Sep 9Thursday, September 10, 2026 · all days
1.DeepSeek v4.1 Flash(twitter.com)
747 points by Liwink 10 hours ago | 399 comments | permalink
tl;dr: DeepSeek has announced DeepSeek-V4.1-Flash, the smallest model in a new architecture family featuring native visual understanding. The company claims improvements in capability, inference speed, and throughput, with the architecture designed to scale to larger models.
HN Discussion:
  • DeepSeek's transparency and technical detail is refreshing compared to safety-focused competitors
  • Admiration for DeepSeek's fearless innovation and clever research ideas at frontier scale
  • Model size has grown too large to reasonably run locally, undermining the 'flash' label
  • Pricing has increased and may not be justified by performance versus competitors
  • Practical strengths noted: strong performance, fewer refusals, and useful as a backup model
2.iPhone Duo(apple.com)
1336 points by thecosmicfrog 22 hours ago | 2331 comments | permalink
tl;dr: Apple's iPhone Duo is the company's first foldable iPhone, featuring a 7.6" inner Super Retina XDR display with nano-texture finish and a 5.4" outer display sharing the same aspect ratio, powered by the A20 Pro chip with a dual-battery system delivering up to 31 hours of video playback. It includes a 48MP dual Fusion camera system with foldable-specific features like Smart Take, Duo Preview, and Duo FaceTime, uses Touch ID in the side button, and is eSIM-only. Available October 23, 2026, starting at 256GB with titanium construction and IP68 rating.
HN Discussion:
  • Enthusiastic about the Duo's design and Apple's new keynote direction under Ternus
  • ~Cautiously optimistic but waiting for later versions before adopting a foldable
  • Foldables are a compromised form factor that serves no clear purpose
  • Price is too high to justify, especially for a first-generation Apple product
  • Hopeful the Duo will push developers to properly support foldable app layouts
3.Show HN: What if the speed of light was 5 km/h?(rivendell.dmitrybrant.com)
523 points by dmitrybrant 14 hours ago | 222 comments | permalink
tl;dr: An interactive browser demo simulates walking through a park where the speed of light is set to 5 km/h, making relativistic effects visible at ordinary walking speeds. Users can move with WASD/arrow keys and toggle effects like length contraction, time dilation, Terrell rotation, and Doppler shifting (blueshift ahead, redshift behind).
HN Discussion:
  • Praises the demo as more accurate than the prior MIT slower-speed-of-light game
  • Reflects on how slow light actually is at cosmic scales, reinforcing the demo's premise
  • ~Questions whether the simulation correctly models Thomas rotation from non-parallel boosts
  • Points to prior/similar works like Mr. Tompkins, MIT's game, and Velocity Raptor
  • Criticizes lack of mobile support and controls
4.Shopify acquires Tailwind(tailwindcss.com)
1108 points by EdwinHoksberg 1 day ago | 434 comments | permalink
tl;dr: Tailwind CSS is being acquired by Shopify, which has been a major early adopter and heavy user of the framework. The open-source projects remain MIT-licensed and actively maintained, but Tailwind is winding down its commercial business—existing Tailwind Plus and ui.sh customers keep access, but new signups are closed as the team refocuses on Tailwind CSS within Shopify.
HN Discussion:
  • AI's impact on dev tool business models made this acquisition inevitable
  • ~Tailwind is unnecessary; vanilla CSS or simpler alternatives work better
  • Well-wishes and gratitude to the Tailwind team for their contributions
  • Questioning Shopify's strategic rationale for acquiring dev tools
  • DevTools companies must offer scale services to survive the LLM era
5.What do Visa and Mastercard do? An intro to card networks(tautology.town)
630 points by evakhoury 1 day ago | 378 comments | permalink
tl;dr: Visa and Mastercard operate card networks with four core functions: routing transaction messages between issuers and acquirers via their telecom infrastructure, coordinating banking relationships to settle payments (including across currencies), setting interchange fees and incentives that shape network behavior, and enforcing rules plus arbitrating disputes. Notably, in a typical 2.5% US merchant fee, the issuing bank pockets ~2% while the network itself takes only ~0.15%—which is why issuers fund lavish rewards programs. EU regulation caps interchange at 0.3%, explaining the region's weaker rewards culture and stronger alternative payment adoption.
