1. Two billion email addresses were exposed
Total comment counts : 20
Summary
An analysis of the world’s largest credential-stuffing corpus confirms the “2 Billion Email Addresses” headline isn’t exaggerated: about 1,957,476,021 unique emails, plus 1.3 billion passwords (625 million new). The dataset, far larger than the previously reported 183M addresses, is the most extensive processed. Credential stuffing lists come from breaches, are sold and reused, letting attackers log into unrelated sites due to password reuse. Verification with Have I Been Pwned subscribers showed some real, active passwords—not previously seen in Pwned Passwords—underscoring the data’s legitimacy. The work demonstrates processing scale and aims to render these credentials useless on a massive scale.
Overall Comments Summary
- Main point: The thread discusses how data breaches have become ubiquitous, exposing extensive personal information and prompting diverse coping strategies and debates about breach notifications and privacy tools.
- Concern: The main worry is that personal data is routinely exposed with few effective remedies, and barriers like paid access and unclear visibility into what was breached hinder timely, actionable responses.
- Perspectives: Views range from cynical frustration with corporate data practices to practical privacy measures (password managers, aliases, disposable emails) and calls for public-sector breach notification as a public good.
- Overall sentiment: Mixed
2. Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model
Total comment counts : 31
Summary
error
Overall Comments Summary
- Main point: The discussion centers on OpenRouter’s moonshotai/kimi-k2-thinking and related open-source LLMs, their pricing, and the potential for local or modest-cluster deployment, in the context of broader AI competition.
- Concern: The main worry is that AI may become pay-to-play, widening inequities, and there are questions about whether smaller, locally runnable models can match larger frontier models in performance and reasoning.
- Perspectives: Viewpoints range from enthusiasm for more competition and open-source progress to skepticism about pricing, data residency, hardware requirements, and the practicality of running giant models locally.
- Overall sentiment: Mixed, with cautious optimism.
3. Universe’s expansion ‘is now slowing, not speeding up’
Total comment counts : 6
Summary
New study questions cosmic acceleration, suggesting the universe’s expansion may already be slowing rather than speeding up. An analysis of 300+ Type Ia supernovae found their brightness depends on the age of their progenitors—younger ones are fainter, older ones brighter. Correcting for this bias, the data no longer fit the standard ΛCDM model and align with DESI BAO+CMB results, implying dark energy weakens over time. When combined, these results indicate the universe has already entered a decelerating expansion phase. Ongoing tests, including future Rubin Observatory data, aim to confirm this paradigm shift.
Overall Comments Summary
- Main point: The discussion centers on evaluating the credibility and novelty of a cosmological claim that the universe may be an infinitely oscillating system with a big bounce.
- Concern: The primary worry is whether the claim is credible and truly new, given unclear dating, which could mislead people about cosmology and its practical implications.
- Perspectives: Opinions vary from skepticism about credibility and whether this is a real new development to curiosity about implications for space travel and the meaning of “now.”
- Overall sentiment: Mixed
4. Show HN: I scraped 3B Goodreads reviews to train a better recommendation model
Total comment counts : 27
Summary
No books have been selected yet. The user is prompted to search for and select books using the options above.
Overall Comments Summary
- Main point: Users praise the fast, data-driven book recommender but flag data gaps, overreliance on series, and several UI/interaction issues that hinder discovery.
- Concern: If data completeness, filtering controls, and explainability aren’t improved, user trust and continued engagement may decline.
- Perspectives: Viewpoints range from those seeking stronger discovery controls, negative feedback options, and explanations for recommendations to those who are largely satisfied and want only minor UI tweaks.
- Overall sentiment: Mixed.
5. You Should Write An Agent
Total comment counts : 5
Summary
Some ideas in computing are easy in the abstract, others only by doing. The author argues LLM agents are a big idea that’s incredibly approachable: a minimal HTTP API, a looping call to the LLM, and a simple tool interface. The context is just a list of strings; the LLM is stateless; tool calls are triggered by a prompt and handled by a small loop. Even a toy like this can teach you how agents work, and you can expand with more tools, persistence, and language choices. MCP isn’t essential.
Overall Comments Summary
- Main point: Building AI agents and personal assistants is increasingly an engineering-heavy endeavour, prompting a search for universal, language-agnostic architectural patterns beyond ad hoc libraries.
- Concern: There is a risk of reinventing the wheel, creating spaghetti architectures, and relying on unproven hacks like RAG, which can lead to disappointment and wasted effort.
- Perspectives: Viewpoints range from frustration with hype and grifting to a desire for concrete, proven, language-agnostic patterns for agent design (analogous to MVC or GoF patterns).
