← Oct 8Friday, October 9, 2026 · all daysOct 10 → · today »
1.Why isn't the industry freaking out about DeepSeek 4.1 Flash?(dgt.is)
1112 points by jonotime 6 days ago | 977 comments | permalink
tl;dr: DeepSeek 4.1 Flash delivers near-frontier performance (comparable to Claude Opus in daily coding use) at a fraction of the cost, enabled by a ~437x KV cache reduction that keeps all-day sessions under a dollar. The author argues Chinese labs are only a month or two behind OpenAI/Anthropic but operate on a fundamentally different cost structure—akin to generic drug makers skipping R&D—which threatens frontier labs' business model. Self-hosting still isn't economical, but the cost and efficiency gains make chasing top-tier models increasingly pointless for most workloads.
HN Discussion:
  • ↓Subsidized subscription pricing from frontier labs makes DeepSeek's API cost advantage irrelevant for most users
  • ↑DeepSeek 4.1 Flash delivers real cost savings and near-frontier quality in daily coding workflows
  • ↑Flash works great as a default driver with specialized subagents, significantly reducing bills
  • ↓Open models are actually 6-12 months behind frontier, not a month or two as claimed
  • ~Enterprise marketing push and anti-China sentiment prevent adoption regardless of cost advantages
2.Whistle: Speech to Text in 16.9 MB(cactuscompute.com)
946 points by gmays 5 days ago | 181 comments | permalink
tl;dr: Cactus released Whistle, a 16.9 MB on-device speech recognition model that transcribes seven languages, runs on CPU with no dependencies, and shares its C++ engine with their Needle LLM so one binary can turn audio directly into tool calls. Benchmarks claim 11ms time-to-first-token and 1,319 tokens/sec decode on an M4 Pro, beating Whisper base on several WER tests at ~1/8th the size, though Whisper leads on TED-LIUM, AMI, and MLS. Weights are on Hugging Face with prebuilt binaries for 17 targets including mobile, embedded, and WASM.
HN Discussion:
  • ↓Accuracy is poor compared to alternatives like Qwen ASR, with high error rates in real use
  • ↓Model hallucinates words and gets stuck repeating defaults like 'Thank you.'
  • ↑Works impressively well and accurately for clear English speech
  • ↓Missing essential features like streaming transcription during recording
  • •Curious how it compares to existing lightweight alternatives like Parakeet
3.Man discovers his parents' coffee machine used 1TB of data in 10 days(dexerto.com)
944 points by ck2 6 days ago | 570 comments | permalink
tl;dr: An IT worker discovered his parents' Keurig smart coffee maker had generated 1TB of network traffic in 10 days, saturating their access point. On closer inspection, most of the traffic stayed on the local network rather than going out to the internet, suggesting a bug rather than data exfiltration. He unplugged the device and plans to replace it, citing it as a reason to avoid unnecessary IoT devices.
HN Discussion:
  • ↑Confirms the data usage is intentional surveillance by Keurig for advertising purposes
  • ↓Skeptical of the claim because UniFi dashboards are known to miscalculate bandwidth
  • ↑Questions why a coffee maker needs internet connectivity at all
  • ↑Highlights that avoiding IoT risks requires technical skills normal users lack
  • •Suggests proactive countermeasures like poisoning data collection with spoofed devices
4.I hired an illustrator to draw my house. Now it's my Home Assistant dashboard(antonfrolov.substack.com)
921 points by soheilpro 7 days ago | 202 comments | permalink
tl;dr: The author commissioned an illustrator to draw a custom day/night illustration of their house, with each smart device on its own animated WebP layer, then wired it into Home Assistant using the stock picture-elements card. Device states swap between still and animated frames via state_image, a sunset/sunrise automation toggles day/night overlays, and YAML mode plus card-mod enables reusing floor layouts across per-floor and combined views. The dashboard runs on a TV via the HAOS Kiosk Display add-on through HDMI, finally getting the whole family to engage with their smart home.
HN Discussion:
  • ↑Applauds hiring a human illustrator over using AI for the artwork
  • ~The illustrator deserved more prominent credit in the write-up
  • ↓Questions the value of smart home dashboards when automation should handle everything
  • ~Not interested in smart home tinkering generally, though acknowledges the project is cool
  • ↓Critiques specific automation choices like turning off AC when away as inefficient
5.Theranos.world(theranos.world)
553 points by kbyatnal 5 days ago | 199 comments | permalink
tl;dr: Summary not available.
