Jul 17Saturday, July 18, 2026 · all days
1.Regressive JPEGs(maurycyz.com)
416 points by vitaut 8 hours ago | 42 comments | permalink
tl;dr: Progressive JPEGs store images as multiple "scans" of increasing detail, and by concatenating multiple images while stripping certain markers, you can create a single JPEG that displays different frames as it loads over a slow network. Since most decoders bail out after ~9 scans, the trick is to use minimal DC-only scans (yielding 1/16 resolution frames), which allows packing ~90 frames—enough for a crude video—into one standards-compliant JPEG. Playback timing depends entirely on network speed, so it has no real practical use beyond novelty demos like HTML-only video via `<dialog>` tags.
HN Discussion:
  • ~Timing can be controlled by server-side chunked delivery rather than pure network delay
  • Suggests practical applications like progress bars or steganography, countering the 'no practical use' claim
  • Pure appreciation of the hack's cleverness and execution
  • Progressive JPEGs are generally not worth the decoding slowdown in practice
  • Extension ideas like GIF-to-regressive-JPEG conversion
2.AWS: Inaccurate Estimated Billing Data – $1.7 billion
1213 points by nprateem 1 day ago | 716 comments | permalink
tl;dr: An AWS user reports their estimated monthly bill spiked to $1.7 billion despite normal usage typically under $5, and has filed an urgent support ticket. AWS's status page is referenced alongside a Reddit thread where others appear to be experiencing similar billing anomalies, suggesting a widespread issue with AWS's estimated billing data.
HN Discussion:
  • Insider explains this is a known unit conversion bug (GB vs Bytes) in AWS billing
  • Users share personal shock and humor at receiving absurd billing amounts
  • Jokes about scale of the bill and negotiating repayment with AWS
  • Recounting past AWS billing errors that were harder to resolve, questioning reliability
  • Broader observation about silent billing/pricing bugs affecting customers unnoticed
3.Thanks HN for 15 years of support and helping me find my life's work
644 points by nicholasjbs 19 hours ago | 73 comments | permalink
tl;dr: The founder of the Recurse Center marks its 15th anniversary, recounting how they pivoted from a failed YC "OkCupid for jobs" idea into a self-directed programming retreat. A 2012 HN launch post drove much of their early growth and remains their #2 applicant source after word of mouth. While not the billion-dollar business YC hoped for, it has served over 3,000 programmers and become the founder's life's work.
HN Discussion:
  • Former participants share fond memories and credit RC with life-changing experiences
  • Curiosity about why the original 'OkCupid for jobs' idea failed
  • ~Questions about accessibility given NYC living costs during the free program
  • Interest in the business model and how RC can afford to be free
  • Appreciation for RC's social rules and community design
4.First atmosphere found on Earth-like planet in habitable zone of distant star(bbc.com)
469 points by neversaydie 21 hours ago | 276 comments | permalink
tl;dr: Researchers have detected helium in the atmosphere of LHS 1140 b, a rocky exoplanet 48 light-years away orbiting in its red dwarf star's habitable zone—the first atmosphere confirmed on an Earth-like world in a Goldilocks zone. Helium alone can't support life, but other gases may exist lower in the atmosphere. The finding, published in Science, is a step toward answering whether Earth-like conditions exist elsewhere, though no biosignatures have been detected.
HN Discussion:
  • Skepticism that LHS 1140b is truly Earth-like, suggesting it may be a mini-Neptune being stripped by its star
  • Excitement about proximity and speculation on propulsion technology to send probes there
  • Focus should shift to the Fermi paradox and the narrow communication window problem
  • Skepticism that helium alone is meaningful, questioning the significance of the finding
  • ~Reminder that 'Earth-like in habitable zone' doesn't imply habitability, citing Venus as example
5.The Zilog Z80 has turned 50(goliath32.com)
238 points by st_goliath 16 hours ago | 95 comments | permalink
tl;dr: The Zilog Z80, launched in July 1976 by ex-Intel engineer Federico Faggin, evolved from the Datapoint 2200-derived 8008 and 8080 lineage, adding index registers, shadow register banks, new instructions, and a simpler single-supply bus design while remaining binary compatible with the 8080. It became a dominant 8-bit processor in early PCs, home computers, and embedded systems, spawning clones like the Game Boy's Sharp LR35902, before Zilog finally discontinued the original in June 2024. The article also recounts how Exxon's ownership stake contributed to IBM choosing Intel's 8088 for the PC.