- Overall sentiment: Mixed
6. Swift on FreeBSD Preview
Total comment counts : 12
Summary
Swift released a preview toolchain for FreeBSD 14.3+ (x86_64) with a development compiler and runtimes in a tarball. It’s not a released product yet and depends on several libraries. Aarch64 support and broader coverage for FreeBSD 14 minor versions are under consideration. Users are encouraged to file issues on GitHub and provide feedback; volunteers are welcome. On FreeBSD 15, swift may fail; a temporary workaround is to install compat14x-amd64. The team aims for ABI stability across dot releases, similar to Linux, and invites community participation.
Overall Comments Summary
- Main point: The discussion centers on Swift expanding beyond Apple to FreeBSD, Linux, Android, and other platforms, and the implications for cross‑platform support, packaging, and adoption.
- Concern: The main worry is whether this cross‑platform push can be reliably sustained, including who will maintain FreeBSD ports, why certain dependencies (like Python) are required, and whether Apple will back the effort long‑term.
- Perspectives: Opinions range from enthusiastic optimism about broader platform reach and container/OCI efforts to skepticism about maintenance burden, dependency complexity, and the practicality of relying on non‑native UI toolkits.
- Overall sentiment: Mixed (cautiously optimistic)
7. ICC ditches Microsoft 365 for openDesk
Total comment counts : 23
Summary
Het Internationaal Strafhof (ICC) stapt van Microsoft 365 over op Open Desk, een Europees opensource kantoorpakket. Handelsblatt zegt dat dit mogelijk een trend in de Europese publieke sector inzet. Microsoft bevestigt de overstap maar stelt dat de dienstverlening blijft. Open Desk is ontwikkeld door Zentrum Digitale Souveränität (Zendis) voor het Duitse ministerie van Binnenlandse Zaken en maakt deel uit van DC-EDIC voor Europese digitale autonomie. Ook Nederlandse ambtenaren experimenteren met Open Desk via Mijn Bureau (Rijksoverheid, Amsterdam, VNG). De VNG vraagt om meer regie op technologie; de Digitaliseringsstrategie streeft naar digitale weerbaarheid en autonomie.
Overall Comments Summary
- Main point: The discussion centers on openDesk as an open-source office suite for the public sector, weighing its architecture, transparency, licensing, production readiness, and the political context around data sovereignty and US tech influence.
- Concern: Key worries include lack of visible roadmaps and public code repositories, unclear licensing/ownership of components, potential reliance on proprietary Microsoft services, and data security implications given geopolitical tensions.
- Perspectives: Some participants champion openDesk and open-source governance while criticizing opaque licensing and reliance on proprietary SaaS; others question production-readiness and transparency, and several point to geopolitical/policy considerations shaping technology decisions.
- Overall sentiment: Mixed
8. LLMs Encode How Difficult Problems Are
Total comment counts : 4
Summary
The message urges readers to support open access by donating to arXiv to keep science accessible. It explains arXivLabs as a framework for collaborators to build and share new features on arXiv, guided by values of openness, community, excellence, and user data privacy. arXiv only partners with those who share these values and invites ideas for valuable projects through arXivLabs, with a reminder to check arXiv’s operational status.
Overall Comments Summary
- Main point: The thread questions whether Claude and similar LLMs accurately estimate task difficulty and how training-data distribution shapes their performance, noting cases where they overestimate complexity yet solve tasks quickly and others where simple questions fail.
- Concern: The main worry is that miscalibrated difficulty judgments and biased training data could lead to missed opportunities and unpredictable outputs.
- Perspectives: Viewpoints range from skepticism about the reliability of self-estimated difficulty and data bias to cautious optimism about occasional strong performance and relevance to related research on certainty and data representations.
- Overall sentiment: Mixed
9. Open Source Implementation of Apple’s Private Compute Cloud
Total comment counts : 15
Summary
OpenPCC is an open-source framework for provably private AI inference, inspired by Apple’s Private Cloud Compute, but fully auditable and deployable on your own infrastructure. It lets users run open or custom AI models without exposing prompts, outputs, or logs, using encrypted streaming, hardware attestation, and unlinkable requests to enforce privacy. The project aims to become a transparent, community-governed standard for AI data privacy. A managed service, CONFSEC, is being built on OpenPCC. The repository includes a Go client, a C library, in-memory services, and a test client; mage is used for dev commands.
Overall Comments Summary
- Main point: The discussion evaluates the privacy guarantees of a privacy-preserving inference scheme (OpenPCC/PrivateMode) and whether it offers real benefits over simpler direct-inference setups, while highlighting a need for clearer documentation and practical use cases.
- Concern: The scheme may still allow the inference provider to read plaintext and the purported privacy benefits may be questionable in practice, with additional worries about documentation, transparency, branding, and real-world applicability.
- Perspectives: Some participants are optimistic and see potential privacy advantages and open-source value, while others are skeptical about practicality, demand source code and clearer guarantees, and compare to existing approaches like Azure Confidential Inference.
- Overall sentiment: Mixed