Hacker News
1Ten advances in mathematics and theoretical computer science
- OpenAI announced that their new AI model Astra solved ten major mathematical problems that had been unsolved for decades or longer, spanning fields like geometry, coding theory, and group theory. They're providing 100,000 researchers free access to ChatGPT and emphasize honest attribution when AI generates mathematical proofs, taking responsibility for correctness while crediting the AI system's contributions.
Transparency and experimental methodology
Users question OpenAI's lack of disclosure about total experimental costs, number of problems attempted, and success rates. Critics compare this to p-hacking and call for full reproducibility standards in AI-powered mathematics, including model details, prompts, and inference settings.Significance and impact of AI mathematical breakthroughs
Discussion about whether these results represent genuine mathematical advances or just effective problem-solving. Users debate the importance of the proofs, with some noting these are major breakthroughs in their fields, while others question if AI is just combining existing tools rather than creating new mathematical insights.Cost-effectiveness and resource allocation
Debate over the true costs of these mathematical discoveries, including compute, salaries, and infrastructure. Users compare AI costs to human mathematician expenses and discuss whether the reported $2,000 figure is misleading without full experimental context.
2How to Exist
- The article describes humans' difficulty with simply existing in the present moment without distraction. It proposes a meditation practice of sitting still and breathing while embracing uncomfortable feelings for a few seconds at a time, gradually building tolerance for existence without needing to escape through activities, thoughts, or distractions.
Meditation experiences and effectiveness
Users share mixed results with meditation - one practiced 6+ months with no benefits and quit after breaking their streak, while others found success with different techniques like transcendental meditation or breathing exercises for sleep.Historical shift from task-based to time-based work
Discussion of how the industrial revolution changed work from task-oriented (farming, seasonal activities) to time-controlled labor, affecting human psychology and our relationship with productivity and rest.Difficulty with doing nothing and modern restlessness
Debate about whether people genuinely struggle with inactivity or if the famous electric shock study reflects curiosity rather than discomfort, plus discussion of caffeine's role in modern inability to relax.
3How Google helped destroy adoption of RSS feeds (2023)
- Google has systematically undermined RSS adoption through an "embrace, extend, and extinguish" strategy. The company removed RSS buttons from Chrome, acquired and limited FeedBurner, shut down Google Reader in 2013, temporarily removed RSS from Google Alerts and Chrome extensions, and eliminated RSS support from Google News. These actions have damaged user confidence in RSS feeds and contributed to their declining usage, despite RSS being vital to the open web.
Google's strategic killing of RSS
Users criticize Google for shutting down Reader despite "declining usage" while simultaneously pushing the unpopular Google+. Many see this as a deliberate move to force users into walled gardens and eliminate RSS's ad-resistant, decentralized nature.Modern web's ad-driven degradation
Discussion centers on how today's internet prioritizes ad delivery over user experience, with everything from browsers to protocols designed around advertising. Users nostalgically compare the current landscape to the early 2000s web.RSS alternatives and current usage
Users share various RSS readers like NetNewsWire, discuss self-hosted solutions like FreshRSS, and debate why many sites don't offer RSS feeds. Some argue RSS remains viable despite mainstream abandonment.
4Run Kimi K3 using 29 GB of RAM at 0.50 tok/s
- WASTE is a streaming inference engine that runs the 2.78 trillion parameter Kimi K3 model on consumer laptops with 64GB RAM at 0.45-0.62 tokens per second. It keeps the model trunk in memory while streaming experts from disk, using remaining RAM as cache to make frontier-scale AI models accessible without cloud services.
LLM-generated content criticism
Users criticized the project's README and documentation as being obviously written by an LLM, calling it incomprehensible and lacking proper human review. The author defended using AI tools for development but agreed to rewrite the README manually.Performance and practicality concerns
Many users found the 0.5 tokens/second speed impractical for real use, comparing costs to cloud services and questioning the value proposition. Some suggested it's more useful as a proof-of-concept for future development.Technical implementation details
Discussion focused on custom memory management versus standard approaches like llama.cpp's mmap, quantization methods, and comparing this project to similar efforts like deltafin with different optimization trade-offs.
5Cursor removed cost information from the usage page and CSV export
- Cursor users are frustrated that the usage page has switched from showing dollar amounts to token amounts, making it difficult to track spending. The company intentionally removed dollar breakdowns for individual and Teams plans, citing confusion over included usage versus actual charges. Users can still find spending data in Dashboard > Spending or by exporting CSV files, but many want the old dollar-based usage display restored as they relied on it for budget monitoring.
