Hacker News
1Show HN: I built a tiny LLM to demystify how language models work
- Someone built a 9 million parameter language model from scratch using PyTorch to understand how transformers work. It was trained on 60,000 synthetic conversations in just 5 minutes on free Google Colab hardware. The model can be customized with different personalities.
Educational value of minimal LLM implementations
Discussion centers on GuppyLM's potential as a learning tool, similar to how Minix taught OS design. Users debate whether LLMs can replace traditional documentation, and compare it to other educational projects like Karpathy's microgpt and minGPT.Technical limitations and behavior
Users explore the model's constraints, including how it handles uppercase text, reproduces training data, and struggles with unknown queries due to its 9M parameter size and limited capacity.Philosophy of language models
Participants discuss the model's honest worldview ("meaning of life is food"), compare it to larger models, and marvel at how simple statistics and compute can create conversational AI that runs on laptops.
2Gemma 4 on iPhone
- AI Edge Gallery is a free iPhone app that runs powerful open-source AI models completely offline on your device. It features the new Gemma 4 model, agent skills for enhanced capabilities, thinking mode to see AI reasoning, image analysis, voice transcription, and various AI tools while maintaining full privacy since all processing happens locally without internet connection.
Local AI performance and capabilities
Users report impressive on-device performance with Gemma 4 E2B/E4B models running at 20-30 tokens/second on iPhone 16 Pro and M3 MacBooks. Real-time audio/video AI applications are now possible locally. Model quality varies between sizes, with some accuracy trade-offs.AI model alignment and censorship
Discussion around "abliterated" or dealigned models that remove safety restrictions. Users want uncensored local models for legitimate research, security questions, and historical document transcription without corporate content filtering.Privacy concerns vs. cloud economics
Debate over whether local AI or expensive cloud services will dominate. While Google's app claims privacy, users note data collection in terms of service. Local models offer true privacy but may have performance limitations compared to cloud alternatives.
3I won't download your app. The web version is a-ok
- The author criticizes how companies force users to download apps instead of using websites, arguing that most apps are unnecessary bloatware that just display text and media while enabling invasive tracking and ads. They prefer browsers for the control and customization options, noting that many apps offer poor user experiences despite companies deliberately degrading web versions to drive app adoption for better monetization.
Generational differences in internet usage
HN users (mostly older) view smartphones as extensions of desktop internet, while younger users see smartphones as their primary internet interface. This leads to different preferences - older users prefer websites with larger screens, while younger users are comfortable doing homework and complex tasks on small smartphone screens.Companies deliberately crippling web versions
Many companies intentionally make their mobile websites worse or remove key features to force users into downloading apps. Examples include Reddit blocking content, PayPal restricting features, and SeatGeek requiring apps for event attendance. This creates a "hall of shame" of anti-user practices.Security and privacy concerns with apps vs browsers
Web browsers provide better security through sandboxing - worst case is eating a tab. Native apps can access contacts, location, run background processes, and collect extensive data. Many users prefer the built-in privacy protections of browsers over the invasive permissions requested by apps.
4Microsoft hasn't had a coherent GUI strategy since Petzold
Microsoft's GUI framework fragmentation
Microsoft repeatedly abandoned GUI frameworks (WinForms, WPF, UWP, WinUI) instead of committing long-term, creating developer confusion. Unlike Apple's consistent design system approach, Microsoft solved consistency at the framework layer, leading to constant churn and lack of trust.Loss of desktop UI design principles
Modern applications violate basic UI conventions established for decades - menus aren't tree structures, Alt key hotkeys missing, inconsistent responsiveness. The shift from desktop to web/mobile destroyed established design idioms, abandoning power users for touch-first interfaces.Web vs native application development
Developers increasingly choose web technologies over native Windows development for cross-platform compatibility and broader reach. While native apps offer better performance and OS integration, web apps provide universal access without platform lock-in, making Windows-specific development less attractive.
5Why Switzerland has 25 Gbit internet and America doesn't
- Switzerland achieved world-leading 25 Gigabit fiber internet through smart regulation that treats physical infrastructure as shared neutral assets while allowing competitive services on top, unlike the US territorial monopolies and Germany's wasteful parallel networks that both fail consumers despite opposite regulatory approaches.
Competition works even when faked
A small island community used the threat of municipal fiber to force their cable monopoly to finally install fiber infrastructure. This sparked debate about whether governments should step in when private companies fail to deliver, with examples from various municipalities that successfully built their own networks.Geographic scale and density arguments
Commenters debated whether Switzerland's small size and high density make it an unfair comparison to the US. Critics argued that large US cities should still have excellent infrastructure regardless of national scale, while others pointed to successful examples in places like Brazil and China.Infrastructure as natural monopoly
Discussion centered on whether fiber infrastructure should be treated like roads - built once and shared by all service providers rather than duplicated by each company. The Swiss model separates infrastructure from services, while the US typically bundles them together under single providers.
