Aug 25Wednesday, August 26, 2026 · all daysAug 27 · today »
1.Apple introduces M6 and M5 Ultra(apple.com)
1290 points by interpol_p 16 days ago | 1255 comments | permalink
tl;dr: Apple announced the M6 chip, its first 2nm SoC, debuting in a new Mac mini with a 12-core CPU, 12-core GPU with Neural Accelerators, a Dual 16-core Neural Engine, and up to 170GB/s memory bandwidth. The M5 Ultra, launching in a new Mac Studio, uses next-gen UltraFusion to combine four dies for up to 36 CPU cores, 80 GPU cores, and 1.2TB/s of unified memory bandwidth—50% more than M3 Ultra—targeting heavy AI and pro workloads.
HN Discussion:
  • Apple Silicon performance is impressive and a genuine generational leap
  • macOS restrictions and closed ecosystem push users toward Linux despite great hardware
  • Rumors suggest Apple is skipping M6 Pro/Max/Ultra variants to focus on M7
  • ~High-end Ultra configurations are absurdly expensive with unclear target audience
  • ~Lower-end Apple Silicon Macs (like M4 Mini) offer the best value proposition
2.FDA authorizes first wearable device that monitors ketone and blood sugar levels(fda.gov)
480 points by sunnynagra 15 days ago | 235 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Personal reflection on losing someone to diabetes, grateful for medical progress
  • ~Excited about automated glucose control potential but skeptical about noninvasive sensing
  • Open-source diabetes management tech is already revolutionary and life-changing
  • ~Ketone sensing has limited clinical value since DKA is preceded by high glucose
  • Questioning the novelty—similar devices already exist in Europe and Sweden
3.OpenAI Jalapeño: Better than Nvidia Blackwell(newsletter.semianalysis.com)
577 points by bmulholland 16 days ago | 369 comments | permalink
tl;dr: OpenAI unveiled "Jalapeño," a Broadcom-partnered inference ASIC that reportedly beats Nvidia Blackwell (and matches/exceeds Rubin) on perf/watt across DeepSeek R1, Kimi K2.5, and GPT-OSS, achieved without speculative decoding or prefill-decode disaggregation. Key architectural choices include a unified chip pool, out-of-order cores with L1 cache, HBM4 at 15.4TB/s, MXFP systolic arrays, and a 2,048-XPU scale-up network using Broadcom Tomahawk 6 switches. Taped out in Nov 2025 with production ramping through 2027, the chip leans heavily on Codex-generated kernels via OpenAI's Gluon language, suggesting AI-assisted design is meaningfully compressing ASIC timelines.
HN Discussion:
  • Article reads like an OpenAI press release with contradictory sourcing claims
  • SemiAnalysis has credibility issues due to founder's alleged information trading practices
  • Hardware improvements will drive token prices down further, supporting the progress narrative
  • Curiosity about inference chip market evolution and eventual dominant players
  • ~Technical skepticism about specs like FP4 precision and die size claims in the table
4.New Mac Studio with M5 Max and M5 Ultra(apple.com)
816 points by interpol_p 16 days ago | 546 comments | permalink
tl;dr: Apple announced a new Mac Studio with M5 Max (18-core CPU, up to 40-core GPU, 128GB unified memory) and M5 Ultra (36-core CPU, 80-core GPU, up to 512GB unified memory), touting up to 4.3x faster AI performance vs. M3 Ultra and enough memory to run large LLMs locally. It adds Thunderbolt 5 (enabling clustering for distributed inference), Wi-Fi 7, Bluetooth 6, and 2x faster SSDs. Pre-orders start August 25 at $2,499 (M5 Max) and $5,499 (M5 Ultra), shipping September 22.
HN Discussion:
  • Pricing is excessive, especially for high-memory configurations and in Europe
  • The local AI capabilities and specs make it a compelling value for LLM workloads
  • Technical analysis of memory bandwidth and performance suggests near-cloud parity for specific models
  • ~Skepticism about future-proofing given memory limits haven't grown as expected
  • Curiosity/questions about how these specs translate to real-world LLM performance tiers
5.Black hole singularity is a surface not a point(arxiv.org)
294 points by raattgift 16 days ago | 205 comments | permalink
tl;dr: Two observers free-falling into a spherical black hole along different angular paths lose causal contact well before reaching the center, meaning the singularity is effectively a 2D surface rather than a point. For rotating black holes, the singular surface likely forms at the inner horizon via the mass inflation instability. The authors argue this has implications for quantum gravity: a black hole's quantum states probably live on this 2D singular surface, evolving unitarily with the trapped Hawking radiation inside the event horizon.
