-
Shopify is moving from React Native back to Swift and Kotlin (861 points by fnthawar2)
Shopify announced it is moving its mobile apps from React Native back to fully native Swift and Kotlin. The company had gone all-in on React Native in 2020 to avoid building features twice and to let non-mobile developers contribute. It says that bet worked well, but coding models have improved so much that maintaining native code no longer carries the same cost. Because one of the core assumptions behind the 2020 decision changed, Shopify reevaluated its mobile stack and concluded native is now the future.
-
More questions about whether researchers can trust OpenAI with unpublished math (736 points by pred_)
A Mathstodon thread by Andreas Thom raises further questions about whether mathematical researchers can trust OpenAI with unpublished mathematics. The discussion reflects concerns about confidentiality, consent, and what may happen to pre-publication problems, proofs, or research inputs shared with a closed AI lab. It also touches on broader anxieties about training data provenance and evaluation practices in frontier AI. The post is a social-media discussion rather than a formal paper, but its high score shows strong community interest in research ethics and data governance.
-
YuE2 · Frontier Music with Symbolic Planning (62 points by sexy_seedbox)
YuE2 is a music-generation system that unifies symbolic planning with audio synthesis. It first creates an editable symbolic score covering melody, rhythm, and chords, then renders that plan into a full song with vocals and accompaniment. The demo supports listening to generated songs, inspecting their scores in ABC notation, cover-style editing, and conversational agentic music editing. YuE2 claims frontier quality, with a best-of-8 score of 6.9632 on SongBench and performance said to rival Suno v5.
-
Mexican student creates an acoustic fire extinguisher to put out fire in seconds (81 points by rguiscard)
A 16-year-old Mexican student, Ángela Karime Venegas Hernández, created an acoustic fire extinguisher that uses sound waves to put out flames. The device combines a 12-volt battery, a frequency generator, and a speaker emitting 30 pulses per second, which pushes oxygen away from the fire. She tested it more than 100 times and found it worked on wood, flammable liquids, cooking grease, and electronic equipment in five to eight seconds. Named Vortex Tech, the project is taking her to an international fair as a residue-free, non-polluting firefighting approach.
-
Don't let anyone take away your big box of cables (371 points by Brajeshwar)
Jim Nielsen’s blog post celebrates the humble “Big Box of Cables” that many technical people keep for years. It responds to a skeet by Tyler Gaw about finally needing two cables that had sat at the bottom of his box for over a decade. Nielsen prints the quote, tapes it to his family’s “FAMILY TECHNO BOX,” and reflects on why keeping old cables is useful and meaningful. The piece is a humorous but sincere defense of practical hoarding and preparedness in tech culture.
-
OpenAI Agents API (174 points by aquir)
The OpenAI Agents API documentation introduces a platform for building stateful, tool-using AI agents. It covers core concepts such as the Responses API, conversation state, background mode, streaming, WebSocket mode, mid-turn steering, multi-agent workflows, webhooks, file inputs, compaction, and token counting. It also documents SDKs, CLI, Agent Builder, evals, safety, fine-tuning, and migration from the legacy Assistants API. The overall message is that agents are becoming first-class production primitives, not just chat-completion wrappers.
-
Thelio Mira AI Linux Workstation: 192 GB GPU Memory (59 points by jonifico)
System76’s Thelio Mira AI is a Linux workstation aimed at local AI development. It can be configured with up to a 16-core AMD Ryzen 9000 CPU, 192 GB DDR5 RAM, and dual NVIDIA RTX Pro 6000 Blackwell GPUs with 192 GB of total GPU memory. The system emphasizes sustained performance, liquid cooling, open-source hardware, and a choice of Pop!_OS or Ubuntu. It targets workloads such as multi-GPU training, fine-tuning, computer vision, image generation, simulations, and inference, with a listed price starting around $3,299.
-
Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra (366 points by seelos)
Cognition announced SWE-2, a new coding model that pushes the cost-performance frontier. It scores 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while reportedly being 64% cheaper. The model is post-trained from Kimi K33, a 2.8T-parameter base, and Cognition scaled RL to the multi-trillion-parameter regime with an algorithm that trains all reasoning-effort levels in one run. SWE-2 beats SWE-1.7 and Grok 4.6 on score and cost on several benchmarks, matches GPT-5.6 Sol and Fable 5/5.1 at a fraction of their price, and comes within a few points of GPT-6 Astra at a quarter of the cost.
-
Technique for Manipulating Satellite Photos Now Reveals Ancient Images (2025) (287 points by gumby)
NASA’s decorrelation stretch technique, originally developed for satellite imagery, is now helping reveal ancient images. The method heightens subtle contrasts in digital images, making faded features easier to see. It has been used to discover about 200 paintings in the Angkor Wat complex, including faint depictions of horseback riders and musical ensembles, as well as rock art in other settings. The NASA Spinoff article traces how a satellite-imaging algorithm became a widely used tool for archaeology and cultural heritage.
