Published on July 16, 2026 at 06:00 CEST (UTC+2)
Inkling: Our Open-Weights Model (781 points by vimarsh6739)
Inkling: Our Open-Weights Model
Thinking Machines Lab released Inkling, a Mixture-of-Experts transformer with 975B total parameters (41B active), capable of processing text, images, audio, and video with up to 1M tokens of context. The model was pretrained on 45 trillion tokens and is designed to be a broad, balanced foundation model that prioritizes customizability over raw performance. The team also previewed Inkling-Small (12B active) and made the full weights available on Tinker for fine-tuning, emphasizing open access to let others adapt the model to their own needs.
SQLite should have (Rust-style) editions (153 points by gnyeki)
SQLite should have (Rust-style) editions
The author argues that SQLite, while excellent as an embedded database, suffers from dangerous default settings—most critically that foreign key constraints are disabled by default, leading to potential data inconsistency. They propose adopting a Rust-style edition system where users explicitly opt into different sets of defaults (e.g., strict mode, legacy mode), making safe behavior the default without breaking existing codebases.
G# – A modern .NET language with Go, Kotlin, and Swift ergonomics (55 points by serial_dev)
G# – A modern .NET language with Go, Kotlin, and Swift ergonomics
G# is a new programming language that compiles directly to .NET managed assemblies, featuring a syntax inspired by Go, Kotlin, Swift, and TypeScript. Key features include structured concurrency with async/await, precise width-bearing numeric types (e.g., int32, float64), pattern matching, and full tooling support (compiler, MSBuild SDK, VS Code extension, language server). It aims to bring modern ergonomics to the .NET ecosystem.
Grok Build is open source (317 points by skp1995)
Grok Build is open source
SpaceXAI (xAI) open-sourced Grok Build, a terminal-based AI coding agent that runs as a full-screen TUI. It can understand codebases, edit files, execute shell commands, search the web, and manage long-running tasks—either interactively, headlessly for scripting/CI, or embedded via the Agent Client Protocol (ACP). The repository is written in Rust and includes a comprehensive set of tools for extensibility.
Governments, companies, nonprofits should invest in free, open source AI [pdf] (119 points by bilsbie)
Governments, companies, nonprofits should invest in free, open source AI [pdf]
This PDF from the Siegel Endowment argues that stakeholders should invest in free and open-source AI models and tools to prevent monopolization and ensure broad access to AI capabilities. It advocates for public funding and corporate commitment to open-weight models, emphasizing the benefits of transparency, customization, and democratic control over AI development.
Bluesky Trademarks ATProto (31 points by chaosharmonic)
Bluesky Trademarks ATProto
Bluesky acquired the trademark for “ATPROTOCOL” (and variants) from another company that was threatening legal action to prevent use of the term. The move is defensive—Bluesky intends to protect the ecosystem from bad actors while making the protocol freely available. The trademark will be enforced only against those who would abuse it, and no licensing fees are planned.
Stripe and Advent have made a joint offer to acquire PayPal – sources (385 points by rvz)
Stripe and Advent have made a joint offer to acquire PayPal – sources
According to Reuters sources, Stripe partnered with private equity firm Advent International to make a bid exceeding $53 billion to acquire PayPal. The deal would reshape the payments landscape, combining Stripe’s modern developer-focused platform with PayPal’s massive user base and merchant network. The offer signals consolidation in the fintech sector.
The Tokio/Rayon Trap and Why Async/Await Fails Concurrency (11 points by LAC-Tech)
The Tokio/Rayon Trap and Why Async/Await Fails Concurrency
The article critiques async/await for conflating asynchrony with concurrency, citing Rich Hickey’s distinction between “easy” and “simple” and Rob Pike’s warning about multiple concurrency implementations. It argues that while async/await is syntactically easy, it creates hidden complexity, colored functions, and forces developers to manage scheduling—contrasting with Go’s goroutines and channels, which provide a single, unified concurrency model.
Guerrilla London Bus Ads Mock Kylie Jenner's Meta Glasses Campaign (11 points by decimalenough)
Guerrilla London Bus Ads Mock Kylie Jenner's Meta Glasses Campaign
Activist group Everyone Hates Elon (EHE) installed lenticular bus ads near Meta’s London HQ that flip between a promotional shot of Kylie Jenner wearing Meta’s camera glasses and an X-ray skeletal image with the tagline “We're always watching.” The spoof references the film They Live and highlights ongoing privacy and surveillance concerns around Meta’s smart glasses, sparking debate about consent and corporate surveillance.