Token efficiency comparison across AI agent harnesses
User compared 6 different AI harnesses using GPT 5.6 Sol, finding dramatic differences in token usage for identical tasks - from 172K tokens (smol) to 5M+ tokens (Claude Code). Discussion revealed that popular harnesses often spam context windows with unnecessary prompts and tools, significantly increasing costs.Cursor's value proposition questioned
Users debate whether Cursor IDE still offers value compared to command-line agents like Claude Code and Codex. Many have switched away citing high costs, while others appreciate Cursor's integrated experience and multi-model support for code editing and review.Cursor removes cost transparency features
Cursor removed dollar cost indicators and usage graphs from their interface, making it harder for users to track spending. Users criticized this as deliberately hiding costs to increase revenue, while Cursor staff claimed it was to reduce confusion about plan vs. on-demand pricing.
6Severance
- Mark announces layoffs affecting 7% of workforce, including everyone on the call, due to challenging macroeconomic conditions. Team members express shock and frustration, with some leaving the call. Christine explains severance includes two weeks of tokens and grief counseling prompts.
Layoff experiences and meeting tactics
Users shared personal experiences of being laid off, noting warning signs like HR-heavy meetings and muted participants. Discussion included timing strategies companies use and awkward moments during termination calls.AI-driven dystopian workplace future
Comments explored whether the transcript depicted AI agents or humans in a token-dependent economy, referencing games like iLands where AI characters face termination when tokens run out, creating disturbing parallels.Corporate callousness and employee treatment
Discussion focused on companies prioritizing AI tool access over healthcare coverage, laying off employees at month-end to minimize benefits, and the general dehumanization of termination processes.
7A Surveillance Treaty in Disguise: Canada Signs UN Cybercrime Convention
- Canada signed the UN Convention against Cybercrime despite originally opposing it alongside the US and EU. The treaty, initiated by Russia, functions more as a cross-border surveillance agreement than cybercrime protection. Human rights groups and Canadian experts warned it enables transnational repression and lacks adequate safeguards, yet the government signed without explanation after skipping the initial ceremony with key allies.
Wysiwyg politics versus political signaling
Users debate whether transparent "what you see is what you get" politics is desirable but impractical, with discussion on how politicians signal to different audiences simultaneously while making backroom deals that contradict public statements.Digital rights and privacy legislation
Commenters express frustration about lack of substantive digital rights protections, discussing the need for constitutional amendments covering right to repair, freedom from surveillance, and general purpose computing rights.Identity verification and internet anonymity
Extended discussion on how cybercrime treaties will likely lead to mandatory identity verification online, comparing this to existing financial surveillance systems and debating whether anonymous internet access can survive.
8AI doesn't generate working products, that's still your job
- AI has made building software prototypes incredibly accessible, but the gap between "this works" and production-ready software remains unchanged. The hard problems were never about writing syntax but about judgment, system design, and understanding failure modes. Computer science fundamentals are more valuable than ever because they provide the mental models needed to evaluate and improve AI-generated code. While AI automates mechanical coding tasks, it amplifies the capabilities of engineers with deep knowledge rather than replacing the need for understanding. The engineers who succeed will treat AI as a force multiplier on expertise, not a substitute for it.
Quality degradation over time
LLMs produce code that looks good individually but creates subtle architectural problems over time. Each change seems logical but the overall codebase becomes a mess that's hard to maintain, similar to how LLM-written documents feel "off" despite each section making sense.Limited production readiness
While LLMs excel at prototypes and simple CRUD apps, they struggle with complex systems requiring good architecture. The "million dollar prompt" test reveals LLMs didn't actually build production-ready code, and few examples exist of high-quality complex codebases generated by AI.Workflow approaches
Users find success using LLMs as assistants rather than autonomous agents. Step-by-step collaboration where humans handle design decisions and LLMs write implementation code works better than fully automated "vibe coding" which creates unmaintainable messes.
9Flint: A Visualization Language for the AI Era
- Flint is a new visualization language designed specifically for artificial intelligence applications and the modern AI era.
ggplot and Grammar of Graphics superiority
Users praise ggplot's API as the best charting solution even in the AI era, discussing the "Grammar of Graphics" book that inspired it. They compare it favorably to matplotlib/pyplot and appreciate its ability to recreate historical hand-drawn charts with character.questioning the need for another charting abstraction
Multiple users question why Flint is necessary when existing solutions like Plotly, ECharts, and Vega-Lite already work well with LLMs. They see it as reinventing the wheel rather than providing clear advantages over mature charting libraries.AI integration skepticism
Discussion centers on whether "for the AI era" is just marketing buzzwords, with users pointing out that LLMs can already generate charts in existing formats and questioning what specific benefits Flint provides for AI applications.