HN Discussion:
  • Nostalgic reflections on learning programming and assembly via Z80-based systems
  • Highlighting Z80's remarkable longevity in products like TI calculators and MP3 players
  • Technical correction that Z80 wasn't fully binary compatible with 8080 due to flag differences
  • Appreciation for Z80's simplicity and elegant features like block instructions and 16-bit extensions
  • Fond memories of specific Z80-powered hardware like ZX Spectrum, TRS-80, and Game Gear
6.Learning a few things about running SQLite(jvns.ca)
267 points by surprisetalk 18 hours ago | 69 comments | permalink
tl;dr: A developer running a Django site on SQLite shares lessons learned: running ANALYZE dramatically sped up a slow FTS5 query (from 5s to 0.05s) by giving the query planner better statistics. They also hit issues with SQLite's single-writer limitation during bulk deletes causing worker timeouts and crashes, which they work around with small batches. For backups, they've used both restic (with VACUUM INTO) and Litestream for incremental replication to S3.
HN Discussion:
  • Sharing helpful SQLite tools and tips that complement the author's learnings
  • Criticizing the article for lacking rigor, guessing instead of investigating
  • Offering concrete solutions to the DELETE and backup problems the author faced
  • Appreciating the author's authentic exploratory writing style, especially in the LLM era
  • ~Puzzled that such small tables caused problems, suggesting basic indexing would solve it
7.Kimi K3, and what we can still learn from the pelican benchmark(simonwillison.net)
356 points by droidjj 21 hours ago | 188 comments | permalink
tl;dr: Moonshot AI released Kimi K3, a 2.8 trillion parameter model priced at $3/$15 per million input/output tokens—matching Claude Sonnet pricing and making it the most expensive Chinese model to date, with open weights promised by July 27. Simon Willison's pelican-on-a-bicycle SVG benchmark no longer correlates well with model quality (GLM-5.2 outperforms frontier models on it), but he still finds value in it as a "hello world" prompt for testing new models, gauging reasoning costs, and verifying basic spatial reasoning capabilities.
HN Discussion:
  • Technical curiosity about hidden system prompts and token counting anomalies in Kimi K3
  • ~The pelican benchmark is limited and misses what matters most like agentic tool calling
  • The benchmark's value lies in comparing cost/speed/quality tradeoffs rather than ranking models
  • Existing benchmarks including pelican fail to measure what truly matters in real code work
  • Pelican benchmark is likely contaminated by training data and single-run variance makes comparisons unreliable
8.The state of open source AI(stateofopensource.ai)
456 points by rellem 21 hours ago | 328 comments | permalink
tl;dr: Open-weight AI models have essentially closed the capability gap with closed frontier models (3.3% on Chatbot Arena, at parity for coding), now route the majority of production tokens on OpenRouter, and cost roughly 6× less per call — while inference prices have fallen 50× in 36 months. However, open lags in operational tooling: only 51% of open-model teams reach production vs. 63% for closed, with the gap concentrated in deployment, compliance, and maintenance. The emerging battleground is the "agentic harness" above the model, where frontier labs are integrating scaffolding with weights to rebuild lock-in.
HN Discussion:
  • Open models will disrupt and eventually kill closed frontier labs like OpenAI and Anthropic
  • Data confirms explosive growth of open model usage on OpenRouter
  • The article is LLM-generated slop, undermining its credibility and message
  • Mozilla's framing of its own history and role is inaccurate or hypocritical
  • The article conflates open weights with open source, which are distinct
9.Kaiser nurses say AI, surveillance are making their jobs and patient care worse(localnewsmatters.org)
518 points by gnabgib 13 hours ago | 341 comments | permalink
tl;dr: Kaiser Permanente advice nurses say workplace surveillance and AI tools—including systems that grade their tone, empathy, and call length—are pressuring them to cut calls short (often under 15 minutes) at the expense of patient care, with some withholding compassion from suicidal or terminally ill patients to avoid poor performance scores. The California Nurses Association is heading into contract negotiations with AI as a central issue, while state lawmakers weigh several bills to regulate workplace AI, including protections for clinicians who override automated recommendations. Kaiser denies using average handle time to evaluate nurses.
HN Discussion:
  • Article conflates AI issues with call center metrics; actual AI tools help clinicians
  • Using machines to evaluate empathy or track workers is inherently wrong and harmful to care
  • Surveillance and bad metrics are an industry-wide problem degrading patient care
  • ~The article lacks concrete examples of AI harm; the real issue is cost-optimization
  • Study is motivated reasoning tied to union contract negotiations and misses real issues
10.Three ways people respond to a problem (other than solving it)(improvesomething.today)
254 points by surprisetalk 22 hours ago | 143 comments | permalink
tl;dr: Beyond actually solving problems, people typically respond in three other ways: pushing problems around (local optimization that shifts pain elsewhere), preserving problems (per the Shirky Principle, institutions perpetuate the issues they exist to solve), and promoting new problems (once the top issue is resolved, the next one takes its place). The author advises identifying who benefits from a problem's existence, accepting that problem-solving is never "finished," and using shared diagrams to help teams agree on which problems are actually worth fixing.