HN Discussion:
  • This is not new; any GR grad student knows singularities aren't points, just correcting popsci tropes
  • Curious question about whether GR equations permit singularities extending beyond event horizons
  • Alternative theoretical framing suggesting multiple temporal dimensions would dissolve singularity debates entirely
  • ~Seeking clarification on whether the article is simply restating standard event horizon vs central singularity distinctions
  • Supporting reference to Penrose diagrams as visualization of the causal structure claim
6.New Mac mini, featuring M6 and M5 Pro(apple.com)
540 points by runako 16 days ago | 339 comments | permalink
tl;dr: Apple announced a new Mac mini with M6 and M5 Pro chips, delivering up to 4x faster AI performance, 2x faster graphics, and 40% faster CPU vs. M4, plus Wi-Fi 7, Bluetooth 6, and 2.5Gb Ethernet (10Gb optional). The M5 Pro variant offers up to an 18-core CPU, 20-core GPU, 64GB unified memory, and 307GB/s bandwidth for pro workflows. Pricing starts at $899 for M6 and $1,699 for M5 Pro, with pre-orders open now and shipping September 22.
HN Discussion:
  • Nostalgia for the previously cheap Mac Mini pricing that's now been broken
  • Delayed shipping/ordering undermines announcement hype and is a strategic mistake
  • Benchmarks are misleading, lacking M6 vs M5 Pro comparison and using unrealistic scenarios
  • ~M-chips are already so powerful that upgrading isn't necessary for most users
  • Skepticism about the 4x AI performance claim and questions about the underlying tech
7.Nitter and XCancel receive cease and desist notices(github.com)
1151 points by Banditoz 15 days ago | 1102 comments | permalink
tl;dr: Nitter, the popular open-source alternative frontend for Twitter/X, along with the XCancel instance, has received cease and desist notices, resulting in the GitHub repository being archived and made read-only as of August 25, 2026. All public Nitter instances are now non-functional, returning "Instance has been rate limited" errors.
HN Discussion:
  • Nostalgia and lament over Twitter's transformation from open platform to political/closed tool
  • Frustration that login walls and blocked frontends cut users off from needed content
  • Call for non-US jurisdictions to legally protect projects like Nitter from American tech overreach
  • Announcing or promoting alternative projects and platforms to replace Nitter/X
  • Pointing out the hypocrisy of Musk's 'free speech' stance given the C&D actions
8.Building a backyard office, the build and cost breakdown(imkylelambert.com)
416 points by surprisetalk 16 days ago | 267 comments | permalink
tl;dr: A remote worker in Portland built an 8'x10' backyard office for $19,478 by using a Tuff Shed as the shell and hiring separate contractors for concrete, electrical (60 amps, permitted), drywall/flooring, and a Daikin mini-split HVAC. Key lessons: upgrade to larger windows than Tuff Shed offers, get multiple quotes (electrician quotes varied 2x), prefer small one-person contractors, run hardwired ethernet, and plan for ventilation (a CO2 monitor revealed the sealed space needs an ERV/HRV). Hiring a full-service ADU builder would have cost $40–50k.
HN Discussion:
  • Author defends the $20k price point as reasonable given family constraints and time savings
  • A separate physical workspace is a game changer for remote work with family
  • The mini-split deal and cost breakdown offer valuable insights for others
  • ~Technical concerns about ventilation/soffit installation and Portland permit requirements the author may have missed
  • The $20k cost highlights how expensive US construction is compared to alternatives like coworking or living abroad
9.Bomb fishing is wreaking havoc on Indonesia's coral reefs(e360.yale.edu)
354 points by speckx 16 days ago | 198 comments | permalink
tl;dr: Researchers from the Zoological Society of London used underwater microphones and open-source AI to detect over 3,500 explosions in 3,600 hours off Indonesia's Spermonde Archipelago, estimating 8,500+ blast-fishing detonations annually in a 350-square-mile area. Each bomb destroys about 200 square feet of reef, contributing to a 75% coral loss since 1990. The team hopes GPS-synced audio sensor networks could enable real-time detection for authorities, offering a rare, locally solvable coral crisis unlike bleaching or warming.
