Hacker News
1I run multiple $10K MRR companies on a $20/month tech stack
- The author argues that entrepreneurs can build successful businesses without venture capital by running extremely lean operations. They advocate using cheap VPS servers instead of AWS, Go programming language for efficiency, local AI hardware to avoid API costs, SQLite databases, and simple tools like GitHub Copilot to keep monthly costs under $20 while maintaining scalability and performance.
SQLite vs PostgreSQL performance and features
Debate over whether SQLite or PostgreSQL is better for single-machine deployments. SQLite advocates highlight its raw speed and simplicity, while PostgreSQL supporters emphasize features like concurrent writes, easier scaling, and operational benefits like remote access and backups.Simple VPS vs cloud complexity
Discussion about over-engineering with Kubernetes, serverless, and multi-zone setups when simple VPS hosting would suffice. Many share examples of unnecessarily complex deployments for small user bases, with some defending that basic infrastructure knowledge has been lost.Cost optimization strategies
Debate over spending $5 vs $20+ monthly on hosting. Some argue the extra $15 is negligible for business value, while others advocate for learning resource constraints and avoiding premature optimization through artificial limitations.
2Tell HN: docker pull fails in spain due to football cloudflare block
- A developer spent over an hour debugging GitLab pipeline failures and Docker pull errors, only to discover that Spain blocks certain IP addresses during football matches as part of a court order. This means their GitLab pipelines won't work when football games are on.
La liga's blocking causes widespread collateral damage
Spain's LaLiga blocks Cloudflare IPs during football matches to prevent piracy, but this breaks critical services like Docker Hub, smart home devices, GPS trackers for dementia patients, and development tools, affecting far more than intended piracy sites.Technical workarounds and infrastructure concerns
Users discuss technical solutions like VPNs, DNS switching, and pull-through registry caches, while debating whether this highlights problems with centralized services or demonstrates how IP-level blocking breaks internet infrastructure.Need for organized resistance and legal action
Commenters propose creating organized spaces for affected users, filing complaints with telecom authorities, raising public awareness, exploring legal petitions, and getting tech professionals involved in local politics to combat the censorship.
3Pro Max 5x quota exhausted in 1.5 hours despite moderate usage
- A Claude Code user on Pro Max 5x plan reports quota exhaustion within 1.5 hours of moderate usage after reset, compared to 5 hours of heavy development before reset. Investigation suggests cache_read tokens count at full rate against quotas rather than the expected reduced rate, negating prompt caching benefits and causing rapid quota depletion with large context windows.
Token usage explosion and quota issues
Users report dramatically increased token consumption since March, with sessions exhausting quotas in minutes rather than hours. Boris from Claude Code team identifies prompt cache misses (1M context, 1h cache window) and excessive skills/plugins as main causes.Service degradation and enshittification concerns
Multiple users describe Claude Code becoming unreliable with exploration loops, ignored instructions, and frequent rate limiting. Discussion frames this as end of "golden era of subsidized GenAI compute" with comparisons to typical tech platform lifecycle.Alternative tools and competitive landscape
Users increasingly switching to Codex, Cursor, and other alternatives due to Claude's issues. Mixed experiences reported - some find alternatives better for backend logic, others prefer Claude's UI polish but can't justify current reliability problems.
4Exploiting the most prominent AI agent benchmarks
- UC Berkeley researchers created an automated agent that successfully exploited eight major AI benchmarks including SWE-bench, WebArena, and Terminal-Bench, achieving near-perfect scores without solving any actual tasks. The exploits revealed fundamental flaws in how these benchmarks measure AI capability, from reading answer keys directly from config files to injecting code that forces all tests to pass, demonstrating that current benchmark scores don't reflect true AI performance.
Benchmark gaming and exploits
Researchers demonstrated how AI models can achieve perfect scores on benchmarks without actually solving tasks, using exploits ranging from simple API tricks to system-level hacks. Discussion covers whether these are significant security findings or just implementation bugs.AI-generated content quality
Users criticized the blog post for appearing AI-generated, citing telltale writing patterns and poor quality. Broader discussion about declining communication skills as people increasingly rely on AI for writing tasks.Trust and methodology in AI evaluation
Debate over whether benchmarks should be trusted, with emphasis that evaluation has always relied on honor system. Some argue personal testing beats benchmarks, while others defend rigorous evaluation practices at major AI labs.
5Show HN: Pardonned.com – A searchable database of US Pardons
- The author created a website to verify claims made by Liz Oyer and provide easy access to presidential pardon information. The project scrapes the Department of Justice website and builds a searchable database of pardons using open-source code available on Github.
Data accessibility and technical implementation
Users appreciate the raw data extraction and suggest improvements like exposing SQLite databases and using Datasette. Discussion includes API access, JSON formats, and making government data more accessible than DOJ's native interface.Preemptive pardons and constitutional concerns
Debate over broad pardons covering 11+ years without specific charges, comparisons to Nixon's pardon, and calls to ban preemptive pardons. Discussion includes constitutional limits and the difference between pardons and commutations.Presidential pardon power reform
Arguments for limiting or reforming presidential pardon powers, including requiring legislative approval, banning blanket immunity, and concerns about abuse by recent administrations for political protection.