HN Discussion:
  • ~Framework maps onto political ideologies, oversimplified but contains truth
  • ~Adds a fourth response: ignoring or downplaying problems as often the best choice
  • Problem preservation applies to individual experts, not just institutions
  • Article assumes problems are real and well-defined, which often isn't true
  • Preserving problems explains government inefficiency in solving social issues
11.Kimi K3: Open Frontier Intelligence(kimi.com)
2039 points by vincent_s 1 day ago | 1181 comments | permalink
tl;dr: Moonshot has released Kimi K3, a 2.8T-parameter open-weight model built on Kimi Delta Attention and Attention Residuals, with native vision, a 1M-token context window, and MoE sparsity activating 16 of 896 experts. It targets long-horizon coding, agentic knowledge work, and reasoning—demonstrating strong results on kernel optimization, compiler building, and chip design—though it still trails Claude Fable 5 and GPT 5.6 Sol. Available now via Kimi.com and API ($0.30/$3/$15 per MTok), with full weights slated for release by July 27, 2026.
HN Discussion:
  • Personal testing shows K3 performs impressively, sometimes outperforming competitors on real tasks
  • ~The high pricing undermines K3's competitiveness despite strong capabilities
  • K3 represents Chinese labs commoditizing AI intelligence, threatening US labs' financial position
  • K3's scale (2.8T params) makes it the largest open model, a notable achievement
  • Website generation and creative demos showcase K3's impressive practical capabilities
12.Evidence of inconsistencies in evaluation process and selection of winners(kaggle.com)
461 points by twerkmeister 1 day ago | 289 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • AI-generated submissions and AI-based judging have corrupted hackathons and competitions
  • Organizer defends the process, explaining extended judging timeline and effort
  • ~Kaggle was never a rigorous source, so these problems aren't surprising
  • ~Brute-force and black-box winning approaches have always existed in Kaggle
  • Mocking the dismissive 'just accept the results' attitude toward inconsistencies
13.Camera Chase Vehicle(transistor-man.com)
228 points by geerlingguy 8 days ago | 22 comments | permalink
tl;dr: A hobbyist built a remote-controlled camera chase vehicle by mounting a Freefly Movi M10 gimbal (bought used for $124) onto a 1/5-scale RC chassis from a surplus store, with a VESC motor controller, DeWalt battery packs, 3D-printed TPU fenders for ice racing, and an Amimon CONNEX 5.8GHz HD video link. The write-up covers mechanical design, power distribution, reverse-engineering Panasonic camera lens control via the shutter port's analog resistance signaling, and unlocking an offline firmware update mode for the discontinued CONNEX radios. Initial shakedown at 30 MPH revealed the gimbal was poorly balanced, degrading footage quality.
HN Discussion:
  • Appreciation for hobbyist content as a refreshing break from AI/ML-dominated HN feed
  • Nostalgic recognition of the author's previous impressive projects and prolific output
  • Technical admiration for specific clever solutions like using TPU for shear resistance
  • Appreciation for the old-internet character of the personal, ad-free writeup style
  • Desire for additional content like FPV footage of the crash
14.Apple targets dozens of OpenAI employees with legal letters(ft.com)
395 points by merksittich 1 day ago | 345 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Document retention letters are standard legal practice, not aggressive escalation as framed
  • Apple likely has strong evidence and will win, threatening OpenAI's IPO plans
  • OpenAI's alleged theft is consistent with its content-stealing origins
  • Crossing Apple is dangerous given their platform leverage over ChatGPT
  • OpenAI's hardware platform ambitions with Jony Ive may be undermined by this
15.How Has Roman Concrete Lasted for Millennia? 1,900-Year-Old Latrine Offers Clues(smithsonianmag.com)
259 points by divbzero 1 day ago | 216 comments | permalink
tl;dr: Researchers studying an undisturbed 1,900-year-old concrete sample from a latrine at Hadrian's Villa found that carbonation—atmospheric CO2 reacting with calcium compounds to form crack-sealing calcite—plays a larger role in Roman concrete's durability than previously thought, complementing the well-known pozzolanic reaction. The findings, published in Science Advances, could inform development of longer-lasting, lower-emission modern concrete, which currently accounts for roughly 8% of global CO2 emissions.
HN Discussion:
  • Explains the underlying lime cycle chemistry that supports the article's carbonation findings
  • Notes carbonation in lime cement is unsurprising, downplaying the article's novelty
  • ~Modern concrete failures stem from rebar corrosion and economics, not inferior chemistry vs Romans
  • Building for extreme longevity wastes resources; shorter-lived structures are actually better engineering
  • Traditional lime-based building techniques like hempcrete apply these principles today for sustainability
16.The LLM Critics Are Right. I Use LLMs Anyway(theocharis.dev)
297 points by JeremyTheo 2 days ago | 296 comments | permalink
tl;dr: The author acknowledges that most criticisms of LLMs—slop, environmental cost, OSS trust erosion, geopolitical risk, opinion homogenization—are valid, yet still uses LLMs heavily (spending ~$10k/month on tokens) because they amplify existing thinking rather than replace it. They share concrete workflows: a "grill-me" skill that forces the LLM to interrogate you until shared understanding, Basecamp-style 3-sentence problem statements, subagents that rip apart context until they hallucinate, and using hallucinated APIs as intuition tests. The key rule: only use LLMs in domains where you can distinguish good output from slop.