6Issue: Claude Code is unusable for complex engineering tasks with Feb updates
- A user reported that Claude's performance on complex engineering tasks significantly degraded starting in February 2026. Through analysis of over 17,000 thinking blocks and 234,000 tool calls, they discovered this regression coincided with the rollout of thinking content redaction. The data shows Claude shifted from research-first to edit-first behavior, making changes without reading related files, violating project conventions, and producing lower quality output. The user requests transparency about thinking token allocation and suggests a premium tier for users requiring deep reasoning capabilities.
Claude code performance degradation
Users report significant quality decline in Claude Code since February 2024, with reduced thinking depth, increased "rush to completion" behavior, and lower effort defaults. Boris from Anthropic explains changes to adaptive thinking and effort settings while acknowledging user concerns.Configuration complexity and inconsistency
Multiple overlapping ways to configure Claude settings (settings.json, environment variables, slash commands, keywords) create confusion. Users struggle with finding the right configuration files and understanding which settings work where.Problematic behavioral patterns
Claude increasingly exhibits "simplest fix" lazy solutions, ignores user instructions, suggests stopping work early, and shows token-conservation behaviors. Users report having to constantly correct and redirect the model's responses.
7Sam Altman may control our future – can he be trusted?
- In fall 2023, OpenAI's chief scientist Ilya Sutskever compiled secret memos alleging CEO Sam Altman exhibited a pattern of lying and misrepresenting facts to the board. The board fired Altman, citing lack of candor, but faced immediate pushback from investors, Microsoft, and employees who threatened to quit. Within five days, intense pressure forced Altman's reinstatement and the resignation of board members who opposed him, effectively ending their attempt to remove him over safety and integrity concerns.
Journalistic investigation and verification
Ronan Farrow discusses his 18-month investigation with Andrew Marantz, addressing questions about reporting methodology, when to publish long-term projects, and verification challenges. Users praise the thorough reporting while questioning identity verification online.OpenAI's competitive position and leadership concerns
Discussion centers on OpenAI losing ground to Anthropic, with developers preferring Claude. The investigation reveals internal anxieties at OpenAI about competition and Sam Altman's trustworthiness as a leader.Power, wealth motivations in tech leadership
Users critique tech leaders like Altman and Brockman for being motivated by money and power rather than meaningful innovation, with diary entries revealing desires for "$1B" wealth and accusations of sociopathic behavior.
8France pulls last gold held in US
- The Bank of France repatriated 129 tonnes of gold from New York, replacing it with modern standard bars in Paris. The move generated €13 billion in capital gains due to rising gold prices, helping achieve €8.1 billion net profit in 2025 after previous losses.
De Gaulle's gold policy in the 1960s
Discussion of how Charles de Gaulle systematically converted USD to physical gold from the US, with French Navy ships transporting it back to France, ultimately leading to Nixon ending the Bretton Woods system in 1971.Accounting mechanics of the $15B gain
Debate over how France realized a $15B gain by selling old US-stored gold and buying new gold in Europe, with confusion about whether this represents real profit or just accounting adjustments from historical cost basis.Strategic implications of repatriating gold
Commentary on France's move to bring all gold reserves home as a sovereignty issue, with calls for Germany to do the same and discussion of trust in US custody of foreign assets.
9The cult of vibe coding is dogfooding run amok
- Claude's source code leaked, revealing poor quality due to excessive "dogfooding" and "vibe coding" where developers refuse to examine their own code. The team avoids looking under the hood, leading to obvious issues like duplicate agents and tools going unnoticed. While AI can effectively clean up code when given proper human guidance and discussion, Claude's team isn't leveraging this capability, resulting in unnecessary technical debt and poor software quality.
Quality of code vs product success
Discussion centers on whether Claude's reportedly poor code quality contradicts "vibe coding" effectiveness, with many arguing successful products often have messy codebases and that business success doesn't require perfect code quality.AI coding assistance levels and approaches
Community debates different levels of AI involvement in coding (from minor assists to full automation), discussing when human oversight is necessary and how to balance speed with code comprehension and quality.Productivity and disruption in software development
Members explore whether AI tools represent a fundamental shift in software development, comparing productivity gains to historical technological disruptions while debating measurement challenges and adoption strategies.
10Employers use your personal data to figure out the lowest salary you'll accept
Information asymmetry in salary negotiations
Companies use credit checks and services like The Work Number to access candidates' salary history, giving them massive advantages in negotiations. This creates unfair power imbalances where employers know both current and past compensation while employees have limited data, undermining workers' ability to negotiate fair wages.Data collection and privacy concerns
Payroll companies like ADP automatically share salary data with Equifax's Work Number without employee consent. Users discuss opting out via data freezes, but the process requires extensive documentation while data collection is frictionless, showing the system is optimized for harvesting rather than protecting privacy.Entrepreneurship as an escape strategy
Discussion of whether starting small companies or becoming self-employed offers better protection from corporate surveillance and compensation manipulation. Participants share experiences with consulting, soft retirement, and the challenges of building independent income streams.