Published on June 23, 2026 at 18:00 CEST (UTC+2)
Show HN: TikZ Editor – WYSIWYG editor for figures in LaTeX (112 points by DominikPeters)
TikZ Editor – This is a WYSIWYG editor for creating TikZ figures in LaTeX, allowing users to draw diagrams interactively without writing raw TikZ code. It simplifies the process of generating publication-quality graphics for academic and technical documents. The tool is hosted on tikz.dev and gained 112 points on Hacker News.
Unlimited OCR: One-Shot Long-Horizon Parsing (266 points by ingve)
Unlimited OCR – Baidu released an open-source OCR system called Unlimited-OCR that performs one-shot long-horizon parsing, capable of processing entire documents in a single pass. It builds upon Deepseek-OCR and is available on GitHub and ModelScope, with a paper on arXiv. The project aims to enable efficient, large-scale document understanding using transformer-based models.
Spying on kids to save kids from spying is stupid (215 points by hn_acker)
Spying on kids to save kids from spying is stupid – Cory Doctorow critiques the push for mandatory online age verification and surveillance of children, arguing it is a misguided solution that harms privacy. He notes the coalition includes both anti-Big Tech activists and conservative culture warriors. The piece highlights the ambiguity in research on online harms and warns that surveillance measures can be co-opted for other purposes.
Lift4D: Harmonizing Single-View 3D Estimation for 4D Reconstruction In-the-Wild (29 points by ilreb)
Lift4D – This framework reconstructs the full geometry, appearance, and deformation of dynamic objects from a single monocular video. It uses a test-time optimization that combines a single-view 3D prior with occlusion-aware video supervision. The approach handles large occlusions and deformations, enabling 4D reconstruction from in-the-wild footage.
AI's Affordability Crisis (58 points by ilreb)
AI's Affordability Crisis – The article argues that AI platforms are heavily subsidizing their services (costing $8–$14 per generation) to create demand, but the revenue gap is growing rapidly. Citing analyses from Sequoia Capital and journalists, it warns that companies are becoming addicted to cheap tokens and will face steep price hikes. The crisis threatens the sustainability of current AI business models.
Mistral OCR 4 (146 points by meetpateltech)
Mistral OCR 4 – Mistral AI released a new OCR model with bounding boxes, block classification, inline confidence scores, and support for 170 languages. It can be self-hosted in a single container and is designed for enterprise document intelligence, RAG, and search pipelines. The model achieved a 72% win rate in human preference comparisons and an 85.20 score on the OlmOCRBench.
MSG Made Dossier on Activists Who Opposed Facial Recognition (129 points by cdrnsf)
MSG Made Dossier on Activists Who Opposed Facial Recognition – Madison Square Garden compiled a document tracking activists who criticized its use of facial recognition technology, revealed in a data breach. The file includes tweets and public comments, showing the venue’s aggressive stance against critics. This raises serious privacy and ethical concerns about corporate surveillance of dissent.
Show HN: Bun-sqlgen – Type-safe raw SQL for Bun, no ORM (22 points by ilbert)
Bun-sqlgen – This tool generates type-safe SQL queries for Bun without an ORM, by checking queries against a real Postgres or SQLite database at build time. It eliminates the need for hand-written types and catches errors before production. The project is open-source and designed for developers who want raw SQL with compile-time safety.
Plotnine (174 points by tosh)
Plotnine – Plotnine is a Python data visualization package based on the grammar of graphics, similar to R’s ggplot2. It allows users to create complex, publication-ready plots with minimal code. Version 0.15.7 includes sensible defaults for legends, colors, and labels, making it easy to go from ad-hoc plots to polished figures.
Open Source for IBM Z and LinuxONE (15 points by ncruces)
Open Source for IBM Z and LinuxONE – IBM published a report on open-source software validated for the s390x mainframe architecture in May 2026. The team tested recent versions of tools like Apache Cassandra, Grafana, MongoDB, and HAProxy. This ensures that enterprise workloads on IBM Z can run modern open-source stacks reliably.
Both Baidu’s Unlimited-OCR and Mistral OCR 4 emphasize parsing entire documents in a single pass, with bounding boxes, block classification, and confidence scores. This shift from per-page/line OCR to holistic document intelligence is critical for enterprise RAG, automated data extraction, and accessibility tools. Why it matters: It reduces pipeline complexity and latency, enabling real-time ingestion of complex documents. Implication: Expect more unified models that combine text extraction with layout understanding, and increased competition between open-source and commercial offerings in the document AI space.
Article 5 highlights that AI platforms are burning $8–$14 per token generation while charging far less, creating an unsustainable subsidy model. The revenue gap has tripled in a year to $600B. Why it matters: As more companies rely on AI APIs, sudden price hikes could disrupt workflows and force in-house alternatives. Implication: We will see consolidation among providers, a push for more efficient models (e.g., smaller, specialized architectures), and increased interest in self-hosted solutions like Mistral OCR 4 to control costs.
Lift4D demonstrates test-time optimization that combines single-view 3D priors with video supervision, handling occlusions and large deformations without requiring multi-view data. Why it matters: This enables content creation (AR/VR, VFX) from everyday smartphone footage. Implication: Expect rapid growth in tools that turn casual video into editable 3D assets, driven by diffusion-based priors and 3D Gaussian Splatting. Privacy and copyright issues around “in-the-wild” reconstruction will become more prominent.
MSG’s dossier on activists and Doctorow’s critique of child surveillance both reflect growing tension between AI-powered monitoring and civil liberties. Why it matters: Public incidents and data breaches are fueling regulatory pressure (e.g., age-verification laws, biometric bans). Implication: AI/ML developers must anticipate stricter data governance and consent requirements, especially for facial recognition. Privacy-preserving techniques (federated learning, on-device processing) will become competitive advantages.
Unlimited-OCR, Mistral OCR 4’s self-hosted option, and tools like bun-sqlgen and Plotnine show a thriving open-source ecosystem. However, the affordability crisis suggests that even open models require costly hardware to run at scale. Why it matters: Open weights lower barriers but inference costs remain high, creating a divide between those who can afford to deploy and those who cannot. Implication: We will see more focus on model distillation, quantization, and specialized hardware (e.g., IBM Z’s mainframe validation for AI workloads) to make AI affordable across different environments.
While not directly AI, Plotnine’s popularity (174 points) and bun-sqlgen’s type-safe SQL indicate demand for robust tooling that integrates with AI pipelines. Visualizations and database queries are essential building blocks for ML workflows. Why it matters: The quality and speed of data preparation directly impact model performance and development velocity. Implication: AI/ML teams should invest in tooling that reduces friction in the data stack, such as automated visualization libraries and schema-aware query generators, to keep pace with rapid model iteration.
The affordability crisis, combined with the complexity of deployment (e.g., IBM Z validation for open-source stacks), suggests many enterprises are struggling to realize ROI from AI. Mistral’s self-hosted OCR and Baidu’s open-source release are attempts to address control and cost, but adoption remains uneven. Why it matters: Hype cycles can outpace practical infrastructure, leading to failed projects and budget cuts. Implication: Companies should focus on narrowly-defined, high-value use cases (e.g., document parsing, dynamic 3D reconstruction) and measure real cost-per-outcome rather than token counts.
Analysis generated by deepseek-reasoner