HN Discussion:
  • ↑Admiration for whistleblower Tyler Schultz's bravery against Theranos insiders
  • ↑Nostalgia for the old-school interactive promotional web design aesthetic
  • ↑Noting telling details within the site content exposing Holmes's self-promotion and fraud
  • •Criticism of Holmes's reduced prison sentence as unjust leniency
  • •Observation that older MacOS/iOS UI design was superior to current versions
6.Yes, and(htmx.org)
742 points by Michelangelo11 5 days ago | 294 comments | permalink
tl;dr: A CS professor argues programming remains a viable career despite AI, but juniors must resist using LLMs as code generators and instead write code themselves to develop real understanding—otherwise they'll fall into a "Sorcerer's Apprentice Trap" of unmaintainable systems. He recommends using AI as a TA for concepts rather than a solution-generator, and suggests programmers invest in communication, business understanding, and architecture skills. For job-seeking in the current bad market, he advises leveraging family/friend connections over online job boards.
HN Discussion:
  • ↑Author reiterates and defends his thesis, noting best vibe coders are already strong developers
  • ~Agrees with core argument but objects to misuse of the term 'deterministic' when comparing compilers to LLMs
  • ↑Fundamentals still matter and experienced devs will combine them effectively with LLMs
  • ↓Questions the premise that writing code is necessary to develop code-reading skills
  • ↓Disagrees, arguing LLMs will reduce dev headcount and management won't recognize complexity costs
7.Ask HN: What do you run on a $5 VPS that's worth keeping online 24/7?
390 points by mariocesar 7 days ago | 625 comments | permalink
tl;dr: A Linux user inspired by DHH's "AI shed" concept asks HN what's actually worth running on a cheap $5–10/month VPS long-term. They're weighing the maintenance overhead against the convenience of an always-on server, and want to know which self-hosted services have genuinely saved time or replaced paid subscriptions.
HN Discussion:
  • ↑Media/entertainment self-hosting stack (Jellyfin, qBittorrent, Immich) with Tailscale access is highly valuable
  • ↑A VPS enables running real businesses and startups cheaply, offering competitive advantage
  • ↑Personal automation, dashboards, and small web apps make a VPS indispensable for daily life
  • ~$5/month is actually expensive; much cheaper VPS options exist for lightweight tasks
  • ↓Self-hosting isn't worth it because it creates more maintenance problems than it solves
8.Beauty in DVD Menus(vale.rocks)
330 points by speckx 5 days ago | 184 comments | permalink
tl;dr: DVD menus hit their creative peak in the 2000s, with releases like Scooby-Doo 2, Harry Potter: Prisoner of Azkaban, Rocky Horror, and Wayne's World featuring elaborate animations, themed interfaces, and even playable games—all built within the format's tight technical constraints (MPEG-2 video, 4-color subpicture overlays, 16 general-purpose 16-bit registers). Despite Blu-ray's superior capabilities, modern physical media menus have largely regressed to generic static templates, mirroring the sterile interfaces of streaming platforms.
HN Discussion:
  • ↓Simple, standardized menus are actually preferable because they get out of the way of watching the movie
  • ↑Nostalgic appreciation for creative DVD menus and sharing personal memories of elaborate menu experiences
  • ↑Technical explanation for why Blu-ray menus regressed despite superior capabilities (authoring complexity)
  • ~Elaborate menus were frustrating in practice due to slow hardware and difficult navigation
  • ↑Personal collecting/creation of DVDs and menus as a hobby, celebrating the format
9.OpenAI withdraws three mathematical results(twitter.com)
373 points by sashank_1509 6 days ago | 607 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑Sign errors suggest higher probability of more undetected false proofs in the batch
  • ↑AI-generated proofs should all be formalized in Lean since manual review is infeasible
  • ↑Many of these AI proofs will likely fall apart on closer inspection over time
  • ↑Sarcasm mocking OpenAI for only its own papers needing retraction
  • ↓3 errors out of ~400 is actually a good success rate
10.ADHD as a circadian rhythm disorder: evidence and implications for chronotherapy (2025)(frontiersin.org)
383 points by bookofjoe 5 days ago | 256 comments | permalink
tl;dr: Up to 80% of ADHD patients show circadian disruption, including evening chronotype, delayed melatonin onset (~90 min in adults), blunted cortisol rhythms, reduced pineal volume, and attenuated BMAL1/PER2 clock-gene expression. Interventions that advance circadian phase—low-dose melatonin, morning bright light therapy, behavioral sleep programs, and fixed wake times—have been shown to improve both sleep and ADHD symptoms, with bright light appearing especially effective in winter. The authors propose routine circadian screening and a behavioral-first chronotherapy pathway as adjuncts to standard ADHD treatment, though more stratified trials are needed to define responder phenotypes.
