-
Shopify moves back to Native from React Native (274 points by fnthawar2)
Shopify is moving its mobile apps from React Native back to fully native Swift and Kotlin. The company says its 2020 bet on React Native was successful, but LLM-based coding agents have changed the cost calculus for building and maintaining native apps. Because building the same feature twice in Swift and Kotlin is now less expensive, Shopify reevaluated its mobile stack from first principles. It concludes that native is now the future of mobile at Shopify.
-
Rust Is Tier-1 Language at Microsoft (208 points by mmastrac)
A Rust Foundation guest post argues that Rust has become a tier-1 language at Microsoft. The content preview is unavailable, so details cannot be verified, but the title indicates Microsoft treats Rust as a first-class language for internal engineering. This likely covers tooling, support, adoption, and use cases across Microsoft products. It signals Rust's growing acceptance in large-scale systems software.
-
Hitachi launches CO2 heat pump water heaters with solar-friendly tariff controls (178 points by thelastgallon)
Hitachi has launched its Y-series EcoCute CO2 heat pump water heaters for the Japanese market. The range includes 370-liter and 460-liter fully automatic models with direct mains-pressure hot-water delivery, launching from November 2026. They add support for daytime electricity tariffs and PV-linked energy management. The units use natural refrigerant CO2 and come with a five-year manufacturer warranty.
-
Amazon pilots ad services in ChatGPT (74 points by thm)
Amazon is piloting advertising services inside ChatGPT in partnership with OpenAI. Advertisers can extend campaigns into conversational experiences through Amazon DSP as a managed service, while OpenAI handles placement through its own systems. The article is aimed at marketers figuring out what conversational ad inventory means for campaigns and measurement. Delta Vacations is mentioned as an example in the rollout.
-
What algorithm did Windows XP use to choose your initial user picture? (249 points by soheilpro)
Raymond Chen explains how Windows XP selected the initial user account picture. XP chose randomly from the default pictures in the User Account Pictures directory. The algorithm used RtlRandomEx, seeded with GetTickCount(), and a one-pass random selection method. The post responds to people trying to reverse-engineer the exact RNG behavior.
-
List of references on Sony websites to players "owning" their digital games (191 points by haunter)
A Consumer Rights Wiki page tracks a lawsuit alleging Sony misleads PlayStation Store customers about digital game ownership. Four California buyers claim they paid for games but received only limited, revocable licenses, violating California's digital-goods transparency law. The complaint points to checkout language such as “Buy Now” and “Confirm Purchase” and allegedly unclear license disclosures. Sony has moved to compel arbitration or dismiss, arguing reasonable consumers would not be misled; the case remains active.
-
>10x More Efficient Pretraining (21 points by ronfriedhaber)
Magic reports that its pretraining recipe is more than 10x more compute-efficient than leading open-weight base models. It says it matches DeepSeek V4 Pro Base using roughly 50x fewer FLOPs, about half of GPT-3's pretraining compute, or around $0.5M on GB200 hardware. Scaling 10x to about $4M reportedly outperformed all publicly available open base models on perplexity evaluations. Magic argues pretraining, agentic RL, and long-context work are sufficient to build superhuman coding agents and automate AI R&D. The blog post focuses on pretraining, with long-context covered separately.
-
To write non-fiction, draw the trunk, then the rest of the tree (39 points by DanielVZ)
The author recounts an Ancient History professor accusing him of being carried by a friend on a group essay because he often slept in class. He says his writing actually carried the group, which leads into his advice on non-fiction writing. The core metaphor is to draw the trunk of a piece first—the main argument or structure—and then add the branches and details. The post argues this trunk-first approach makes non-fiction clearer and easier to write.
-
More questions about whether researchers can trust OpenAI with unpublished math (8 points by pred_)
A Mathstodon thread raises further questions about whether researchers can trust OpenAI with unpublished mathematics. The preview is heavily truncated and requires JavaScript, so only the title and initial post snippet are visible. The discussion likely concerns confidentiality, priority, and the risk of sharing unpublished results with an AI lab. It reflects broader tensions between AI companies and academic researchers over data, credit, and competitive advantage.
-
Stockfish 19 (174 points by atiedebee)
Stockfish 19 is a new major release of the open-source chess engine. It gains up to 44 Elo over Stockfish 18 and wins more than three times as many game pairs as it loses. The release introduces universal binaries that auto-detect CPU features and an upgraded SFNNv16 NNUE network architecture and training process. The project continues to rely on community testing and contributions to improve engine strength.