HN Discussion:
  • Credit card fees are excessive and the rewards-funded model is problematic for merchants and consumers
  • ~Card networks are outdated legacy middlemen being surpassed by app-based payment systems elsewhere
  • Card networks provide remarkable value by enabling global trust and low-friction transactions across jurisdictions
  • Firsthand industry experience confirms the article's summary is accurate
  • Practical observations and questions about card system mechanics (data sharing, name verification, dispute leverage)
6.Growing proof that autonomous cars save lives(spectrum.ieee.org)
424 points by bookofjoe 23 hours ago | 744 comments | permalink
tl;dr: IIHS and Waymo studies show autonomous vehicles crash significantly less than human drivers, with Waymo robotaxis involved in 68% fewer crashes overall and 81% fewer injury crashes per mile across four U.S. cities. ADAS features like automatic emergency braking are also cutting pedestrian crashes by 27% and rear-end crashes by 50%. However, the AV industry still lacks national safety standards and consistent crash reporting, prompting NHTSA to partner with SAE on a three-year effort to develop performance benchmarks.
HN Discussion:
  • ~Public transit, biking, and walking should be prioritized over autonomous cars for safety
  • Autonomous cars only need to be better than humans, not perfect, and have reached that bar
  • Waymo's statistics are misleading by comparing to average drivers rather than rideshare drivers
  • ~ADAS assistance is sufficient; full self-driving isn't necessary to save lives
  • Economic incentives like insurance costs will naturally push adoption of autonomous vehicles
7.No Man's Sky Cosmos(nomanssky.com)
432 points by Limb 1 day ago | 437 comments | permalink
tl;dr: No Man's Sky's 10-year anniversary update, Cosmos (7.0), overhauls deep-space exploration with asteroid belts, derelict hulks, outposts, reachable stars, and an interactive star system map, plus expanded six-degrees-of-freedom spacewalking and a Corvette Tractor Beam. Players can now become Space Station directors—customizing interiors, exteriors, and adding rooms—and form or join galactic alliances with leaderboards. The update also adds orbital base construction, a nostalgia-themed "Our Journey Continues" expedition with developer commentary from Sean Murray, and significant rendering upgrades including DLSS 4.5, XeSS 3, and improved terrain tessellation.
HN Discussion:
  • NMS feels like an impressive tech demo lacking substantive gameplay or soul
  • Hello Games deserves celebration for their post-launch turnaround and continued free updates
  • The game has abundant content and variety for those who engage with it
  • ~NMS is a chill, cozy niche game misjudged by players expecting stakes and challenge
  • Trailers overhype and misrepresent the actual in-game experience
8.GPT-6 Astra, looped transformers, and hidden reasoning(magazine.sebastianraschka.com)
493 points by ModelForge 1 day ago | 158 comments | permalink
tl;dr: OpenAI's rumored GPT-6 Astra reportedly uses "looped transformers" (recurrent depth), an architectural technique where intermediate representations pass through the same transformer blocks multiple times, effectively increasing depth without adding parameters—an approach seen in models like Nanbeige, Ouro, and Universal Transformers. The author argues this looping likely isn't the reason Astra produces shorter/less monitorable reasoning traces; more capable models simply need less scratchpad, similar to how larger models in a family use fewer tokens than smaller siblings. Astra's biggest practical leap appears to be in computer-use tasks, likely trained via RL on OpenAI's reported fleet of Mac Minis.
HN Discussion:
  • Provides supporting academic references on CoT computational requirements and prior work on universal transformers
  • Questions whether looped transformers inherently hide reasoning and asks about labs using latent recursive reasoning
  • Argues that looping a full transformer is by-definition hidden reasoning, pushing back on the article's framing
  • Praises the article's clarity and endorses the author as a trusted source on LLM internals
  • Hands-on experimentation with looped transformer models confirms the article's claims about compute intensity
9.AirPods 5(apple.com)
499 points by awad 23 hours ago | 437 comments | permalink
tl;dr: Apple announced AirPods 5, priced at $129 ($149 with wireless charging case), featuring open-ear Active Noise Cancellation that reduces external noise up to 50% more than AirPods 4, a new multiport acoustic architecture, and improved Adaptive EQ. They integrate with Siri AI (rolling out in iOS 27 beta) for hands-free personal context queries, head gesture responses, and Live Translation. Pre-orders start today with in-store availability September 18.
HN Discussion:
  • Wireless Bluetooth audio quality is significantly worse than older wired options
  • Apple's ecosystem lock-in punishes non-iPhone users who buy AirPods
  • Apple is marketing existing/basic features as new innovations
  • Competitors like Huawei and Sony offer better products at lower prices
  • The $129 pricing is surprisingly good value for the features offered
10.iPhone 18 Pro and iPhone 18 Pro Max(apple.com)
398 points by meetpateltech 23 hours ago | 462 comments | permalink
tl;dr: Apple announced the iPhone 18 Pro and Pro Max, featuring a 48MP Fusion Main camera with variable aperture, the A20 Pro chip with a new vapor chamber for improved sustained performance, and a smaller Dynamic Island. The Pro Max delivers Apple's largest-ever battery life increase, with eSIM-only models gaining extra capacity. Starting at $1,199 and $1,299 respectively, they ship September 18 with iOS 27, Apple Intelligence, and a beta Siri AI (not available in the EU at launch).