-
Google will buy half the electricity of a nuclear power plant (123 points by lukaspetersson)
Google announced its largest single investment in Europe: a €13bn package for AI infrastructure in Finland. The plan funds three new data centers, expansion of an existing site, and energy projects to support growing AI demand. As part of the deal, Google will buy up to half the electricity produced by one of Finland’s nuclear power plants. The move highlights how AI expansion is increasingly tied to large-scale energy procurement and national infrastructure strategy.
AI/ML Insights & Trends
1. Coding agents are rewriting platform economics and architecture bets. - Trend: LLM coding agents are reducing the cost of writing and maintaining platform-specific code, leading companies to reverse earlier cross-platform decisions. Shopify’s move from React Native back to Swift and Kotlin is a clear example: the main reason for React Native was avoiding duplicate native work, but AI-assisted coding has weakened that constraint. - Why it matters: Architecture choices that once traded off developer productivity against native quality may no longer hold when agents can generate, translate, and maintain code across platforms. This affects mobile, web, backend, and infrastructure teams alike. - Implications/actionable takeaways: Engineering leaders should revisit long-standing framework and abstraction decisions. Teams should measure AI-assisted velocity per platform, account for generated-code review and security, and expect more organizations to choose performance-native stacks when AI lowers the maintenance burden.
2. Agent APIs are maturing from demos into production infrastructure. - Trend: OpenAI’s Agents API documentation shows agent-native primitives such as conversation state, background mode, streaming, WebSocket mode, mid-turn steering, multi-agent workflows, webhooks, file inputs, compaction, and evals. Agent Builder, safety, and migration from the legacy Assistants API further signal platformization. - Why it matters: Reliable agents require memory, orchestration, tool calling, observability, and evaluation. Standardized APIs reduce the amount of custom plumbing teams must build before deploying agents in real products. - Implications/actionable takeaways: Builders should design agents around explicit state, tool permissions, and evals from day one. Invest in tracing, guardrails, and regression testing, and consider vendor lock-in when choosing agent runtimes. Multi-agent workflow design is becoming a core engineering skill.
3. Coding-model competition is shifting to the cost-performance Pareto frontier. - Trend: Cognition’s SWE-2 claims near-frontier coding performance at a much lower cost, using RL scaled to multi-trillion-parameter training and a single run that trains all reasoning-effort levels. It competes with models such as Fable 5.1, GPT-5.6 Sol, and GPT-6 Astra on both score and price. - Why it matters: For coding agents, raw benchmark leadership matters less than cost per successful task, latency, reliability, and integration quality. Cheaper near-frontier models make large-scale deployment of coding assistants more economically viable. - Implications/actionable takeaways: Evaluate coding models on your own repositories and workflows, not just public leaderboards. Track cost, latency, pass rates, and human review burden. Expect rapid commoditization of basic coding assistance, with differentiation moving toward agents that can safely plan, edit, test, and ship code.
4. Local AI workstations are becoming a practical complement to cloud AI. - Trend: System76’s Thelio Mira AI offers up to 192 GB of GPU memory with dual NVIDIA RTX Pro 6000 Blackwell GPUs, targeting local training, fine-tuning, image generation, simulations, and inference. - Why it matters: Not all AI work needs hyperscale cloud. Local high-VRAM machines can improve privacy, latency, data control, and cost predictability for sensitive or iterative workloads. - Implications/actionable takeaways: Organizations should consider hybrid architectures: cloud for frontier-scale training and elastic inference, local workstations for private fine-tuning, prototyping, and regulated data. Evaluate total cost of ownership, power, cooling, and staff expertise before assuming cloud-only is always cheaper.
5. AI compute growth is increasingly an energy and infrastructure problem. - Trend: Google’s €13bn Finland investment includes three new data centers and a deal to buy up to half the electricity from a Finnish nuclear power plant. - Why it matters: AI data centers require massive, reliable, and increasingly clean power. Energy availability, cost, and permitting can constrain AI capacity as much as chips or algorithms. - Implications/actionable takeaways: Energy strategy is becoming part of AI strategy. Companies should plan for power procurement, sustainability reporting, and regional infrastructure risk. Policymakers and utilities will face pressure to expand generation and grid capacity, while AI firms may compete for long-term energy contracts.
6. Generative AI is moving toward symbolic, editable, and agentic creative workflows. - Trend: YuE2 combines symbolic planning with audio generation, producing an editable score before rendering vocals and accompaniment. It also supports conversational, agentic music editing. - Why it matters: Raw generative output is difficult to control, revise, and integrate into professional workflows. Symbolic intermediate representations make AI-generated creative work more interpretable and editable. - **
Analysis generated by deepseek-reasoner