AI/ML Insights & Trends
- LLM coding agents are changing platform strategy.
- Trend/insight: Shopify’s move back to native Swift and Kotlin shows that AI coding agents can lower the cost of maintaining multiple native codebases, weakening the traditional cross-platform advantage of React Native.
- Why it matters for AI/ML development: If LLMs make code generation, translation, and refactoring cheap enough, architecture decisions will increasingly optimize for runtime performance, UX, and platform integration rather than developer duplication.
-
Actionable takeaways: Reassess mobile and frontend stacks with AI-assisted engineering in mind; measure AI-generated code maintenance costs; keep native modules modular so AI agents can work across Swift, Kotlin, and shared business logic.
-
Algorithmic efficiency is becoming a primary competitive axis in foundation models.
- Trend/insight: Magic claims >10x more compute-efficient pretraining, matching DeepSeek V4 Pro Base with far fewer FLOPs and much lower cost, then scaling further to beat open base models on perplexity.
- Why it matters for AI/ML development: As frontier compute remains expensive and concentrated, better data recipes, architectures, and training methods can let smaller teams compete without 100k-chip clusters.
-
Actionable takeaways: Invest in compute-efficient training pipelines, scaling-law measurement, data quality, and evaluation per FLOP; combine pretraining gains with agentic RL and long-context work rather than relying on scale alone.
-
Conversational assistants are becoming ad and commerce surfaces.
- Trend/insight: Amazon’s pilot with OpenAI places conversational ad inventory inside ChatGPT via Amazon DSP, with OpenAI handling placement.
- Why it matters for AI/ML development: AI assistants are evolving from answer engines into discovery, recommendation, and transaction channels, which changes how products are marketed and measured.
-
Actionable takeaways: Prepare structured product feeds, conversational creative, consent and privacy controls, and attribution models; avoid over-dependence on a single assistant platform; build measurement for assistant-driven conversions.
-
Trust and governance for AI-research collaborations remain unresolved.
- Trend/insight: The Mathstodon discussion about researchers trusting OpenAI with unpublished math highlights fears around confidentiality, priority, and data use.
- Why it matters for AI/ML development: AI labs increasingly seek expert data and evaluations, but researchers need guarantees that unpublished work will not be leaked, scooped, or silently absorbed into training.
-
Actionable takeaways: Use clear data-sharing agreements, provenance tracking, opt-outs, secure evaluation environments, and independent oversight; publish model/data governance policies that academic partners can verify.
-
Open-source neural engines keep advancing through hybrid ML and community iteration.
- Trend/insight: Stockfish 19’s SFNNv16 NNUE upgrade and universal binaries deliver up to 44 Elo over Stockfish 18, showing continued gains from neural architecture, training, and distributed testing.
- Why it matters for AI/ML development: Efficient neural inference combined with domain-specific algorithms can outperform pure scale in constrained, real-time settings.
-
Actionable takeaways: Study NNUE-style hybrid approaches for edge or latency-sensitive ML; invest in community benchmarking and automated testing; package models with CPU feature autodetection for broad deployment.
-
Systems languages like Rust are moving into tier-1 status for critical infrastructure.
- Trend/insight: The Rust Foundation post frames Rust as a tier-1 language at Microsoft, signaling adoption in performance- and safety-critical software.
- Why it matters for AI/ML development: AI infrastructure—runtimes, data pipelines, inference servers, and orchestration layers—needs memory safety, concurrency, and low overhead.
-
Actionable takeaways: Consider Rust for high-performance ML infrastructure and tooling; train teams on Rust/C++/Python interop; use Rust where safety and concurrency bugs are costly.
-
Digital ownership and licensing transparency are becoming critical as AI and digital goods converge.
- Trend/insight: The Sony PlayStation digital ownership lawsuit challenges “Buy Now” language when consumers receive revocable licenses, not ownership.
- Why it matters for AI/ML development: Similar ambiguity is emerging around AI-generated assets, model weights, datasets, and commercial licenses; users and regulators will demand clearer rights.
- Actionable takeaways: Audit terms of service and licensing language; track data and model provenance; make consumer or enterprise rights explicit; prepare for regulation around digital goods and AI-generated content.
Analysis generated by deepseek-reasoner