Published on July 18, 2026 at 18:00 CEST (UTC+2)
GPT-5.6 used a prompt to close a 30-year gap in convex optimization (226 points by mbustamanter)
GPT-5.6 used a prompt to close a 30-year gap in convex optimization
This article reports that OpenAI’s GPT-5.6, using a similar prompt to a previous CDC proof announcement, managed to close a long-standing gap in convex optimization. The solution was formally verified in the Lean theorem prover, adding credibility to the result. The post highlights the model’s ability to tackle advanced mathematical problems that had remained unsolved for decades, signaling a new capability for AI in rigorous mathematics. It also sparks discussion about how prompting strategies can unlock surprising reasoning abilities in large language models.
Gleam Is Now on Tangled (18 points by nerdypepper)
Gleam Is Now on Tangled
Gleam, a friendly, type-safe programming language for building scalable systems, has been added to Tangled — a decentralized, self-hosted version control and collaboration platform. The language’s repository on Tangled includes its compiler, tooling, and documentation. This move reflects a growing interest in alternative hosting platforms beyond GitHub, especially those emphasizing decentralization and user control. For Gleam developers, it provides a new, community-driven way to contribute and track changes.
Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help? (128 points by couAUIA)
Fable 5 vs. GPT-5.6 Sol on an NP-Hard Problem: Does /goal help?
The author compares Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on a real-world NP-hard fiber-network optimization problem. Fable 5 outperformed GPT-5.6 Sol, achieving the best solution overall with remarkable consistency. The /goal mode (a prompt technique to encourage targeted reasoning) did not consistently improve performance — sometimes it helped, sometimes it harmed. The post emphasizes that raw model intelligence, not just prompt engineering, drives results on complex, constraint-heavy tasks.
Is this the end of the once-mighty GoPro? (69 points by aanet)
Is this the end of the once-mighty GoPro?
This article examines GoPro’s declining fortunes in the action-camera market, speculating whether the company can survive amid fierce competition from smartphones, cheaper rivals, and shifting consumer preferences. It notes that the GoPro HERO line, once dominant, now struggles to differentiate itself. The piece reflects broader trends in hardware: commoditization, market saturation, and the challenge of maintaining relevance when software and ecosystem matter more than specs.
Tech note: making your own V-I plots at home (25 points by zdw)
Tech note: making your own V-I plots at home
Written by security researcher Michał Zalewski (lcamtuf), this tech note explains how to accurately capture voltage-current (V-I) curves of semiconductor devices using inexpensive lab equipment. The author critiques the fake or retraced diagrams common in electronics textbooks and shows how to obtain real, reproducible measurements. The post is a practical guide for hobbyists and engineers, emphasizing the importance of empirical data over idealized schematics.
LG monitors silently install software through Windows Update without consent (602 points by baranul)
LG monitors silently install software through Windows Update without consent
LG monitors have been found to push driver or utility software onto Windows machines via Windows Update without explicit user consent. Users report unexpected installations and system changes, raising privacy and control concerns. The story underscores ongoing tensions between hardware vendors, OS update mechanisms, and user autonomy, reminiscent of similar controversies with other peripherals.
Regressive JPEGs (537 points by vitaut)
Regressive JPEGs
This project (likely an art or research experiment) explores “regressive” JPEGs — images that degrade in quality through iterative re-compression, creating a visual decay effect. The technique may illustrate lossy compression artifacts or serve as a commentary on digital image fidelity. Given its high upvote count, it resonates with those interested in generative art, compression algorithms, or the aesthetics of data loss.
EU ban on destruction of unsold clothes and shoes enters into application (81 points by robtherobber)
EU ban on destruction of unsold clothes and shoes enters into application
As of July 19, 2026, large companies in the EU are prohibited from destroying unsold clothing, accessories, and footwear; medium-sized companies must comply by 2030. The regulation aims to reduce waste, encourage reuse and recycling, and promote a circular economy. Businesses must prioritize selling, donating, or refurbishing unsold goods. This is the first such ban at a major regional scale and could influence global supply chain practices.