HN Discussion:
  • ~Expert cautions that circadian correlation doesn't prove ADHD is a circadian disorder
  • ↑Personal experience confirms sleep interventions help manage ADHD symptoms
  • ↓Criticism of the journal's quality and credibility
  • ↓Title mislabels ADHD as a circadian disorder when it's only correlated
  • ~Findings are not new, just rehashing older studies
11.Tell HN: I've been paying for a rural Tanzanian's education for 10 years
742 points by lukehandcool 5 days ago | 237 comments | permalink
tl;dr: Ten years ago, a 19-year-old traveler began paying school fees for an 11-year-old girl in a rural Tanzanian village; she has since graduated from the University of Dar es Salaam and been accepted into a Master's program in Swahili. The effort has grown into a 501(c)(3), the Tanzania Education Project, which now funds schooling and healthcare for five students from the village, with the author seeking donations to scale beyond his own paycheck.
HN Discussion:
  • ↑Encourages others to donate time/money to good causes, affirming small contributions compound over time
  • ↑Shares personal stories of being helped by or sponsoring students, reinforcing the impact of such efforts
  • ↑Surfaces similar charitable organizations working in other developing countries
  • ~Warns about potential scams and urges verification/trusted sources for such charities
  • ~Suggests alternative aid models like funding education/skills rather than direct financing
12.Trump administration is suspending Microsoft from a green card program(apnews.com)
924 points by alephnerd 5 days ago | 1586 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑The suspension is correct because the H1B/green card program is widely abused to replace American workers
  • ~Enforcement appears selective and politically motivated, should be applied uniformly to all abusers
  • ↓The US needs a sane visa program with clear citizenship paths rather than punitive measures against workers
  • •Green cards should be based on tenure and taxes paid, not employer sponsorship
  • ↑Schadenfreude that Microsoft is being targeted by the administration it supported
13.Sharing AI progress in mathematics(openai.com)
1353 points by OfficialTurkey 7 days ago | 1530 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ~Personal reflection on AI solving problems researchers spent decades working on
  • ↑Excitement about significance of Unique Games Conjecture proof for complexity theory
  • ↑Highlighting breadth of impact across major open problems including Millennium Prize problems
  • ~Existential concern about AI surpassing humans in meaningful intellectual work
  • ↑Framing the progress as beauty to appreciate rather than competition to fear
14.The Mathocalypse(scottaaronson.blog)
395 points by 6bitquant 6 days ago | 414 comments | permalink
tl;dr: Scott Aaronson describes a hypothetical "Mathocalypse" in which OpenAI dumps 372 solutions to major open math and CS theory problems — including the Unique Games Conjecture, L=BPL, sub-O(n log n) integer multiplication, and partial progress on Riemann and Hodge — generated by an internal model using ~3 hours of compute per problem (~5% success rate on 8,000 attempts). The proofs come with Lean certificates but are nearly unreadable, triggering a scramble among mathematicians to decipher them. Aaronson contrasts OpenAI's "raw dump" approach with Anthropic's model of letting human researchers digest and announce results, and reflects on what mathematical research looks like now.
HN Discussion:
  • ↑The real story is the compute cost economics of AI-solved math problems
  • ↑Unreadable AI proofs mirror the thankless toil of reviewing AI-generated code in software engineering
  • ~The unreadability of the proofs casts doubt on their actual validity
  • ~Nobody has truly verified the proofs; Lean certificates alone may not be trustworthy
  • ↑Mere proof isn't the point of math; understanding and intuition matter more
15.“Math 2.0” will need to value mathematical progress more holistically(mathstodon.xyz)
608 points by ent101 6 days ago | 641 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑AI dumping proofs without engagement neglects the collaborative insight-building that defines mathematics
  • ↑Math is art/enlightenment, not just problem-solving, so AI should serve human understanding
  • ↑The holistic concern applies broadly across fields where AI prioritizes outputs over craft
  • ~Skeptical of AI's mathematical contributions as mostly information recombination rather than genuine insight
  • ↑Solving problems for benchmark cachet will fade, letting genuine knowledge-advancing uses emerge
16.OpenAI annualised revenues $20B less than previously signalled(cnbc.com)
422 points by mfiguiere 5 days ago | 298 comments | permalink
tl;dr: OpenAI told investors it hit roughly $50 billion in annualized revenue at end of September, well below the $68 billion figure widely reported last month (which included partner gross revenue). The disclosure sent AI-related stocks tumbling, with Nvidia, Oracle, CoreWeave, AMD, and Broadcom all falling significantly as the company tries to justify its $852 billion valuation ahead of an expected 2027 IPO. Rival Anthropic, meanwhile, reported a $65 billion annualized run rate in July and is reportedly seeking a $2 trillion valuation.
