibiotic window. The CDC doesn't recommend tick testing due to false positive/negative risks, and experts caution the test indicates tick infection—not human infection—since only 1-5% of tick bites result in Lyme.
HN Discussion:
  • Lateral flow test has inferior accuracy compared to PCR-based lab tests, claims unreviewed
  • Blood tests for humans would be more useful than testing the tick itself
  • Product will sell well given growing tick concerns among outdoor families
  • Prophylactic antibiotic advice is impractical in high-tick regions
  • Asking what viable alternatives exist if this test isn't the solution
20.Semaglutide linked to lower predicted dementia risk(alz-journals.onlinelibrary.wiley.com)
498 points by randycupertino 6 days ago | 393 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ~Questions whether benefits come from semaglutide itself or just weight loss/calorie reduction
  • Criticizes the study as industry-funded biomarker research that doesn't reflect real clinical outcomes
  • Enthusiastically endorses GLP-1 drugs based on personal or general health benefits
  • Accepts findings as good news given diabetes is a known dementia risk factor
  • Provides clarifying summary of what the study actually showed