HN Discussion:
  • Excitement about the new authenticity/signed image feature for verifying real photos
  • Enthusiasm for the technical upgrades like 2nm chip, vapor chamber, and improved battery
  • ~Desire for more genuinely 'pro' features like Thunderbolt, multiple eSIMs, or PCIE
  • Underwhelmed by marginal improvements that don't justify the high price
  • Concern about missing specs like RAM and memory bandwidth in the announcement
11.Automattic's board forces CEO Matt Mullenweg into leave of absence(techcrunch.com)
406 points by LeoPanthera 16 hours ago | 279 comments | permalink
tl;dr: Automattic's board voted to place CEO Matt Mullenweg on paid leave against his will, installing CFO Mark Davies as interim CEO; Mullenweg says he was given the resolution 50 minutes before the meeting and denied time for legal review. The move follows a turbulent stretch including the ongoing WP Engine lawsuit, mass severance-driven departures in 2024, and a 16% layoff in April 2025. Mullenweg remains on the board and continues to lead the WordPress.org open-source project.
HN Discussion:
  • Removing Mullenweg is overdue given his harmful behavior and unforced errors as CEO
  • This is a major, difficult decision that will likely trigger retaliation from Mullenweg
  • Mullenweg needs a break or professional help after 23 years leading WordPress
  • The WordPress/Automattic community and industry itself is inherently dramatic and gossipy
  • Publicly accusing board members of conspiracy signals Mullenweg won't be returning
12.Desert Ant Labs: local, fast models that run on device(desertant.com)
477 points by willwhitedc 1 day ago | 99 comments | permalink
tl;dr: Desert Ant Labs, a European AI startup, launched 18 small, specialized on-device models for audio, vision, and text tasks—accessible via a single SDK for Swift, Kotlin, and JavaScript, and free up to 100k monthly active devices. Their models include Voz (transcription 4.7x faster than Whisper), Clear (9MB audio enhancement), Redact (PII masking in 27 languages), and Tongue (2MB language ID), all claiming to outperform much larger cloud and local alternatives on speed and often quality. The team previously built the Detail video app and trained these models to replace their own cloud API dependencies.
HN Discussion:
  • Enthusiastic about on-device small models as a powerful alternative to cloud LLMs
  • ~Questions the business model viability of free local models with no per-call billing
  • Disappointed by lack of Python SDK and web/cross-platform support limiting usefulness
  • Skeptical of marketing claims, noting Voz is just repackaged Parakeet v3 and demos are unconvincing
  • ~Discounts the offering due to LLM-generated marketing copy despite liking the concept
13.Qwen 3.8 follows GPT-5.5 Pro reasoning prefills(gist.github.com)
232 points by wsxiaoys 23 hours ago | 90 comments | permalink
tl;dr: When prefilled with the first 1% of GPT-5.5 Pro's reasoning, Qwen3.8 A95B's output overlap with the teacher's answer jumped +18.18 points (including +27 on STEM), while DeepSeek V4 Flash, Inkling, and Kimi K3 barely budged. Since Qwen showed little response to Opus 4.8 prefills in a prior experiment, the author suggests Qwen may have been trained on outputs from GPT-5.5 Pro or a closely related model.
HN Discussion:
  • Explains and contextualizes the methodology, noting the authors' prior work on CoT recovery
  • ~Questions the methodology since the only available GPT-5.5 traces come from the stolen-thoughts leak that Qwen could have trained on
  • Proposes alternative explanation: both models may have trained on the same benchmark solutions rather than one distilling from the other
  • ~Argues the result doesn't justify strong conclusions about distillation extent or nature, and defends training on model outputs
  • Directly challenges the logical leap from the data to the conclusion about GPT-lineage training
14.Apple Watch Series 12(apple.com)
268 points by Lealen 22 hours ago | 358 comments | permalink
tl;dr: Apple announced the Apple Watch Series 12, featuring a new S11 chip, an upgraded Health Sensing System with more accurate heart rate and HRV tracking, and a readiness score that feeds into a redesigned Health app with a "Longevity" tab and Health Age metric. New Audio Intelligence features (coming later in 2026) include Sound Recognition, Live Rewind (replays the last 15 seconds of conversation as text), and Siri Recap for conversation summaries, plus Siri AI powered by Apple Intelligence on the wrist. It starts at $399, with pre-orders today and availability September 18.