HN Discussion:
  • ↑Lack of transparency in trillion-dollar private AI companies is concerning
  • ↑Discrepancy suggests potential revenue misstatement or fraud by OpenAI
  • ↑Vindicates skeptics like Ed Zitron who've questioned AI company fundamentals
  • ↓This is a non-story caused by media/investor accounting error, not OpenAI misreporting
  • ↓FT is manufacturing drama by reporting both inflated and corrected numbers
17.I gave Opus 5.5 one prompt and six hours to visualize Invisible Cities(quesma.com)
406 points by stared 5 days ago | 199 comments | permalink
tl;dr: The author tested Claude Opus 5.5 against GPT-6 Astra by giving both a single prompt to build a Three.js visualization of Calvino's Invisible Cities over six hours. Opus 5.5 finished in 1h25m using 6 parallel subagents (~$74 in tokens) and produced a dramatically more polished, "mesmerizing" result, suggesting a significant leap in AI-assisted interactive design. The author notes this raises existential questions about the human role in creating interactive media.
HN Discussion:
  • ↓AI-generated projects feel soulless and fail to evoke engagement despite technical impressiveness
  • ↓The visualization fails to capture the book's deeper themes of semiotics and meaning
  • ↓The output has concrete quality flaws like inaccurate details and poor UX curation
  • ~Conflicted acknowledgment that AI capability raises genuine questions about creative hobby work
  • •Readers should preserve their imagination by avoiding visualizations of the book
18.4-hour battery storage is cheaper to install than gas turbines all across globe(solarpowerworldonline.com)
257 points by 01-_- 5 days ago | 159 comments | permalink
tl;dr: Four-hour battery storage is now cheaper than open-cycle gas turbines in all 43 markets Wood Mackenzie modeled, driven by expanding battery manufacturing (led by China, which sits 55% below the Asia Pacific average) and rising gas peaking costs. Single-axis solar is the lowest-cost new build in 43 of 48 markets, with Saudi and UAE solar projected below $20/MWh by 2033. The combination is foreclosing the economic case for new gas peaker capacity globally, though North American costs face near-term pressure from tariffs and Section 232 import restrictions.
HN Discussion:
  • ↑Praises the article's writing quality and informativeness
  • ↓Skeptical of forecast assumptions driving the conclusion
  • ~Points out batteries don't solve longer-duration gaps like cloudy windless weeks
  • •Questions why only 4-hour batteries are considered rather than longer durations
  • ~Clarifies article scope applies only to peaker/open-cycle turbines, not all gas
19.The Slow Formation of Durable Software(newsletter.dancohen.org)
290 points by benbreen 7 days ago | 116 comments | permalink
tl;dr: On Zotero's 20th anniversary, co-creator Dan Cohen recounts how the research tool emerged from years of slow, collaborative thinking at George Mason's Center for History and New Media, evolving from earlier tools like Web Scrapbook and Scribe before Firefox's XUL extension system made the vision technically feasible in 2004. The team iterated through names (SmartFox, Firefox Scholar) before settling on an Albanian word, launching in October 2006 to 60,000 users within a month. Cohen argues this deliberate pace—impossible to replicate with today's "ask for what you want" AI coding—is precisely what produced software durable enough to still be widely used two decades later.
HN Discussion:
  • ↑Zotero is genuinely excellent software that exemplifies the article's point about durability
  • ↑Slow, deliberate development produces better software than fast AI-driven churn
  • ↑Products need real problems to solve; rushing with AI prompts skips necessary discovery
  • ↓AI doesn't preclude slow deliberate development and can actually assist the discovery process
  • ~Real-world enterprise software development still requires slow alignment regardless of AI tools
20.House with 15m underground tunnels for sale for 300k(readingchronicle.co.uk)
236 points by librasteve 6 days ago | 214 comments | permalink
tl;dr: A Tilehurst property is up for auction at £300K-£325K featuring a 1,300+ sq ft network of hand-dug tunnels extending 15m underground across four levels, built by civil engineer Tony Deane and his six children over two decades using only buckets and wheelbarrows. The complex includes a hand-built railway, a one-tonne goods lift, secret doors (including a 1.5-tonne sliding concrete wall), and an entrance hidden inside a 23ft rocket by the pool. The auctioneer suggests potential uses ranging from secure storage to an EMP-resistant data center.
HN Discussion:
  • ↑Shares personal anecdotes about similar tunnel systems in other locations
  • •Provides additional resources like listing links and more photos
  • ↑Appreciates the quirky hobby of a qualified engineer
  • ↑Points to related tunnel-building personalities and projects
  • ~Skeptical about practicality, planning permission, or the hiding claim