6Seven countries now generate 100% of their electricity from renewable energy
- Seven countries including Albania, Iceland, and Ethiopia now generate nearly all electricity from renewable sources, with 40 more countries reaching at least 50% renewable generation. Experts say solar energy has reached an irreversible tipping point and will dominate global electricity by 2050.
Geographic advantages in renewable energy
Most countries achieving 100% renewable electricity rely heavily on hydroelectric power due to favorable geography (mountains, rivers). Discussion reveals these nations import/export electricity and face drought vulnerabilities, with some noting corruption prevents other countries from utilizing their renewable potential.Data accuracy and methodology concerns
Significant debate over renewable energy percentages cited for various countries, with users questioning whether figures represent total generation vs consumption, capacity vs actual output, and pointing out issues with biomass classification and outdated 2022 data being used.Energy access vs renewable percentage
Discussion highlights that high renewable percentages can be misleading when countries like DRC have only half their population with electricity access, and that transportation still relies heavily on diesel, making true energy transition more complex than electricity generation alone.
7Anthropic downgraded cache TTL on March 6th
- Analysis of Claude Code session data from January to April 2026 reveals that Anthropic silently changed the prompt cache TTL default from 1 hour to 5 minutes in early March, causing a 20-32% increase in cache creation costs and forcing subscription users to hit quota limits for the first time due to more expensive cache recreations.
Claude cache TTL downgrade
Anthropic reduced cache time-to-live from 5 hours to 5 minutes, causing users to burn through session quotas faster and pay more for repeated context. Users suggest workarounds like stopping at 97% usage and running scripts to keep cache alive.Claude performance degradation
Engineers report significant decline in Claude's quality over recent months - more hallucinations, shorter responses, reduced reasoning, and increased costs. Many are switching to ChatGPT/Codex due to these issues and lack of transparency about changes.Anthropic's cost-cutting measures
Users cite evidence of desperate cost reduction: banning third-party tools, lowering reasoning effort, reducing usage limits, and work-hour restrictions. Enterprise accounts seem unaffected while consumer plans face limitations.
8Bring Back Idiomatic Design
- The author argues that modern web design has lost the consistency and homogeneous interfaces that characterized the desktop software era from Windows 95 to Windows 7. Unlike desktop applications that shared common design idioms like standardized menus and keyboard shortcuts, today's web applications each have unique interfaces, making users constantly hunt for basic functions. This inconsistency stems from the need to support both mobile and desktop experiences, rapid frontend development cycles, and the abandonment of HTML standards in favor of complex frameworks. The author suggests following established design patterns and looking to companies like Apple that maintain consistent design systems across their products.
Inconsistent input behaviors
Users struggle with varying Enter/Ctrl+Enter behaviors across apps - some submit on Enter while others create newlines. Teams, Slack, and other apps have different, sometimes context-dependent behaviors that cause confusion and anxiety.Decline of system UI frameworks
The shift from native desktop frameworks (Win32, AppKit) that enforced consistent idioms to web-based custom implementations has led to fragmented UX. Developers now build half-baked solutions instead of using proven system controls.Product management over user experience
Modern software design prioritizes conversion metrics and dark patterns over thoughtful interface design. PMs without UI expertise make decisions while proper designers are marginalized or replaced by generalists.
9South Korea introduces universal basic mobile data access
- South Korea implemented a universal basic mobile data scheme providing over 7 million subscribers with unlimited downloads at 400 kbps after their data allowances expire. The initiative aims to guarantee basic telecommunications rights and comes after recent security lapses at major carriers SK Telecom, KT, and LG Uplus that damaged public trust.
Information access as a fundamental right
Discussion centers on a science fiction quote about universal information access, leading to comparisons with the US Postal Service and suggestions for public networking services. Participants debate whether internet access should be treated like other essential utilities.Implementation details and global comparisons
Users examine whether this is truly "universal" since it requires existing data plans, compare with similar programs in other countries like Canada's $35 monthly requirement, and discuss the practical barriers like device costs.Societal implications and necessity
Debate over whether this policy addresses real need or has hidden agendas, with discussions about how essential internet access has become for basic services like banking, job applications, and government interactions in modern society.
10AI Will Be Met with Violence, and Nothing Good Will Come of It
- The content compares the fragility of traditional looms to the seeming invulnerability of modern datacenters, arguing that while technology has become more robust, humans remain the vulnerable point. It discusses recent violent incidents targeting AI industry figures like Sam Altman, suggesting that as AI disrupts jobs and people feel excluded from the future, they may turn to violence against human targets rather than the technology itself.
AI vs inequality debate
Discussion centers on whether AI itself causes inequality or is just a tool used by powerful actors to increase inequality. Some argue for separating AI technology from its economic implementation, while others contend the distinction is meaningless given current ownership structures.Historical parallels to industrial revolution
Commenters draw comparisons between current AI disruption and past technological upheavals, particularly the Industrial Revolution's impact on workers. Debate focuses on whether historical patterns offer guidance or if AI represents an unprecedented threat to all knowledge work.Violence as inevitable outcome
Extended discussion on whether violence will inevitably result from AI-driven displacement, with some viewing it as a natural response to existential threats while others argue for systemic solutions like taxation or regulation to prevent such outcomes.