What AI did to stackoverflow in a graph (196 points by secretslol)
What AI did to stackoverflow in a graph
This query on Stack Exchange Data Explorer likely shows a sharp decline in new questions, answers, or user activity on Stack Overflow coinciding with the rise of AI coding assistants (like ChatGPT, Copilot). The graph visualizes a trend many have observed anecdotally: developers increasingly rely on AI for quick answers, reducing traditional Q&A traffic. It raises questions about the future of community-driven knowledge bases and the role of AI in software development.
Elixir-lang.org has a new design (4 points by bbg2401)
Elixir-lang.org has a new design
The official Elixir programming language website has been redesigned with a modern, clean layout emphasizing the language’s strengths: composability, scalability, fault-tolerance, and developer productivity. The site highlights new features like Elixir v1.20, Livebook, and Numerical Elixir, and showcases adoption by both solo developers and large companies. The redesign reflects the language’s growing maturity and community investment in developer experience.
AI is making genuine contributions to mathematical research
The GPT-5.6 convex optimization result (Article 1) shows that LLMs can solve open problems in pure mathematics, not just code or trivia. This trend matters because it extends AI’s utility to theorem proving and formal verification (Lean). Implication: researchers should treat LLMs as collaborators for generating conjectures or proofs, but rigorous validation (via formal systems) remains essential. The gap between “interesting suggestion” and “verified theorem” is narrowing.
Prompt engineering is not a universal “try harder” button
Article 3’s test of /goal mode on an NP-hard problem reveals that sophisticated prompting does not guarantee better results. Sometimes it worsens performance by locking the model into suboptimal search paths. Takeaway: understanding a model’s inherent reasoning strengths is more important than tweaking prompt syntax. For complex tasks, we need better benchmarks that isolate raw intelligence from prompt sensitivity.
AI is reshaping developer platforms and knowledge sharing
Article 9’s graph shows Stack Overflow traffic declining under AI’s influence. Combined with Article 2 (Gleam on Tangled) and Article 10 (Elixir redesign), we see a shift: developers move toward decentralized, AI-augmented tools and curated documentation. Trend: AI copilots reduce the need for basic Q&A; instead, communities focus on niche, high-level discussions and curated resources. Implications: platform designers must integrate AI assistants while preserving human collaboration.
Hardware vendors and OS updates create new trust challenges
Article 6 (LG monitor software via Windows Update) highlights a tension: automated updates can bypass user consent, eroding trust. In an AI/ML context, similar issues arise when models or agents silently update their behavior, or when hardware uses “AI” features that phone home. Trend: the line between helpful automation and intrusion is blurring. Takeaway: transparency and opt-in mechanisms are not optional — they are foundational for user acceptance of AI-driven systems.
AI models show uneven performance across real-world optimization tasks
Article 3’s comparison between Fable 5 and GPT-5.6 Sol on a fiber-network design problem reveals that model choice matters enormously for NP-hard problems. Fable 5’s consistency was unprecedented. This underscores that “general intelligence” does not yet translate equally to all problem classes. For AI/ML practitioners: benchmark suites must include realistic, constrained optimization tasks (not just coding puzzles) to evaluate models’ practical utility.
Sustainability regulations are forcing companies to rethink product lifecycles
The EU ban on destroying unsold clothes (Article 8) is part of a broader trend toward circular economy legislation. For AI/ML, the same principles apply: training models consumes resources, and “data waste” (unused training data, discarded checkpoints) has an environmental cost. Trend: expect future regulations to mandate model reuse, open-weight sharing, and efficient training practices. Actionable: start measuring and reporting carbon footprints of ML workflows now.
Open-source languages and decentralized platforms gain traction
Gleam’s move to Tangled (Article 2) and Elixir’s redesigned site (Article 10) signal that developers are actively seeking alternatives to centralized, corporate-controlled ecosystems. Combined with AI-generated code, the barrier to creating and distributing new languages lowers. Trend: we may see a proliferation of domain-specific languages (DSLs) optimized for AI-assisted development. Implication: AI tools should support not just Python/JavaScript, but smaller, type-safe languages that prioritize correctness and concurrency.
Analysis generated by deepseek-reasoner