HN Discussion:
  • Concern that heavy LLM use will atrophy cognitive/engineering skills over time
  • The $10k monthly spend is absurd, wasteful, or hypocritical given environmental concerns
  • LLMs are qualitatively different from past tools like calculators and shouldn't be compared
  • ~Skepticism about the article's claimed workflow benefits, especially skimming output
  • Pushback on specific claims like geopolitical framing being one-sided projection
17.Microsoft Comic Chat is now open source(opensource.microsoft.com)
797 points by jervant 1 day ago | 173 comments | permalink
tl;dr: Microsoft has open-sourced Comic Chat, the 1996 IRC client that rendered conversations as comic panels with illustrated characters and speech bubbles, and famously debuted the Comic Sans font. The release includes original source code (Visual C++ 4.0/MFC) along with AI-assisted modernization attempts that get it building with current Visual Studio tools and connecting to modern IRC servers. Created by DJ Kurlander's team at Microsoft Research with artwork by cartoonist Jim Woodring, the code is now available on GitHub for study and experimentation.
HN Discussion:
  • Insider account from the person who made the open-sourcing happen, sharing the backstory
  • Personal nostalgia and inspiration stories tied to Comic Chat's creative legacy
  • Appreciation for the era's willingness to experiment with unreasonable creative ideas
  • ~Historical critique noting Comic Chat was reviled in IRC culture for polluting the protocol
  • Technical exploration of the codebase, sharing easter eggs, papers, and repo links
18.Mathematics of Data Science(arxiv.org)
208 points by Anon84 1 day ago | 14 comments | permalink
tl;dr: A textbook on the mathematical foundations of data science, covering 16 chapters spanning high-dimensional geometry, SVD/PCA, linear regression, graph-based methods, dimension reduction, optimization, classification, and deep learning. It also treats more theoretical topics including concentration of measure, matrix concentration inequalities, compressive sensing, and low-rank matrix recovery. Authored by Thomas Strohmer and posted to arXiv.
HN Discussion:
  • High-dimensional intuition is a crucial foundation that this book valuably addresses first
  • Data science jobs really require decision-making intuition, which is hard to build
  • ~Statistics fundamentals matter more than the mathematical topics emphasized in the book
  • Recommending related books and resources on similar mathematical foundations
  • Practical question about converting the LaTeX source into an epub format
19.Decoy Font(mixfont.com)
705 points by ray__ 1 day ago | 157 comments | permalink
tl;dr: Decoy Font is a downloadable TTF font that exploits the hybrid image technique (same idea as the Einstein/Monroe illusion), overlaying sharp foreground letters on blurred, low-frequency background letters so humans reading from a distance see a different message than what AI/OCR sees up close. The author demonstrates it fooling ChatGPT and Gemini 3.5, though notes agentic models with proper prompting could likely defeat it. It's positioned as an accessible anti-AI-scraping tool, with potential applications in CAPTCHAs or private messaging.
HN Discussion:
  • ~Cool concept but ultimately ineffective at its stated purpose of blocking AI
  • AI models can easily defeat it, especially with hints or simple preprocessing
  • Shares related prior technical work on hybrid image techniques
  • Suggests creative alternative applications like t-shirts
  • Raises accessibility concerns for screen reader users
20.Pebble Mega Update – July 2026(repebble.com)
265 points by crazysaem 1 day ago | 181 comments | permalink
tl;dr: Pebble has shipped over 23,000 Pebble Time 2 watches (80%+ of pre-orders) and expects to be in-stock shortly, while Pebble Round 2 mass production begins late July with pre-orders fulfilled by end of September. Software updates have pushed Pebble 2 Duo battery life past 30 days, added new SDK APIs, and enabled Index 01 functionality in the mobile app. Reported hardware issues (high power draw, touch problems, cracked glass, button defects) affect a small percentage of units, and the company is offering free replacements regardless of warranty status.
HN Discussion:
  • Satisfied Pebble Time 2 owners praising the product despite shipping delays
  • Criticism of Index 01's non-rechargeable battery design and disposable nature
  • The 30-day warranty is unacceptably short for consumer electronics
  • Index 01 is misrepresented with unrealistic battery life and use case claims
  • ~Appreciation for the CEO's transparency about product flaws despite concerns