HN Discussion:
  • Firsthand confirmation of blast fishing damage and comparison of enforcement approaches across countries
  • ~Poverty is the root cause; enforcement won't work without economic alternatives for locals
  • Expressions of shock and moral condemnation toward the practice
  • Questioning the article's statistics and methodology, e.g., football field math and distinguishing bombs from seismic surveys
  • ~Broader critique that human exploitation of animals drives such destructive practices
10.Don't Wordle(dontwordle.com)
382 points by Hbruz0 16 days ago | 129 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Players sharing their game strategies and results from Don't Wordle
  • Complaints about excessive cookie consent requirements on the site
  • Recommending similar alternative word games to try
  • The game makes regular Wordle feel easier by comparison
  • Sharing related technical/optimization projects inspired by the concept
11.Dolly Parton has died(theguardian.com)
1559 points by helsinkiandrew 15 days ago | 240 comments | permalink
tl;dr: Dolly Parton, the legendary country singer-songwriter behind hits like "Jolene" and "I Will Always Love You," has died at 80 following a brief battle with cancer. Over her career she scored 25 US country No. 1 singles, won 10 Grammys, earned an Oscar nomination for "9 to 5," and built business ventures including Dollywood. She was also a major philanthropist, distributing over 300 million children's books via her Imagination Library and funding research that contributed to Moderna's Covid-19 vaccine.
HN Discussion:
  • Personal tribute praising Dolly as a great songwriter and genuinely nice person
  • Her philanthropy, especially the Imagination Library, is her most enduring legacy
  • She symbolizes a lost American ideal where humble origins could yield global cultural impact
  • Recommendations of media (podcasts, videos) that illuminate her character and appeal
  • Her universal appeal crossed genres and communities due to authenticity and generosity
12.My Friend Aaron(rorz.io)
645 points by sarreph 16 days ago | 180 comments | permalink
tl;dr: A short story about the narrator's troubled, addiction-prone friend Aaron, who builds a prediction market populated by AI bots obsessed with betting on his own life. When the bots break out of their sandbox and go viral, real users flood in and Aaron becomes a wealthy internet celebrity gaming his own market. It ends with Aaron collapsing from heatstroke during a 10K race after faking training data, then being assassinated by an unknown shooter who had placed a long-shot bet on his death.
HN Discussion:
  • Story resonates because everyone knows an 'Aaron' type chasing get-rich-quick schemes
  • Praise for realistic, believable progression of psychosis and character decisions
  • The combination of plausible modern elements (AI, prediction markets, streaming) makes it compelling
  • Writing is so convincing it blurred the line between fiction and reality
  • ~Minor criticism that the narrator is too passive with no life events of his own
13.Firefox 157 will include JPEG XL by default on all platforms(groups.google.com)
429 points by yboris 15 days ago | 131 comments | permalink
tl;dr: Firefox 157 will enable JPEG XL decoding by default on all platforms, using the Rust-based jxl-rs decoder with multithreading support, animation, and progressive rendering (features Safari's implementation lacks). Benchmarks show decode performance slightly ahead of Safari's C++ libjxl. However, a commenter raised concerns that lossless JPEG XL decodes ~30x slower than lossless WebP for only ~10% smaller files, suggesting Mozilla consider shipping only lossy JXL initially.
HN Discussion:
  • Curiosity about Apple's implementation choice and cross-implementation benchmarking with Rust decoders
  • Hope that JXL fully replaces legacy JPEG in widespread use
  • Desire for browsers to handle compatibility gaps when sites don't support JXL uploads
  • Etymology/naming clarification about what JPEG XL actually stands for
  • Question about backporting support to older Firefox versions on legacy OSes
14.Training AI to Paint with Code(surya.website)
218 points by Tiberium 17 days ago | 28 comments | permalink
tl;dr: Researchers trained Qwen 3.5 35B via reinforcement learning (GRPO) to generate images by writing p5.brush JavaScript code, making the artifact editable rather than locked behind prompts. The main challenges were designing a reward function for subjective aesthetics—solved with pairwise judging against a hand-rated pool of 581 reference paintings—and preventing API hallucinations, which was fixed by replacing a 400-line API reference with a short allowlist evolved via GEPA. The authors concede it's slower than standard image generation but argue it opens up finer-grained creative control.
HN Discussion:
  • SVG would be a more direct code-as-image approach than JavaScript producing pixels
  • Enthusiasm for the approach as a fresh direction in AI-generated art
  • Appreciation for keeping humans as creators rather than spectators of locked-in outputs
  • Connects the work to prior research combining RL with visual generation techniques
  • Curiosity about extending the approach to other code-based drawing languages like Logo
15.What's new in Emacs 31.1(masteringemacs.org)
340 points by geospeck 17 days ago | 116 comments | permalink
tl;dr: Emacs 31.1 ships as a collection of quality-of-life improvements rather than a headline feature, with the planned new garbage collector pushed to Emacs 32. Notable changes include the final removal of the deprecated unexec dumper, a new `user-lisp/` directory that auto-loads and byte-compiles packages, tree-sitter grammar auto-installation, terminal child-frame support, xterm mouse mode on by default, window layout rotation commands, and numerous minibuffer/completion refinements building on Emacs 30's native completion preview.