LLM Networking with MikroTik (63 points by gregsadetsky)
LLM Networking with MikroTik
The author describes using LLMs (like GPT-4/Claude) to configure MikroTik networking equipment—routers, switches, and IoT devices—with great success. Despite networking’s inherent complexity (OSPF, MPLS, etc.), the LLM handled configuration tasks, troubleshooting, and even explained concepts, making advanced networking accessible to hobbyists. The author dubs this “vibe networking” and praises the combination of affordable hardware and AI assistance.
Open-weight models become a strategic priority
The release of Inkling (975B MoE, fully open weights) and the Siegel Endowment’s call for government/corporate investment in open-source AI signal a growing movement toward transparent, customizable foundation models. Unlike closed APIs, open-weight models allow fine-tuning, auditing, and local deployment. This trend reduces dependency on a few large providers and democratizes AI capabilities for research, startups, and regulated industries. Implication: Organizations should evaluate open-weight models for critical workflows to maintain control and compliance.
AI coding agents move from experiments to production tools
Grok Build’s open-source release (full TUI, extensible via Agent Client Protocol) shows that AI coding agents are maturing beyond simple autocomplete. These agents can edit files, run commands, browse the web, and manage long-running tasks—integrating directly into developer workflows. The Rust codebase and modular design encourage community contributions. Implication: Expect coding agents to become standard infrastructure, similar to CI/CD pipelines; companies should invest in agent-native tooling and evaluate security implications of code-generating agents.
LLMs enable domain experts to bridge skill gaps
The MikroTik networking post demonstrates a powerful pattern: LLMs can translate natural language intent into complex configuration commands (OSPF, MPLS, firewall rules), enabling non-experts to perform advanced tasks. This “vibe networking” approach reduces the learning curve for specialized domains (networking, DevOps, data engineering). Implication: AI platforms should target domain-specific “co-pilots” that combine LLM reasoning with validated tool schemas—opening up new markets for professional and hobbyist users.
Async/await concurrency challenges persist—relevant to AI system design
The critique of async/await (Tokio/Rayon trap) highlights that popular concurrency models hide complexity, leading to performance pitfalls in production. This matters for AI/ML systems, which often involve heavy I/O (data loading, model serving, distributed training) and parallel execution. Languages with simpler concurrency models (e.g., Go’s goroutines) are gaining favor for building robust AI infrastructure. Implication: AI engineering teams should carefully evaluate concurrency model trade-offs when choosing frameworks for inference servers, data pipelines, and agent orchestration—especially where latency and resource isolation are critical.
Surveillance backlash drives adversarial AI-art campaigns
The guerrilla ads against Meta’s camera glasses (using lenticular imagery and They Live references) reflect growing public distrust of AI-powered wearable cameras. As smart glasses become more capable (onboard AI for object recognition, real-time analysis), privacy advocates are employing creative, AI-generated visuals to highlight risks. Implication: Companies deploying AI wearables must invest heavily in transparency, consent mechanisms, and visible privacy indicators—or face reputational damage and regulatory pushback.
Fintech AI meets massive consolidation
Stripe’s joint bid to acquire PayPal for >$53B signals that AI-driven financial platforms are now central to M&A strategy. Both companies have extensive AI/ML pipelines for fraud detection, underwriting, recommendation, and personalization. A combined entity would have unrivalled transaction data and AI talent. Implication: Expect further consolidation among fintech firms as AI becomes a key differentiator; smaller players should focus on niche AI models or specialized verticals to avoid being absorbed.
Open protocol dynamics shift with defensive trademarking
Bluesky’s trademark acquisition for AT Protocol is a defensive move to protect an open standard from legal threats. This reflects a broader trend: as open protocols (e.g., ActivityPub, Matrix) gain adoption, trademark and patent trolls become a risk. Open-source AI projects should proactively secure trademarks and patents not to restrict use, but to prevent bad actors from hijacking the ecosystem. Implication: Open-weight model builders and protocol creators should allocate legal budget for defensive IP protection, especially when the technology reaches critical mass.
Analysis generated by deepseek-reasoner