HN Discussion:
  • Always-listening audio features raise serious privacy and legal concerns
  • Apple Watch updates show diminishing returns with no meaningful new features
  • Battery life matters more than smart features, driving users to alternatives like Garmin or Fitbit
  • Apple's environmental commitment is undermined by dropping support for older watches
  • ~The Watch's real value is passive health tracking, not active features
15.How I advertise malicious software on Google Ads(xlii.space)
423 points by xlii 1 day ago | 256 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Google's automated systems make arbitrary decisions with no human recourse, harming legitimate users
  • Google Ads is full of actual scams and malicious content that go unpoliced
  • Public complaints on social media are the only effective appeal mechanism against big tech gatekeepers
  • ~The suspension may have a legitimate technical cause like domain blacklisting or flagged terminology
  • Sharing similar personal experiences of unjust Google account suspensions
16.GNU Radio in the browser(gnuradioworld.com)
214 points by kristianpaul 1 day ago | 27 comments | permalink
tl;dr: GNU Radio Companion has been ported to run entirely in the browser, with the DSP stack and Qt GUI sinks compiled to WebAssembly. It reads/writes standard .grc files, renders live spectrum, waterfall, and constellation plots, and requires no Python, install, or server. It can also communicate with RTL-SDR, PlutoSDR, and HackRF devices via WebUSB.
HN Discussion:
  • Enthusiastic praise for the project as cool and useful
  • Sharing related personal projects using WebUSB/WASM for RF work
  • Criticism of the demo's poor onboarding and unclear documentation
  • ~Reflection on GNU Radio's steep learning curve, willing to retry
  • Appreciation for the interactive GUI reminiscent of MaxMSP
17.OpenAI might have stolen another major proof(twitter.com)
283 points by tamnd 12 hours ago | 10 comments | permalink
tl;dr: Mathematician Andreas Thom alleges OpenAI may have trained its Astra model on private conversations between him and Gábor Kun regarding Gromov's soficity conjecture, which OpenAI later claimed Astra had solved independently. Combined with similar accusations from Levent Alpöge and Tristan Buckmaster, the author argues this suggests a pattern of OpenAI passing off unpublished human mathematical work as AI-generated breakthroughs.
HN Discussion:
  • Pattern of behavior suggests OpenAI likely used private data as alleged
  • Big AI labs are built on IP theft and shouldn't be trusted with private data
  • Claims are weak with no credible evidence, just accusations on social media
  • Researchers should timestamp and publicly document work to protect against AI training theft
  • Thom's own prior public post about the proof undermines his plagiarism claim
18.What will our economic future look like?(anthropic.com)
224 points by oumua_don17 1 day ago | 437 comments | permalink
tl;dr: Anthropic's Economics team built a scenario explorer modeling how AI could reshape the US economy by 2030, based on user inputs about AI capabilities, adoption, and productivity. In modest-to-substantial scenarios, GDP grows 1.6-8.3% with unemployment staying within historical norms, while an "extreme" scenario shows 32% GDP growth but significant knowledge-worker displacement, stagnant or declining wages for those workers, and a shift of economic gains from labor to capital. The model omits factors like robotics, policy responses, and demand shocks, and is presented as a thinking tool rather than a forecast.
HN Discussion:
  • Productivity gains from AI will lead to staff cuts, not better working conditions as the article optimistically implies
  • The model omits crucial negative scenarios like societal harm, economic crisis, or AI being overhyped
  • ~The article's own finding on capital vs labor share confirms troubling wealth concentration
  • The scenario-building is overwrought when existing redistribution theory already addresses the problem
  • The article's uncertainty and hedging make the exercise essentially useless as guidance
19.Muse – Meta’s personal AI agent(ai.meta.com)
648 points by yks 1 day ago | 730 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Meta is targeting mainstream/normie users who don't follow AI closely
  • Sharing technical details about Meta's prompt injection defenses
  • Personal AI assistants lack meaningful use cases and add complication
  • Trusting Meta with deeply personal data is foolish given their track record
  • The product is poorly marketed and targets anxiety-inducing domains people won't trust
20.We built our house for LAN parties (2024)(lanparty.house)
587 points by fittingopposite 4 days ago | 364 comments | permalink
tl;dr: Cloudflare engineer Kenton Varda and Jade Wang built a custom house in Austin featuring 12 wall-mounted PCs in the basement plus 6 more in the office, all fold-away, with 20 identical netbooted machines (i5-13600KF, RTX 4070, 32GB RAM) running from a shared disk image in a dedicated server room with its own A/C. Additional features include DDR pads in the floor, extensive Cat6 wiring with UniFi networking, fiber-optic DisplayPort/USB runs up to 100ft to the stations, cat doors throughout, and a roof deck. It's a follow-up to Varda's 2011 Palo Alto LAN house, scaled up for family life.
HN Discussion:
  • Wholesome admiration for the builders enjoying life and their friend group
  • Nostalgia for LAN parties and suggestions for modern equivalents
  • ~Constructive suggestions for improvement (external blinds for energy, decor for coziness)
  • Feels the showcase resembles wealth-bragging like MTV Cribs
  • Curious about technical details like long-distance HDMI/USB cabling