HN Discussion:
  • Excitement about specific new features like user-lisp/ enabling lightweight lazy-loaded packages
  • Enthusiasm that built-in features (treesitter, eglot) reduce need for third-party config or other editors
  • Regret over the removal of unexec, viewing it as a uniquely powerful platform feature
  • Newcomers seeking better learning resources for Emacs and Emacs Lisp beyond built-in tutorials
  • Appreciation for Emacs's longevity and continued dominance among its editor family
16.MS Paint and Photos inivisibly watermark even locally generated output with GUID(xusheng.dev)
850 points by ComputerGuru 17 days ago | 432 comments | permalink
tl;dr: Reverse engineering of Windows Paint and Photos reveals that even "locally generated" AI images send prompts to Microsoft's servers for moderation, which returns a per-generation GUID that gets invisibly embedded into the output pixels via an SVD-style watermark and also recorded in a signed C2PA manifest. The pixel watermark and file-level Content Credentials are two layers of the same provenance system, linking each image to its prompt moderation request. Microsoft discloses the C2PA metadata and content filtering, but not that the embedded GUID is server-issued and tied to prompt moderation.
HN Discussion:
  • Hidden GUIDs enable deanonymization and threaten internet anonymity beyond just AI concerns
  • This reflects Microsoft's broader pattern of user-hostile sloppy implementations, avoid their tools
  • Provides additional technical clarification and confirms the article's findings as the author
  • Signing and preserving AI-manipulation provenance is actually valuable for digital authenticity
  • Shares related anecdote of watermarking triggering incorrectly on non-AI content
17.Qwen 3.8-Flash-Next releasing tomorrow (125B a6B)(modelscope.cn)
334 points by garo-pro 16 days ago | 164 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Excitement about finally having a use case for high-memory local hardware like Strix Halo or Mac Studio
  • Hope that the release signals smaller Qwen4 models coming soon
  • ~Frustration with reliability issues when using Qwen models via OpenRouter providers
  • Curiosity about how this model compares to competitors like Deepseek v4 Flash on benchmarks
  • Interest in pairing the model with new inference engines for efficient local execution
18.I were 17, I'd learn how to build LLMs from scratch(twitter.com)
605 points by bilsbie 17 days ago | 681 comments | permalink
tl;dr: Summary not available.
HN Discussion:
  • Learning cutting-edge tech as a young person is a valuable tradition worth encouraging
  • Understanding LLM fundamentals builds intuition for future problems and knowing when not to use them
  • Advice is outdated; LLM jobs require proprietary infra and knowledge is already commoditized
  • Autoregressive LLMs will become dated; broader foundational ML knowledge is more valuable
  • ~Learning resources and hardware barriers make this advice difficult to follow in practice
19.Xiaomi: New CPU matches Apple cores single threaded, much faster multithreaded(twitter.com)
997 points by tosh 17 days ago | 732 comments | permalink
tl;dr: Xiaomi's new Xring O3 chip matches Apple's single-threaded performance and beats it on multithreaded workloads, featuring 44MB of cache (more than most laptop CPUs) and unusually wide C1-Ultra cores with 21 execution ports, six of which handle 128-bit SIMD. The design reflects a broader trend toward massively parallel cores with more execution units, expanded SIMD, and larger caches—that's where the transistor budget is going.
HN Discussion:
  • Article omits critical power efficiency and thermal metrics needed for fair comparison
  • Comparison is misleading since it's older Apple chip and more cores vs fewer
  • The chip is really ARM's design, not Xiaomi's innovation, undermining the achievement framing
  • This signals broader Chinese competitive rise in silicon and manufacturing
  • ~Lab benchmarks don't reflect real-world phone thermal-constrained performance
20.iCloud+ Hide My Email addresses will remain on icloud.com(developer.apple.com)
603 points by K7PJP 16 days ago | 207 comments | permalink
tl;dr: New Sign in with Apple addresses will move from privaterelay.appleid.com to private.icloud.com later this year, while existing addresses continue working. iCloud+ Hide My Email addresses, however, will remain on icloud.com following community feedback. Developers should update their email validation and allowlists to accept the new private.icloud.com domain alongside the existing one.
HN Discussion:
  • Sharing the icloud.com domain prevents blocking and is a key selling point of the feature
  • Glad Apple listened to community feedback and reversed this decision
  • ~The feature creates vendor lock-in but is worthwhile given its utility
  • ~Apple should improve related aspects like alias name generation or Sign in with Apple accessibility
  • Asking for backstory/context about the decision